TWI802334B - Multiple-variable predictive maintenance method for component of production tool and computer program product thereof - Google Patents

Abstract

Embodiments of the present invention provide a multiple-variable predictive maintenance method for a component of a production tool and a computer program product thereof, in which a multiple-variable time series prediction (TSP_MVA) and an information criterion algorithm are adapted to build a best vector autoregression model (VAR), thereby forecasting the complicated future trend of accidental shutdown of the component of the production tool. Therefore, the multiple-variable prediction of the present invention can improve the accuracy of prediction compared with the single-variable prediction.

Description

Multivariate predictive maintenance method for production machine components and its computer program products

The present invention relates to a multivariable predictive maintenance method for production machine components and its computer program product, and in particular to a multivariable predictive maintenance method for production machine components based on remaining useful life (Remaining useful life; RUL) prediction and its computer program products.

Production machines are an integral part of any manufacturing plant. The failure of components, modules or devices in the production machine (such as: heaters, pressure modules and throttle valves (Throttle Valve), etc.) will cause abnormal production, resulting in poor product quality and/or reduced production capacity, resulting in major loss.

Generally, the most common method to solve the above problems is regular preventive maintenance (PM). That is, maintenance-related work is performed at preset time intervals. The preset time interval is basically determined according to the mean time between failure (Mean Time between Failure) of the target device (TD). Therefore, how to arrange an appropriate PM plan is usually a task Key issues of the plant. An improperly scheduled PM program can increase maintenance costs or reduce productivity.

The purpose of predictive maintenance is to find out when the target equipment (ie, the components of the production machine) is sick and perform immediate maintenance before the death of the target equipment occurs, so as to avoid unexpected target equipment downtime. In this way, not only the utilization rate and manufacturing quality of the production machine are improved, but also the extra cost of excessive maintenance in preventive maintenance can be reduced.

In order to improve the tool maintenance plan to increase the performance of the fab, the International Sematech Manufacturing Initiative (ISMI) proposes a predictive and preventive maintenance (PPM) index. As defined by ISMI, PPM includes preventive maintenance (PM), condition-based maintenance (Condition-based Maintenance; CbM), predictive maintenance (Predictive Maintenance; PdM) and post-breakdown maintenance (Breakdown Maintenance; BDM). Among them, ISMI advocates that CbM and PdM technologies should be developed and used in the form of a single module or multiple modules so that end users can use these technologies efficiently. CbM is defined as: "Maintenance performed after one or more indicators indicate that the machine is about to fail or that its performance is deteriorating". Fault Detection and Classification (FDC) is a method related to CbM, which is defined as: "monitoring machine and factory data to assess the health of the machine, and issuing an alarm and/or when an error is detected or turn off the machine". On the other hand, PdM is a technology that uses a predictive model to find out the relationship between equipment status information and maintenance information to predict the remaining of the machine or target device (TD) Life (Remaining Useful Life; RUL), in order to achieve the goal of reducing maintenance events of unplanned downtime.

In some conventional technologies, only a single feature is used to predict the aging characteristics of the equipment, and then determine the remaining service life of the equipment. How to increase the accuracy of prediction is a topic of concern to those skilled in the art.

Due to the limitations of conventional algorithms, when the target device is about to die, if the aging characteristics of the target device suddenly rise or become smooth, the exponential model may not be able to keep up with the real-time prediction or even wrongly predict the RUL of the target device.

One object of the present invention is to provide a multi-variable predictive maintenance method for production machine components and its computer program product, so that the RUL of the production machine components can be predicted in real time and accurately, and the maintenance of the production machine components can be carried out in time.

Another object of the present invention is to provide a multi-variable predictive maintenance method and its computer program product for production machine components, by proposing an early warning mechanism and a death correlation index (Death Correlation Index; DCI), so that the production machine components may Immediately perform maintenance in the dead state, and present the possibility of the production machine components entering the dead state in a digital way.

According to an aspect of the present invention, a multivariable predictive maintenance method for production machine components is provided. Firstly, multiple sets of process data used by the components of the production machine to sequentially process multiple workpieces are obtained, wherein each set of process data contains multiple values of parameters. Then, a plurality of event indication values respectively corresponding to the groups of process data are obtained, wherein the event indication values indicate whether the component has an abnormal event when the component is processing each workpiece. Then, the values of these parameters of each set of process data are converted into the values of a plurality of parameter indexes by using a plurality of algorithms respectively. Then, a correlation analysis is performed between each parameter index in the groups of process data and the event indicator values, so as to obtain a plurality of correlation coefficients respectively corresponding to the parameter indexes. Then, select the parameter index corresponding to the largest one of these correlation coefficients as an aging feature, and set an auxiliary aging feature. Next, a first judging step is performed to determine whether the component is in a sick state when processing the workpieces according to the value of the aging characteristic corresponding to each workpiece, wherein once the component is processing one of the workpieces is in the sick state, set this artifact as a sample selection point. Then, a multivariate modeling step is carried out, and the multivariate modeling step includes: N values corresponding to this aging feature and N values corresponding to the auxiliary aging feature in N sets of process data corresponding to the N workpieces before using this sample selection point The values are a set of modeling sample data, where N is a positive integer. Next, perform a Granger causality test on the aging features and the auxiliary aging features to determine the correlation between the aging features and the auxiliary aging features, and delete the auxiliary aging from the group modeling sample data if there is no correlation feature. Then, use the set of modeling sample data to establish an aging characteristic prediction model according to a multivariate time series prediction algorithm, and obtain a plurality of prediction values of the aging characteristics arranged according to the production sequence of the workpieces. These predicted values are then converted into a plurality of remaining useful life (RUL) predicted values using a workpiece processing time for each workpiece and a death specification value of the aging characteristic when the component becomes unusable (RUL _t ), where t represents the tth workpiece, and t is an integer. Then, a second judging step is performed to judge whether the component needs to be replaced or repaired according to the remaining service life prediction values (RUL _t ).

In some embodiments, in the aforementioned first judging step, firstly, a set of conversion formulas are used to convert the values of the aforementioned aging characteristics of each group of process data into a plurality of Device Health Indexes (Device Health Indexes) respectively corresponding to workpieces. ;DHI), the conversion formula of this group is:

； when

;

； when

;

； when

;

； when

;

； when

;

； when

;

為此些組製程資料之該老化特徵的數值的平均 值，其中

為

所對應之轉換值； in

This is the average value of the values of the aging characteristics of these sets of process data, where

for

The corresponding conversion value;

Max y _T is the maximum value of the aging characteristic of these groups of process data, and Max y _{T_ mapping} is the conversion value corresponding to Max y _T ;

Min y _T is the minimum value of this aging characteristic of these groups of process data, and Min y _{T_ mapping} is the conversion value corresponding to Min y _T ;

LSL is the lower specification limit; LCL is the lower control limit; UCL is the upper control limit; USL is the upper specification limit; _{LSL_mapping} is the conversion value corresponding to LSL; _{LCL_mapping} is the conversion value corresponding to LCL; _{UCL_mapping} is the conversion value corresponding to UCL Conversion value; _{USL_mapping} is the conversion value corresponding to USL.

Then, determine whether the device health indices are greater than or equal to a threshold value in sequence, and set the artifact corresponding to the first one of the device health indices greater than or equal to a threshold value as the aforementioned sample selection point.

In some embodiments, in the aforementioned multivariate modeling step, a vector autoregression model (Vector AutoRegression Model; VAR) is firstly used as the aforementioned multivariate time series forecasting algorithm to establish the aforementioned aging characteristic prediction model. Use the partial autocorrelation function (PACF) to select the maximum number of lagging periods for this vector autoregressive model. Then, a white noise test is performed on the values in the modeling sample data, and when these values are white noise, they are added to another set of process data corresponding to the N+1 workpieces before the sample selection point to correspond to this Numerical values for aging features to model sample data. Next, use the maximum number of backward periods of this VAR model to create a plurality of VAR model combinations. Then, an information criterion algorithm is used to calculate the information amount of each vector autoregressive model combination. Then, select one of the vector autoregressive model combinations with the largest amount of information as a best model.

In some embodiments, the aforementioned information criterion algorithm is Bayesian Information Criteria (BIC).

In some embodiments, the aforementioned multivariate modeling step further includes: judging whether the variance of the values in the modeling samples will increase over time, wherein when the variance of these values increases over time When the value is larger, perform logarithmic transformation on each value of the modeling sample data; perform a unit root test on these values to confirm whether these values arranged in sequence are in a steady state, where When these values are not in a steady state, differential conversion is performed on each value of the modeling sample data.

In some embodiments, the aforementioned single root test is Augmented Dickey-Fuller (ADF) test or Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test.

In some embodiments, the aforementioned second determination step includes:

Judging whether ( RUL _t - RUL _{t - 1} )/ RUL _{t - 1} is greater than or equal to a threshold value, and obtaining the first result, wherein t - 1 represents the t-1th workpiece; judging whether RUL _t is less than a maintenance buffer time , and the second result is obtained, wherein the aforementioned components must be repaired during this repair buffer time; when both the first result and the second result are no, the component is in a sick state but not deteriorating rapidly, and maintenance is not required; when the second When the first result is no and the second result is yes, the component has not deteriorated rapidly but its remaining service life is insufficient, and maintenance is required; when the first result is yes and the second result is no, the component has deteriorated rapidly. If the treatment continues For each of the t+i-th workpieces, the first result is yes and the second result is no, the component needs to be inspected or repaired, where i is a positive integer; and when both the first result and the second result are yes , this component requires repair.

In some embodiments, the aforementioned second determination step includes:

Using a set of conversion formulas to convert the values of the aforementioned aging characteristics of each set of process data into a plurality of death correlation indices (Death Correlation Index; DCI) respectively corresponding to the aforementioned components of the workpiece, the set of conversion formulas are:

Among them, y _death is the value of the aging feature corresponding to the component in the dead state, y _{t - 1 is} the value of the aging feature corresponding to the component when processing the t-1th workpiece, and conv is the covariance calculation , Var is the variation calculation;

When the DCI _t is greater than a threshold value, it means that the component is close to a dead state when processing the t-th job, wherein the calculation of the threshold value is based on the standard deviation of the DCI _t .

In some embodiments, the aforementioned component is a heater, a pressure module, a throttle valve, an oil-free bushing or a bearing, and these parameters include: a shaft deflection, a valve opening, a vibration amplitude, A driving voltage, a driving current, a temperature and a pressure.

In some embodiments, the aforesaid parameter index includes: k multiplied frequency (wherein k is greater than 0) after converting to the frequency domain, an overall similarity index (Global Similarity Index; GSI), a kurtosis of a statistical data distribution ( kurtosis), a statistical distribution of skewness (skewness), a standard deviation, a root mean square (root mean square), a mean, a maximum and a minimum.

According to another aspect of the present invention, a computer program product is provided. After the computer program product is loaded and executed, the basic multivariable predictive maintenance method of the aforementioned target device can be completed.

Therefore, by applying the embodiment of the present invention, the RUL of the production machine assembly can be predicted in real time and accurately, and the maintenance of the production machine assembly can be performed in time; and the maintenance can be performed immediately when the production machine assembly may die soon, And the possibility of the production machine components entering the dead state is presented in a digital way.

200: Process information

210: Data quality check

220: Convert to parameter index

222,224,226,228: curves

230: Algorithm library

240: DHI module

250: The first judgment step (DHI<0.7)

260: Applying multivariate RUL predictive modeling steps

270: Early warning mode

280: DCI mode

300: Select modeling sample data

302: Does the variance of the values in the modeling sample increase over time?

304: Perform logarithmic transformation on the numerical values of the modeling sample data

306: Whether the value of the modeling sample data is a steady state

308: Perform differential conversion on the numerical value of the modeling sample data

309-1: Perform Granger Causality Tests

309-2: Remove Auxiliary Aging Features from Group Modeling Sample Data (z _T )

310:Using PACF to select the maximum backward period of VAR( p ) model

312: Perform white noise verification on the numerical values of the modeling sample data

314: Add the value of another aging feature to the modeling sample data

316: Establish VAR model combination

318: Calculate the amount of information of the VAR model combination

320: Select the best VAR model

324:Removing insignificant predictive components in the best VAR model

326: Test the residual of the best VAR model

328:Confirm the best VAR model

400: Whether the RUL _t drop is greater than or equal to the threshold

410: Whether RUL _t is less than the maintenance buffer time

420: Whether RUL _t is less than the maintenance buffer time

510,520: predicted value

530: actual value

540: death specification value

550: sick specification value

560,570:RUL

580:DCI Threshold

DCI: Death Correlation Index

RUL: remaining useful life

TSP: Time Series Forecasting Algorithm

TSP _MVA : Multivariate Time Series Forecasting Algorithm

For a more complete understanding of the embodiments and advantages thereof, reference is now made to the following description taken in conjunction with the accompanying drawings, in which

[FIG. 1A] is a schematic block diagram illustrating a multi-variable predictive maintenance method for production machine components according to some embodiments of the present invention;

[FIG. 1B] is a diagram illustrating the relationship between event indication value and parameter index (average opening) according to some embodiments of the present invention;

[FIG. 1C] is a diagram illustrating the relationship between an event indication value and a parameter index (1/4 octave frequency) according to some embodiments of the present invention;

[FIG. 2A] and [FIG. 2B] are flow diagrams illustrating the steps of multivariate modeling according to some embodiments of the present invention;

[FIG. 3A] is a schematic diagram showing the partial autocorrelation function of the aging characteristics of an embodiment of the present invention;

[FIG. 3B] is a schematic diagram showing the partial autocorrelation function of the assisted aging feature according to an embodiment of the present invention;

[FIG. 4] is a schematic block diagram illustrating an early warning mode according to some embodiments of the present invention;

[Figure 5] shows the prediction results of the axial deflection of the component;

[Figure 6] shows the prediction results of the RUL of the components;

[Figure 7] shows the prediction result of DCI(TSP), one of the components; and

[Fig. 8] shows the prediction result of another DCI (TSP _MVA ) of the module.

Embodiments of the invention are discussed in detail below. It should be appreciated, however, that the embodiments provide many applicable inventive concepts that can be implemented in various in a wide variety of specific content. The specific embodiments discussed are illustrative only and do not limit the scope of the invention.

Because a single variable cannot represent the full picture of machine aging information, in order to solve the problem of single variable inaccuracy, the embodiment of the present invention proposes a multivariate time series prediction algorithm (TSP _MVA ), and uses the application information criterion algorithm to establish the most The best vector autoregression model (Vector AutoRegression Model; VAR) to predict the complex future trend of unexpected downtime of production machine components. In addition, the known time series prediction (Time Series Prediction; TSP) technology only uses one variable to make predictions, and the prediction accuracy is limited, while the multivariate time series prediction of the present invention expands the limitation that only one variable can be used, The predicted aging characteristics depend not only on past aging characteristics, but also on other characteristics (auxiliary aging characteristics), so the use of multiple variables can improve the prediction accuracy. In addition, the embodiment of the present invention also proposes an early warning mechanism, so that maintenance can be performed immediately when the production machine components may die soon, and a death correlation index (DCI) is provided to present the possibility of the production machine components entering death with data .

Please refer to FIG. 1A . FIG. 1A is a schematic block diagram illustrating a multi-variable predictive maintenance method for production machine components according to some embodiments of the present invention. Firstly, multiple groups of process data 200 used when the components of the production machine process multiple workpieces in sequence are obtained, wherein each group of process data 200 includes the values of multiple parameters. A production machine is a machine that processes workpieces on a production line. The production line can be, for example: semiconductor production line, TFT-LCD production line, tool machining production line, etc.; the workpiece can be, for example: wafer, glass substrate, wheel frame, screw, etc.; the machine can be, for example: thin film deposition machine, positive photoresist coating Laying machine, exposure machine, developing machine, etching machine, photoresist removal machine, process Tools, etc.; components can be, for example: heaters, pressure modules, throttle valves, oil-free bushings or bearings; parameters can be, for example: valve opening, vibration amplitude, driving voltage, driving current, temperature and/or pressure wait. The above description is only for illustration, so the embodiments of the present invention are not limited thereto. It is worth mentioning that the value of each parameter in each set of process data 200 is a time-series data value in a period of workpiece processing time, that is, a curve of parameter value versus time.

Next, a data quality check is performed on the process data 200 (step 210 ) to confirm whether the quality of the process data 200 is good. If the quality of the process data 200 is not good, it is necessary to obtain the process data used when the components of the production machine process other workpieces sequentially. Step 210 may adopt a process data quality assessment method similar to that used in US Patent No. 8095484B2. The embodiments of the present invention refer to the relevant provisions of the US Patent No. 8095484B2 (Incorporated by reference).

Then, use a plurality of algorithms in an algorithm library 230 to convert the value of each parameter of each set of process data into values of a plurality of parameter indicators (step 220 ). These parameter indicators include: k multiplied frequency after conversion to the frequency domain (where k is greater than 0), the global similarity index (Global Similarity Index; GSI), the kurtosis of the statistical data distribution, and the skewness of the statistical data distribution (skewness), standard deviation (STD), root mean square (root mean square; RMS), average value (Average), maximum value (Max) and minimum value (Min), and the conversion method of these parameter indicators can adopt a The moving window method (Moving Window; MW) is used to determine the number of samples.

For example: each set of process data 200 has a set of time-series data of valve opening and a set of time-series data of vibration amplitude. Statistical Process Control (SPC) in algorithm library 230 can use this The time-series data of the group valve openings are converted into a kurtosis, a skewness and a standard deviation; the Global Similarity Index (GSI) algorithm in the algorithm library 230 can convert the time-series data of the group valve openings into a GSI value; the time-frequency conversion algorithm in the algorithm library 230 can convert the time series data of this group of vibration amplitudes into parameters such as a 1/4 multiplier, a 1/2 multiplier, a 2 multiplier, and a 4th multiplier The numeric value of the indicator. For the GSI algorithm, reference may be made to US Patent No. 8095484B2. The above-mentioned algorithms in the algorithm library 230 are only for illustration, so the embodiments of the present invention are not limited thereto.

On the other hand, the method obtains a plurality of event indication values respectively corresponding to the sets of process data, wherein the event indication values indicate whether an abnormal event occurs in the production machine assembly when the production machine assembly is processing each workpiece. Please refer to FIG. 1B. FIG. 1B is a diagram illustrating the relationship between an event indication value and a parameter index (average opening) according to some embodiments of the present invention, wherein the curve 222 indicates the average opening of the production machine assembly when processing each workpiece. Degree; Curve 224 indicates whether abnormal events occur when the production machine assembly processes each workpiece. As shown in Figure 1B, before the production machine component processes the 289th workpiece, no abnormal event occurs, and its event indication value can be, for example, "0"; after the production machine component processes the 289th workpiece, there is an abnormal event Occurs, its event indication value can be, for example, "1". Please refer to FIG. 1C. FIG. 1C is a diagram illustrating the relationship between an event indication value and a parameter index (1/4 multiplier frequency) according to some embodiments of the present invention, wherein the curve 226 indicates that the production machine assembly processes each workpiece when The vibration amplitude value of 1/4 octave frequency; the curve 228 indicates whether any abnormal event occurs when the production machine assembly processes each workpiece. As shown in Figure 1C, no abnormal events occurred before the production machine components processed the 287th workpiece, and the event indicated The indication value can be, for example, "0"; after the production machine assembly processes the 287th workpiece, an abnormal event occurs, and the event indication value can be, for example, "1".

Then, a correlation analysis is performed between each parameter index and event indication value in these sets of process data, and a plurality of correlation coefficients respectively corresponding to these parameter indexes are obtained, as shown in Table 1. Next, select the parameter index corresponding to the largest of these correlation coefficients as an aging characteristic (y _T ), as shown in Table 1. In addition, an auxiliary aging feature can also be selected, and this auxiliary aging feature can be combined with the aforementioned aging feature (y _T ) to predict the aging feature at the next time point. Since the aging feature (y _T ) and the auxiliary aging feature are merged into a vector, the method proposed here can also be called a vector autoregression model (Vector AutoRegression Model; VAR). The auxiliary aging feature can be any other feature (such as axis deflection) in Table 1 except the opening degree. In some embodiments, the auxiliary aging feature can be selected artificially, for example, temperature can be selected as the auxiliary aging feature. In some embodiments, any feature selection method, such as Adaptive Boosting, can also be used, but the present disclosure is not limited thereto.

Next, the first judgment step 250 is carried out to determine whether the production machine assembly is in a sick state when processing these workpieces according to the value of the aging characteristic corresponding to each workpiece, wherein once the production machine assembly is in a workpiece When is in the sick state, this artifact is set as a sample selection point (t). The following example illustrates an implementation manner of the first determining step, but the embodiment of the present invention is not limited thereto. As shown in FIG. 1A, in the first judging step, at first, the values of the aging characteristics (y _T ) of these groups of process data are input to the device health index (Device Health Index; DHI) module 240, and the DHI module 240 uses A set of conversion formulas respectively converts the value of the aging characteristic (y _T ) of each set of process data into a plurality of device health indices (DHI) respectively corresponding to workpieces. Then, determine whether these device health indexes are greater than or equal to a threshold value (for example: 0.7) in sequence, and set the corresponding artifacts of the first one of these device health indexes greater than or equal to a threshold value as sample selection point. The conversion formula for this set is:

； when

;

； when

;

； when

;

； when

;

； when

;

； when

;

為此些組製程資料之此老化特徵的數值的平均值， 其中

為

所對應之轉換值； in

is the average value of the values of this aging characteristic for these sets of process data, where

for

The corresponding conversion value;

Max y _T is the maximum value of this aging characteristic of these groups of process data, and Max y _{T_ mapping} is the conversion value corresponding to Max y _T ;

Min y _T is the minimum value of this aging characteristic of these groups of process data, and Min y _{T_ mapping} is the conversion value corresponding to Min y _T ;

LSL is the lower specification limit; LCL is the lower control limit; UCL is the upper control limit; USL is the upper specification limit; _{LSL_mapping} is the conversion value corresponding to LSL; _{LCL_mapping} is the conversion value corresponding to LCL; _{UCL_mapping} is the conversion value corresponding to UCL Conversion value; _{USL_mapping} is the conversion value corresponding to USL. The DHI algorithm is similar to US Patent No. 10,242,319B2. The embodiments of the present invention refer to relevant provisions of the US Patent No. 10,242,319B2 (Incorporated by reference).

Next, a multivariate modeling step 260 is performed. First, use the N values corresponding to this aging feature (y _T ) and the N values corresponding to the auxiliary aging features in the N sets of process data corresponding to the N workpieces before the sample selection point (t) as a set of modeling sample data , where N is a positive integer. Then, a Granger causality test (Granger causality test) is performed on the aging feature (y _T ) and the auxiliary aging feature to judge the correlation between the aging feature (y _T ) and the auxiliary aging feature. Delete auxiliary aging features in modeling sample data. For example: suppose there is one aging feature (such as SX1), and there are two auxiliary aging features (such as SY1, SZ1) to be selected later, and the auxiliary aging feature SY1 is related to the aging feature (y _T ), then it is retained; the auxiliary aging feature SZ1 If it is irrelevant to the aging feature (y _T ), it is deleted, so that this set of modeling sample data corresponds to the aging feature SX1 and the auxiliary aging feature SY1 (as shown in FIGS. 3A and 3B ). In one embodiment, the aging feature SX1 and the auxiliary aging features SY1 and SZ1 may be axial deflection, which is the angle at which the axis of the workpiece (such as a bearing) deviates. The aging feature SX1 is the angle that the axis of the workpiece deviates from the X plane, the auxiliary aging feature SY1 is the angle that the axis of the workpiece deviates from the Y plane, and the auxiliary aging feature SZ1 is the angle that the axis of the workpiece deviates from the Z plane, but the present invention It is not limited to this; then, use this group of modeling sample data and establish an aging characteristic prediction model according to a multivariate time series prediction algorithm, and obtain the aging characteristics (y _T ) arranged according to the production sequence of the workpieces Plural number of predicted values. These predicted values are then converted into a plurality of remaining useful lives (Remaining _useful life; RUL) predicted value (RUL _t ), where t represents the t-th workpiece, and t is an integer. It is worth mentioning that the reason why the present invention uses the multivariate time series forecasting algorithm and the multivariate modeling step 260 is that the conventional technology only uses one variable for forecasting, and its forecasting accuracy is limited. However, the multivariate time series prediction of the present invention expands the restriction that only one variable can be used, and the predicted aging characteristics not only depend on past aging characteristics, but also depend on other characteristics (i.e. auxiliary aging characteristics), so multiple variables are used to improve prediction accuracy. The multivariate time series forecasting algorithm and the detailed multivariate modeling step 260 will be described later.

Then a second judging step is performed to judge whether the components of the production machine need to be replaced or repaired according to the remaining service life prediction values (RUL _t ). In some embodiments, the second determination step includes an early warning mode 270 and a DCI mode 280 . The warning mode 270 and the DCI mode 280 will be described later.

Please refer to FIG. 2A and FIG. 2B . FIG. 2A and FIG. 2B are schematic flowcharts illustrating the multivariate modeling step 260 according to some embodiments of the present invention. First, step 300 _is performed to select N values ( Aging characteristic actual value) and N values corresponding to auxiliary aging characteristics as a set of modeling sample data (Y _M ), where N is a positive integer, such as 30, that is, the first 30 workpieces with DHI>0.7, modeling The aging features in the sample data are expressed as Y _M ={ y _{t -30} , y _{t -29} ,..., y _{t -2} , y _{t -1} }, while the auxiliary aging features are expressed as Z _M ={ z _{t - 30} , z _{t -29} ,..., z _{t -2} , z _{t -1} }.

Next, step 302 is performed to determine whether the variance of the value of the aging feature in the group of modeling samples will increase with time. In other words, if y _t =(1+ α )× y _{t - 1} , where α is greater than 0, which means that y _t increases with time and Var( y _t ) grows with α , go to step 304 , otherwise go to step 306 . In step 304, logarithmic transformation is performed on each value of the set of modeling sample data to force the distribution of the increase rate of the data to have a certain degree of regularity, and then step 306 is performed.

In step 306, a single root test (unit root test) is performed on the values in this group of modeling samples to confirm that the values in this group of modeling samples are sequentially Whether the arrayed values are in a steady state, wherein when these values are not in a steady state, differential conversion is performed on each value of the set of modeling sample data (step 308 ). Algorithms for confirming whether the sequence is in a steady state include Augmented Dickey-Fuller test (Augmented Dickey-Fuller test; ADF) test or Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test. The formulas and application methods of the ADF test and the KPSS test are known to those skilled in the art, so they will not be repeated here.

When the values arranged in sequence are not in a steady state (the result of step 306 is no), proceed to step 308 to carry out differential conversion to each value of the set of modeling sample data, so that the values arranged in sequence ( timing) reaches a steady state. The formula for differential conversion is: ▽ ^d y _{t - i} = y _{t - i} - y _{t - i -1} , ▽ ^d z _{t - i} = z _{t - i} - z _{t - i -1} . When the values arranged in sequence are in the steady state (the result of _step 306 is Yes), proceed to step 309-1 _, and execute the Granger causality test (Granger causality test) to judge the correlation between the aging feature (y _T ) and the auxiliary aging feature (z _T ), if not, delete the auxiliary aging feature (z _T ) from the group modeling sample data in step 309-2, Auxiliary aging features (z _T ) are preserved if relevant. The Granger causality test can be understood by those skilled in the art, and will not be described in detail here.

Next, step 310 is performed to select the maximum backward period p of the vector autoregressive (VAR( p )) model by using a partial autocorrelation function (PACF). Its formula is:

B = arg max ( ρρ _k ). (1);

γ _k = cov ( y _t , y _{t - k} ) = E ( y _t - μ )( y _{t - k} - μ ) (3-1);

ρρ _k = Corr ( y _t , y _{t - k} | y _{t -1} , y _{t -2} ,..., y _{t - k +1} ) (3-2);

Where B is the time of the maximum PACF (workpiece number) related to y _{t - 1} ; Var ( y _t ) = Var ( y _{t + k} ) = E( y _t - μ ) ² = γ ₀ .; E [. ] is the expected value function; μ is the mean value (mean) of y _t . For example, refer to FIG. 3A and FIG. 3B , wherein FIG. 3A is a schematic diagram illustrating the partial autocorrelation function of the aging feature SX1 of an embodiment of the present invention; FIG. 3B is a schematic diagram of the auxiliary aging feature SY1 of an embodiment of the present invention. Schematic representation of the partial autocorrelation function. The aging feature SX1 and the auxiliary aging feature SY1 are axis skewness. It can be seen from Figure 3A and Figure 3B that the maximum backward period p of the vector autoregressive (VAR( p )) model selected by the aging feature SX1 through PACF is 3, and the vector autoregressive (VAR( p ) model selected by the auxiliary aging feature SY1 by PACF )) The maximum number of backward periods p for the model is 3. Thus, the present invention can determine the maximum backward period p of the VAR( p ) model through PACF.

Next, proceed to step 312 to perform a white noise test on each value (the value of the aging feature and the value of the auxiliary aging feature) in the set of modeling sample data. The algorithm of the white noise test includes Ljung-Box test and the like. The main purpose of step 312 is to confirm whether the time sequence is white noise (that is, states are not correlated with each other). When these values are white noise, this group of modeling sample data needs a little more data, so add the value corresponding to the aging feature in another set of process data corresponding to the N+1 workpieces before the sample selection point (aging feature actual value, that is, y _{t - (N + 1)} or y _{t - 31} ) and the value of the auxiliary aging feature (the actual value of the auxiliary aging feature, that is, z _{t - (N + 1)} or z _{t - 31} ) to this group of modeling samples data (step 314).

Then, go to step 316, establish a plurality of VAR( p ) model combinations, and determine the number of VAR( p ) model combinations according to the maximum backward period p of the VAR( p ) model. For example: when the maximum backward period p of the aforementioned VAR( p ) model is 3, three model combinations are established: VAR(1), VAR(2), and VAR(3). VAR(1), VAR(2), and VAR(3) are VAR( p ) where p is equal to 1, 2, and 3, respectively. The VAR( p ) model will be described below.

The VAR( p ) model is defined as:

為在時間點t(第t個工件)的老化特徵預測值；b為常數向量；φ _i 為在時間點(第t個工件)之VAR(p)模型的最小平方預估係數，i=1,2,...,p；y _t-i 為在時間點t-i(第t-i個工件)的老化特徵實際值(對應圖3A之老化特徵SX1)；z _t-i 為在時間點t-i(第t-i個工件)的輔助老化特徵實際值(對應圖3B之輔助老化特徵SY1)；β _i 為所要訓練(尋找)的係數；ε _t 為在時間點t(第t個工件)的白噪音項，亦即誤差向量。如果在步驟309-2中刪除了輔助老化特徵，則在式子(4)中的係數β _i 都設為0。 in

is the predicted value of aging characteristics at time point t (the t- th workpiece); b is a constant vector; φ _i is the least square prediction coefficient of the VAR( p ) model at the time point ( t- th workpiece), i =1 ,2,..., p ; y _{t - i} is the actual value of aging characteristics at time point t - i (the t - ith workpiece) (corresponding to the aging characteristics SX1 in Figure 3A); z _{t - i} is the aging characteristic at time The actual value of the auxiliary aging feature at point t - i (the t - ith workpiece ₎ ( corresponding to the auxiliary aging feature SY1 in Figure 3B); β _i is the coefficient to be trained (searched); artifacts), that is, the error vector. If the auxiliary aging feature is deleted in step 309-2, the coefficients β _i in the formula (4) are all set to 0.

Then, proceed to step 318 to use an information criterion algorithm to calculate the information amount of each VAR( p ) model combination, wherein the information criterion algorithm is Bayesian Information Criteria (BIC). The formula of the BIC algorithm is as follows:

Where SSE is the sum of squared errors; M is the size of the modeling sample data.

Table 2 is an example for calculating the information volume of each VAR( p ) model combination VAR(1), VAR(2), and VAR(3) using the BIC algorithm.

Next, proceed to step 320 to select one of the VAR( p ) model combinations VAR(1), VAR(2), and VAR(3) with the largest amount of information (i.e. the smallest BIC) as an optimal model, for example : var(1).

Then, step 322 is performed to remove insignificant predictors (parameters) in the best model. When the prediction coefficient of the predictor is greater than the 95% confidence interval, the predictor is not significant and should be removed. Under the assumption of normal distribution, the 95% confidence interval is equal to 1.96, and the judgment formula of the insignificant predictor (parameter) in the best model is as follows:

| φ _i |>1.96× s . e .( φ _i ) (7);

| β _i |>1.96× s . e .( β _i ) (8);

Where i =1,2,..., p ; s . e .(.) is the standard deviation of the coefficient.

Next, step 324 is performed to re-estimate the weights for the VAR( p ) model after removing the significant predictors (parameters).

Then, go to step 326 to test the residual error of the re-estimated model. The algorithm for this test includes Ljung-Box test and the like. When the residuals of the re-estimated model have been explained, it is confirmed that the re-estimated model is the best model (step 328 ), and a plurality of predicted values of the aging characteristics (y _T ) arranged according to the production order of the workpieces are obtained accordingly. Next, step 330 is performed to convert these predicted values by using the workpiece processing time ( dt ) for processing each workpiece of the production tool assembly and the death specification value of the aging characteristic (y _T ) of the production tool assembly when it cannot be used. The complex remaining useful life (Remaining useful life; RUL) prediction value ( RUL _t ), the formula is RUL _t = k _D - k _t ., where t represents the t -th workpiece; k _t represents the t -th workpiece The corresponding time point (ie the 7th × dt ), where t is an integer k _D represents the time point corresponding to the death standard value of the aging characteristic (y _T ).

After the RUL predicted value ( RUL _t ) is obtained, a second judgment step is performed to judge whether the components of the production machine need to be replaced or repaired according to the RUL _t . As shown in FIG. 1A , in some embodiments, the second determination step includes an early warning mode 270 and a DCI mode 280 .

0.3是否成立？當步驟400的結果為是時，在第一階中進行步驟410或420，以判斷RUL _t 是否小於一維修緩衝時間(Buffer Time；BT)，而獲得一第二結果，其中BT係由生產機台組件的原廠提供，當生產機台組件異常時，必須在此維修緩衝時間(BT)對生產機台組件進行維修或更換。當第一結果和第二結果均為否時，生產機台組件處於生病狀態但未急速惡化，不需進行維修，而顯示例如綠燈。當第一結果為否而第二結果為是時，生產機台組件未急速惡化但其剩餘使用壽命 不足，需進行維修，而顯示例如藍燈。當第一結果為是而第二結果為否時，生產機台組件急速惡化，而顯示例如褐燈。若生產機台組件處理連續第t+i個工件之每一者均顯示例如褐燈時，則需檢查或維修生產機台組件，其中i為正整數。當第一結果和第二結果均為是時，生產機台組件需進行維修，而顯示例如紅燈。 When the predicted value of RUL ( RUL _t ) drops sharply or oscillates close to the state of death, it is difficult for the user to judge whether the components of the production machine need to be replaced or repaired. Therefore, the embodiment of the present invention proposes an early warning mode to solve this problem. Please refer to FIG. 4 , which is a schematic block diagram illustrating an early warning mode according to some embodiments of the present invention. First, step 400 is carried out in the first stage to determine whether the current RUL _t is lower than the previous RUL _{t - 1} and is greater than or equal to a threshold value (for example, 30%), namely ( RUL _{t -1} - RUL _t ) / RUL _{t -1}

Does 0.3 hold? When the result of step 400 is yes, carry out step 410 or 420 in the first stage, to judge whether RUL _t is less than a maintenance buffer time (Buffer Time; BT ), and obtain a second result, wherein BT is by the production machine The original factory provides the components of the machine. When the components of the production machine are abnormal, the components of the production machine must be repaired or replaced within this maintenance buffer time ( BT ). When both the first result and the second result are negative, the production machine components are in a sick state but not deteriorating rapidly, and maintenance is not required, and a green light is displayed, for example. When the first result is no and the second result is yes, the production tool components have not deteriorated rapidly but their remaining service life is not enough, maintenance is required, and a blue light is displayed, for example. When the first result is yes and the second result is no, the production tool assembly deteriorates rapidly, displaying, for example, a brown light. If each of the t+i th consecutive workpieces processed by the production machine assembly displays, for example, a brown light, the production machine assembly needs to be inspected or repaired, wherein i is a positive integer. When both the first result and the second result are yes, the production machine components need to be repaired, and a red light is displayed, for example.

The embodiment of the present invention will be described below by using the axial skewness of the component (the aging characteristic SX1 and the auxiliary aging characteristic SY1 ), the RUL prediction result and the DCI prediction result. Axis skewness represents the angle of deviation of the axis of the component, aging feature SX1 represents the angle of deviation of the axis of the component on the X plane, and auxiliary aging feature SY1 represents the angle of deviation of the axis of the component on the Y plane, wherein the auxiliary aging feature SY1 is based on the aforementioned The result of the Granger causality test in the multivariate modeling step 260 . Please refer to Figures 5, 6, 7, and 8, where Figure 5 shows the predicted results of the axial deflection of the component; Figure 6 shows the predicted results of the RUL of the component; Figure 7 shows the DCI (TSP) of one of the components The prediction results of ; FIG. 8 shows the prediction results of another DCI (TSP _MVA ) of the component. Among them, TSP represents the conventional univariate time series prediction algorithm (only considers a single aging characteristic); TSP _MVA represents the multivariate time series prediction algorithm of the present invention (considers multiple aging characteristics). In Fig. 5, the horizontal axis is time, the vertical axis is the shaft deflection, the curve 510 is the predicted value of the shaft deflection of the component obtained using the TSP _MVA of the present invention; the curve 520 is the shaft of the component obtained using the conventional TSP Predicted value of skewness; point group 530 (composed of "*") is the actual value; straight line 540 is the dead specification value of the axial skewness of the component; straight line 550 is the sick specification value of the axial skewness of the component. In FIG. 6 , the horizontal axis is the workpiece number, and the vertical axis is RUL (number of days). Curves 560 and 570 are the RUL calculated using the conventional TSP and the TSP _MVA of the present invention, respectively. The curve 560 has an early warning for the 106th workpiece in the workpiece number, and the curve 570 has an early warning for the 119th workpiece in the workpiece number. In FIG. 7 and FIG. 8 , the straight line 580 is the DCI threshold ( _DCIT ). When the DCI is less than the DCI threshold value, no maintenance is required; when the DCI is greater than the DCI threshold value, it means that the component is close to a dead state when processing workpieces, and maintenance is required. FIG. 7 shows the DCI calculated by the conventional TSP, and FIG. 8 shows the DCI calculated by the TSP _MVA of the present invention. The conventional TSP issues an early warning 24 days before the time of death, but the TSP _MVA of the present invention issues an early warning 11 days before the time of death. From the comparison of the two, it can be seen that the conventional TSP issues an early warning, so the RUL evaluated by the TSP _MVA of the present invention is relatively accurate.

It can be understood that the multi-variable predictive maintenance method of the production machine components of the present invention is the above-mentioned implementation steps, and the computer program product stored in the present invention for measurement and sampling is used to complete the above-mentioned measurement Sampling method. The order of the implementation steps described in the above embodiments can be adjusted, combined or omitted according to actual needs. The above-mentioned embodiments can be realized by using a computer program product, which can include a machine-readable medium storing a plurality of instructions, and these instructions can program a computer to perform the steps in the above-mentioned embodiments. A machine-readable medium may be, but is not limited to, floppy disk, compact disk, compact disk, magneto-optical disk, read-only memory, random access memory, erasable programmable read-only memory (EPROM), electronically erasable Excluding programmable read-only memory (EEPROM), optical or magnetic cards, flash memory, or any machine-readable medium suitable for storing electronic instructions. Moreover, the embodiment of the present invention can also be downloaded as a computer program product, which can be downloaded by using a communication link The computer program product of the present invention is transferred from the remote computer to the requesting computer by means of a data signal connected (such as a connection such as a network connection).

It should also be noted that the invention may also be described in the context of a manufacturing system. Although the present invention can be implemented in semiconductor fabrication, the present invention is not limited to semiconductor fabrication, but can be applied to other manufacturing industries as well. A manufacturing system is configured to manufacture workpieces or products including but not limited to microprocessors, memory devices, digital signal processors, application specific circuits (ASICs) or other similar devices. The present invention can also be applied to other workpieces or products other than semiconductor devices, such as vehicle wheel frames and screws. A manufacturing system includes one or more processing tools that can be used to form one or more products, or a portion of a product, on or in a workpiece (eg, wafer, glass substrate). Those skilled in the art will appreciate that the processing tools can be any number and any type, including lithography tools, deposition tools, etching tools, grinding tools, annealing tools, tool tools and the like. In an embodiment, the fabrication system also includes a scatterometer, an ellipsometer, a scanning electron microscope, and the like.

To sum up, the embodiment of the present invention can use the aging feature and the auxiliary aging feature to predict the aging feature, which means using multiple variables to predict one variable. Conventional technology only uses one variable to make predictions, and the prediction accuracy is limited. However, the multivariate time series prediction of the present invention expands the limitation of using only one variable. The predicted aging characteristics not only depend on the past aging characteristics, Also depends on other features (auxiliary aging features), so using multivariate can improve prediction accuracy. The embodiment of the present invention can also immediately and accurately predict the RUL of the production machine components, so that the maintenance of the production machine components can be carried out in time; and the maintenance can be performed immediately when the production machine components may die soon. protection, and present the possibility of production machine components entering a dead state in a digital way.

Although the present invention has been disclosed above in terms of implementation, it is not intended to limit the present invention. Anyone skilled in this art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be defined by the appended patent application scope.

200: Process information

210: Data quality check

220: Convert to parameter index

230: Algorithm library

240: DHI module

250: The first judgment step (DHI<0.7)

260: Applying multivariate RUL predictive modeling steps

270: Early warning mode

280: DCI mode

Claims (9)

A multi-variable predictive maintenance method for production machine components, comprising: obtaining multiple sets of process data used or generated when a component of a production machine processes a plurality of workpieces sequentially, wherein each set of process data includes a plurality of The value of the parameter, the value of each of the parameters of each of the sets of process data is a set of time-series data values during a period of processing time when the component of the production machine is processing one of the workpieces, these A set of process data is corresponding to these workpieces in a one-to-one manner; according to whether an abnormal event occurs in the component when processing each of the workpieces, it is determined that a plurality of sets of process data are corresponding to the groups of process data in a one-to-one manner. An event indicator value, wherein the event indicator values indicate whether an abnormal event occurs in the component when the component is processing each of the workpieces; a plurality of algorithms are respectively used to convert each of the parameters of each of the groups of process data The set of time-series data values are converted into values of a plurality of parameter indexes, wherein the parameter indexes correspond to the algorithms in a one-to-one manner; for each of the parameters in the sets of process data Carrying out a correlation analysis between parameter indexes and these event indicator values, and obtaining a plurality of correlation coefficients corresponding to these parameter indexes in a one-to-one manner; selecting a parameter index corresponding to a maximum among these correlation coefficients as a Aging feature, and an auxiliary aging feature is set, the auxiliary aging feature corresponds to any parameter index other than the parameter index of the largest among the correlation coefficients; a first judgment step is carried out, so as to correspond to each of the workpieces The value of the aging characteristic is used to determine whether the component is in a sick state when processing the workpieces, wherein once the component is in the sick state when processing one of the workpieces, the one of the workpieces is Set as a sample selection point; perform a multivariate modeling step, the multivariate modeling step includes: using the N values corresponding to the aging feature in the N groups of process data corresponding to the N workpieces before the sample selection point and The N values corresponding to the auxiliary aging feature are a set of modeling sample data, where N is a positive integer; a Granger causality test (Granger causality test) is performed on the aging feature and the auxiliary aging feature to determine the aging feature Correlation with the auxiliary aging feature, if not relevant, delete the auxiliary aging feature from the set of modeling sample data; to use the set of modeling sample data and according to a multivariate time series forecasting algorithm to establish a an aging characteristic prediction model to obtain a plurality of predicted values of the aging characteristic in order of workpiece production; and using a workpiece processing time for each of the workpieces and a death specification value of the aging characteristic when the component becomes unusable, To convert these predicted values into a plurality of remaining useful life (Remaining useful life; RUL) predicted values ( RUL _t ), wherein t represents the t-th workpiece, and t is an integer; and a second judgment step is carried out, based on These residual service life prediction values ( RUL _t ) are used to determine whether the component needs to be replaced or repaired; wherein, the multivariate modeling step further includes: using a Vector AutoRegression Model (Vector AutoRegression Model; VAR) for the multivariate time Sequence prediction algorithm to establish the aging characteristic prediction model; use a partial autocorrelation function (partial autocorrelation function; PACF) to select the maximum number of backward periods of the vector autoregressive model; the values in the group of modeling sample data Carry out a white noise test, and when the values are white noise, add the values corresponding to the aging characteristics in another set of process data corresponding to the N+1 workpieces before the sample selection point to the set of modeling samples data; use the maximum number of backward periods of the vector autoregressive model to establish a plurality of vector autoregressive model combinations; use an information criterion algorithm to calculate the information amount of each of the vector autoregressive model combinations; and select the One of the vector autoregressive model combinations with the largest amount of information is an optimal model; wherein, the parameters include an axis skewness. The multi-variable predictive maintenance method for production machine components as described in claim 1, wherein the first judgment step includes: using a set of conversion formulas to convert the values of the aging characteristics of each of the sets of process data into corresponding To the multiple Device Health Index (DHI) of these workpieces, the conversion formula of this group is: when

< y _T < UCL,

_; When UCL< yT <USL,

; When USL< y _T ,

; when LCL < y _T <

,

_; When LSL< yT <LCL,

; When Min y _T < y _T <LSL,

;in

is the average value of the aging characteristics of these sets of process data, where

for

The corresponding conversion value; Max y _T is the maximum value of the aging characteristics of these groups of process data, Max y _{T_ mapping} is the conversion value corresponding to Max y _T ; Min y _T is the aging characteristics of these groups of process data Min y _{T_ mapping} is the conversion value corresponding to Min y _T ; LSL is the lower specification limit; LCL is the lower control limit; UCL is the upper control limit; USL is the upper specification limit; _{LSL_mapping} is the conversion value corresponding to LSL; LCL _{_ mapping} is the conversion value corresponding to LCL; UCL _{_ mapping} is the conversion value corresponding to UCL; USL _{_ mapping} is the conversion value corresponding to USL; and sequentially determine whether the device health index is greater than or equal to a threshold value, and The artifact corresponding to the device health index that is greater than or equal to one of the threshold values first is set as the sample selection point. The multivariable predictive maintenance method for production machine components as described in Claim 1, wherein the information criterion algorithm is Bayesian Information Criteria (BIC). The multivariate predictive maintenance method for production machine components as described in claim item 1, wherein the multivariate modeling step further includes: judging whether the variation of the values in the group of modeling samples will increase over time Larger, wherein when the variation of these values is larger and larger over time, logarithmic transformation is performed on each of these values of the set of modeling sample data; and a single root test (unit root test) to confirm whether the numerical values arranged in sequence are in a steady state, wherein when the numerical values are not in a steady state, differential conversion is performed on each of the numerical values of the set of modeling sample data. The multi-variable predictive maintenance method for production machine components as described in claim 1, wherein the second judging step includes: judging whether ( RUL _t - RUL _t-1 )/ RUL _t-1 is greater than or equal to a threshold value, and Obtain a first result, wherein t-1 represents the t-1th workpiece; judge whether RUL _t is less than a maintenance buffer time, and obtain a second result, wherein the component must be repaired within the maintenance buffer time; when the first When the first result and the second result are both negative, the component is in a sick state but not rapidly deteriorating, and maintenance is not required; when the first result is negative and the second result is yes, the component is not rapidly deteriorating but its Insufficient service life remaining, requiring maintenance; when the first result is yes and the second result is no, the component deteriorates rapidly if the first result is yes for each of the t+i th workpieces in a row If the second result is no, then the component needs to be inspected or maintained, wherein i is a positive integer; and when both the first result and the second result are yes, the component needs to be maintained. The multi-variable predictive maintenance method for production machine components as described in claim 1, wherein the second judgment step includes: using a set of conversion formulas to convert the values of the aging characteristics of each of the sets of process data into corresponding To a plurality of death correlation indices (Death Correlation Index; DCI) of the component of the artifacts, the conversion formula of the set is:

Among them, y _death is the value of the aging characteristic corresponding to the component in the dead state, y _t-1 is the value of the aging characteristic corresponding to the component when processing the t-1th workpiece, and conv is the calculation of covariance , Var is the calculation of variation; when the DCI _t is greater than a threshold value, it means that the component is close to a dead state when processing the t-th workpiece, wherein the calculation of the threshold value is based on the standard deviation of DCI _t . The multi-variable predictive maintenance method for production machine components as described in Claim 1, wherein the component is a heater, a pressure module, a throttle valve, an oil-free bushing or a bearing, and these parameters further include: A valve opening, a vibration amplitude, a driving voltage, a driving current, a temperature and a pressure. The multivariable predictive maintenance method for production machine components as described in claim item 1, wherein the parameter indicators include: a k multiplied frequency after conversion to the frequency domain (wherein k is greater than 0), an overall similarity index (Global Similarity Index; GSI), a statistical distribution of kurtosis (kurtosis), a statistical distribution of skewness (skewness), a standard deviation, a root mean square (root mean square), a mean, a maximum and a min. A computer program product, when the computer program product is loaded and executed, it can complete the multi-variable predictive maintenance method for production machine components as described in any one of claims 1 to 8.

Images