KR100997009B1 - The method for dynamic detection and on-time warning of industrial process - Google Patents

Abstract

PURPOSE: A dynamic monitoring about process margin of industrial device and timely alarming method are provided to realize the dynamic monitoring and timely alarming about a stop/protection signal of the industrial device according to the serial statistical learning and prediction model. CONSTITUTION: An outlier is removed from a learning data(S12). The leveling of learning data is performed(S13). The learning data is filter(S14). The cross correlation analysis of the learning data is performed(S15). The dynamic sampling window is applied to the learning data(S16). The correlation analysis and grouping is performed(S17). To make the prediction model for execution mode, a regression analysis is performed(S18).

Description

The method for dynamic detection and on-time warning of industrial process}

The present invention relates to a dynamic monitoring and timely alarm method for the process margin of an industrial facility. More specifically, the present invention relates to a dynamic monitoring and timely alarm for a stop / protection signal of an industrial facility based on a series of statistical learning and prediction models. The process of industrial equipment, which can provide the critical level for the current operating condition and provide timely alarm by providing the confidence interval considering the uncertainty of the process margin according to external conditions as well as the process margin. It relates to a dynamic monitoring and timely alarm method for margin.

Industrial equipment is a device consisting of a plurality of systems and devices to achieve a specific purpose, generally one or more instruments are installed for checking the operation and safety status, it consists of a configuration that can be measured offline or online.

In the above industrial equipment, the efficiency and safety of the equipment will vary depending on external conditions (air temperature, pressure, humidity, seawater or precipitation temperature if cooling water is required), characteristics of the injected fuel, deterioration of the equipment, operating range, etc. In terms of cost, the acceptable range of change in the efficiency and safety of the facility is called process margin. In most industrial installations, stop / protection functions for specific systems or devices are provided to prevent operation beyond this process margin. In order to implement such a stop / protection function, a control device is forcibly stopped when the value of a specific operating variable exceeds the above-mentioned stop / protection signal setting value.

The stop / protection setpoint and the process margin described above are mutually dependent variables. When the stop / protection setpoint is set too high, the process margin becomes relatively large and the cost benefit of operating the industrial equipment increases. In addition, there is a problem that serious accidents can occur and cause long-term facility shutdowns. On the contrary, when the above stop / protection set value is set too low, the probability of an accident is lowered, but the process margin is relatively low, and thus the industrial equipment is frequently shut down, thereby reducing the cost benefit of operating the industrial equipment. You lose.

Therefore, the overall process margin is determined by reflecting these two-sided properties. When the high degree of safety is required, it is necessary to set the process margin to a conservative value inclusive of all external conditions, input fuel, deterioration of equipment, and operating range. It is common.

However, it is very difficult to determine the overall process margin for a variety of situations, such as external conditions, input fuels, deterioration of equipment, operating ranges and the like.

On the other hand, it is common to have procedures in place to provide a preliminary stop / protection setpoint before the value of a particular operating variable approaches the stop / protection setpoint so that the operator can be prepared for the shutdown of the plant or take appropriate measures to normalize it. .

However, such a preliminary stop / protection set point is usually a static value, and once set, the set point is determined as a function of two conditions, as long as the value does not change, or the characteristic of the installation even when the value changes.

Therefore, the preliminary stop / protection set point does not provide information on the urgency of the trend of increasing or decreasing the stop / protection signal.

SUMMARY OF THE INVENTION An object of the present invention is to solve the conventional problems as described above, which can enable dynamic monitoring and timely alarms for stop / protection signals of industrial facilities based on a series of statistical learning and prediction models. It is to provide a dynamic monitoring and timely alarming method for the process margin of the facility.

Another object of the present invention is to provide an emergency level for the current operating state and to provide timely alarm by providing a confidence interval considering the uncertainty of the process margin according to external conditions as well as the process margin. To provide a dynamic monitoring and timely alarm method for process margin.

As a means for achieving the above object, the configuration of the present invention includes a 'learning mode' for performing statistical learning modeling and data to be used in 'execution mode', and data and modeling results prepared in the 'learning mode' and periodic It includes 'run mode' to perform dynamic monitoring and timely alarm on process margin using the operation data inputted into the system.

According to one aspect of the present invention, the learning mode includes the steps of: inputting learning data, removing outliers having high problems in statistical processing from the learning data, and averaging the learning data. And filtering the training data, performing cross correlation, applying a dynamic sampling window, performing correlation analysis and grouping, and using at run time. Performing regression analysis and quadruly verifying the training data for each group to create a predictive model, and finally finding the optimal smoothing coefficient, and ending the learning mode after completing the training data modeling. It is done by

According to one aspect of the present invention, the execution mode includes the steps of inputting driving data, removing a singular point having a high degree of problem in statistical processing from the driving data, taking an average value of driving data, Sampling operation data using a dynamic sampling window, calculating a process margin confidence interval by applying an autocorrelation regression model to signals classified by each group (S25); And combining the process margin confidence interval with the input signal (X _i ± PMCI _i ) and outputting the result to the user interface.

The invention can enable dynamic monitoring and timely alarms for stop / protection signals of industrial facilities, based on a series of statistical learning and prediction models, as well as process margins and uncertainty of process margins due to external conditions. Simultaneous provision of the confidence intervals taken into account has the effect of providing an emergency level for current operating conditions and generating timely alerts.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the accompanying drawings in order to describe in detail enough to enable those skilled in the art to easily carry out the present invention. . Other objects, features, and operational advantages, including the purpose, operation, and effect of the present invention will become more apparent from the description of the preferred embodiments.

For reference, the embodiments disclosed herein are only presented by selecting the most preferred embodiment in order to help those skilled in the art from the various possible examples, the technical spirit of the present invention is not necessarily limited or limited only by this embodiment Rather, various changes, additions, and changes are possible within the scope without departing from the spirit of the present invention, as well as other equivalent embodiments.

1 is an operation flowchart of a dynamic monitoring and timely alarming method for a process margin of an industrial facility according to an embodiment of the present invention.

As shown in Figure 1, the configuration of the dynamic monitoring and timely alarming method for the process margin of the industrial equipment according to an embodiment of the present invention, to perform statistical learning modeling and data to be used in the 'run mode' It includes a 'learning mode' and 'execution mode' to perform dynamic monitoring and timely alarm on process margins using the data and modeling results prepared in the 'learning mode' and the operation data periodically input.

The configuration of the learning mode includes the step of inputting the learning data (S11), the step of removing outliers having high problems in statistical processing from the learning data (S12), and averaging the learning data. Step (S13), filtering the training data (S14), performing cross-correlation (S15), applying a dynamic sampling window (Dynamic Sampling Window), (S16), Performing correlation analysis and grouping (S17), performing regression analysis and quadruple verifying training data for each group (S18) to create a prediction model for use in an execution mode, and an optimal smoothing coefficient (λ) Finally, after completing the training data modeling to find the step comprises the step (S19) of ending the learning mode.

The configuration of the execution mode includes the steps of inputting driving data (S21), removing a singular point having a high problem in statistical processing from the driving data (S22), and taking an average value of the driving data. (S23), step (S24) of sampling operation data using a dynamic sampling window (Dynamic Sampling Window), and process margin confidence interval by applying an autocorrelation regression model to the signals classified by each group Comprising a step (S25) and the combined process margin confidence interval (X _i ± PMCI _i ) to the input signal and outputs to the user interface (S26).

By the above configuration, the operation of the dynamic monitoring and timely alarming method for the process margin of the industrial equipment according to an embodiment of the present invention is as follows.

First, the learning mode will be described step by step as follows.

First, the training data is input (S11). The above learning data are obtained from the normal operating conditions of the industrial equipment without aging or deterioration of efficiency. Prepare operation data in various combinations, such as. All or some of the variables that can be measured on or off-line through the measuring equipment installed in the industrial facilities to be monitored are treated as learning data. There is no requirement for the time for collecting the learning data. However, in order to collect sufficient data under various conditions as described above, it is preferable to collect data for at least one year in a stable operation state after the first installation of the monitoring equipment. . If it is practically impossible to collect data for more than one year under stable operation, alternative methods can be used. This can be done, for example, by collecting data three months after a planned outage has been implemented over many years.

Next, outliers that are highly problematic in statistical processing are removed from the training data (S12). In general, depending on how the training data was collected, the method of eliminating highly problematic singularities in statistical processing will vary. If the training data was obtained from a database where basic signal processing is being performed, it is likely that outliers were removed from the database. However, if the training data is obtained directly from a database or sensor that is not performing enough signal processing, an appropriate singular point removal method should be used. Representative examples of singularity include problems caused by not inputting data at all, such as 'Bad Input', and cases where data is input, such as 'Out of Range', but temporarily larger or smaller than well beyond the normal range. If such data are generated, the reliability of the training data is increased by simultaneously removing the data of all variables acquired at that time under the following conditions.

if x ^L ≤x _i ≤x ^U then x _i OK, otherwise remove x _i

Where x _i is the i-th measured data value, x ^L is the normal lower limit of the variable x, and x ^U is the normal upper limit of the variable x.

Next, the training data is averaged (S13). The method proposed in the present invention aims to monitor a process change that changes in the minimum minutes or hours, rather than monitoring a rapid process change that changes in seconds or less. In a database installed in a monitoring system of an ordinary industrial equipment, data is often acquired in seconds or less. Therefore, in this case, the average value is obtained in units of 1 second, 5 seconds, or 10 seconds. In order to increase the reliability of the training data. The time interval for taking the average value takes into account the characteristics of the industrial installation. If, due to the nature of industrial equipment, a normal transition is completed within the time interval T, the maximum average time interval should not exceed T / 2.

Here, X _j is the j th value based on the sampling interval of the training data, x _i is the i th value based on the data collection interval, and N is the number of data collection intervals within the sampling interval of the training data.

Next, the training data is filtered (S14). If the average value can be taken from more than 10 sufficient data, no additional signal filtering (Signal Filtering or Signal Conditioning) is required. However, if the average value cannot be sufficiently obtained because the state can be changed at very short intervals due to the characteristics of the industrial equipment, or the acquisition period is too long due to the nature of the system acquiring the data, it is necessary to filter the signal. Do.

Signal filtering is a technique for removing noise by passing only a low frequency signal, and there are various methods of filtering a signal, and the present invention does not limit the use of a specific method. However, variables that are relatively slow in change, such as temperature and pressure, are the most variable in Group 1, and flows or water level are faster. Group 2, electrical signals, rotational speed, vibration, etc. Are divided into group 3 to vary the intensity of passing the low frequency signal. This means that only very low frequencies are passed through for group 1 to maximize noise rejection, while cutoff frequencies are set higher for group 2 so that information related to fast-changing processes is not lost, and block for group 3 Set the Cutoff Frequency to the highest value.

Subsequently, cross correlation is performed (S15). The cross-correlation analysis is performed by combining two different ones among all variables constituting the training data. By performing cross-correlation analysis, you can find pairs of variables that have a high correlation with time difference, and time difference can also be analyzed. The reason why two highly correlated variables appear with a time difference is not that all the process changes of the whole system are made instantaneously, but it takes time for the whole to be steady-state, taking place one by one. Because. Even in the case of variable pairs with high correlation, the correlation coefficient becomes smaller if the variation is made with time difference, so that an error may occur during the grouping process.

Where r _XY is the cross-correlation coefficient at the time delay d between the variables X and Y, Xi is the i-th value based on the sampling interval of the training data, Yi is the i-th value based on the sampling interval of the training data, and (where X and other variables), μ _X is the average of the variables X, μ _Y is the average of the variables Y, N is the number of times of the data collection interval in the sampling interval of the learning data, d is a time delay of 0, 1, 2, ..., N-1, and i is 0, 1, 2, ..., N-1.

Next, a dynamic sampling window is applied (S16). The time difference analyzed through cross-correlation analysis is solved using a sampling method to compensate for this. The concept of the dynamic sampling window is illustrated in FIG. 2. As shown in FIG. 2, the dynamic sampling window selects one reference signal from all the signals included in the training data, and changes the time point of sampling using the time delay for each signal. In FIG. 2, variable 1 becomes a reference signal, variable 2 having no time delay is sampled without time delay, and variable 3 having time delay is sampled by applying time delay analyzed through cross-correlation analysis. Will slow down. There is no special requirement for selecting the reference signal, but in the case of power generation facilities, it is preferable to select the main water supply flow rate signal as the reference signal. Using a dynamic sampling window that compensates for time delays can effectively analyze highly correlated variables.

Subsequently, correlation analysis and grouping are performed (S17). The training data contains a mixture of what contains useful information to inform the status of a particular plant and what does not. In addition, not all signals provide status for all equipment in the system, even if they contain useful information. Therefore, it is necessary to group signals containing information useful for checking the status of each target equipment. Grouping can exclude signals that contain information that is not useful from the training data, and reduce the number of signals needed to monitor a particular facility to an appropriate level. Correlation Coefficient is used as the standard value for grouping. Correlation coefficient is analyzed for all pairs of variables. If this value is over the set value, it is regarded as learning data. Otherwise, it is dropped from learning data. The setting value is input by the user.

Where ρ _XY is the correlation coefficient between variables X and Y, Xi is the i-th value based on the sampling interval of the training data, and Yi is the i-th value based on the sampling interval of the training data (variable from X) , μ _X is the mean for variable X, μ _Y is the mean for variable Y, σ _X is the standard deviation for variable X, σ _Y is the standard deviation for variable Y, and N is the sampling interval of the training data. The number of data collection intervals within.

The variables finally included in the training data perform unsupervised grouping according to the correlation of the pairs of variables. In the non-supervision grouping, the number of groups to be classified is not determined in advance, the number of groups is determined according to the degree of correlation of the pair of variables, and the number of variables belonging to the group is determined. Unsupervised grouping is possible by processing the results of correlation analysis as shown in FIG. In FIG. 3, when the correlation coefficient is greater than or equal to the set value, a dot is shown, and when the correlation coefficient is less than or equal to the point, the processing is performed so that there is no point. Group 1 has three variables, group 2 has five variables, group 3 and 4 have two, and finally group 5 has only one variable.

Next, regression analysis is performed to generate a prediction model for use in the execution mode for quadruple verification for each group (S18). Although the present invention does not propose a specific regression analysis method, a regression analysis method capable of auto-correlation is selected in consideration of the usability in the execution mode. Autocorrelation means that the input signal corresponding to the independent variable and the output signal corresponding to the dependent variable are the same, and a diagram of this is shown in FIG. 4. If the autocorrelation regression model is composed of the training data obtained in the normal state of the equipment, when the data obtained in the normal state of the equipment is input, the output signal is the same as the input signal. In the case of input, a deviation occurs between the input signal and the output signal. The autocorrelation model may be configured as parametric or non-parametric. The present invention proposes an autocorrelation regression model based on the concept of non-parametric lazy learning in consideration of the usability in execution mode.

The reliability of the autocorrelation regression model is verified by the concept of 4-fold validation.

First, a good regression model is to reduce the goodness-of-fit and smoothness at the same time.

The goodness of fit is calculated as follows.

The smoothness is calculated as follows.

This is not possible, however, because a higher order polynomial is required to reduce the sum of squares of the residuals. In this case, the smoothness value is increased, and a lower order polynomial is required to reduce the smoothness value, but the residual sum of squares is It becomes bigger. Therefore, a value is needed to balance the goodness of fit and smoothness. In other words, an excellent regression model is a function that minimizes S (λ) with respect to the appropriate weight λ.

S (λ) is calculated as follows.

Here, the weight λ is a value representing a balance between the residual sum of squares and the smoothness, which is called a smoothing parameter (bandwidth). In the present invention, the process of finding such λ uses a grid search method based on a quadratic verification method. A detailed description of the grid search method based on the quadratic verification method is illustrated in FIGS. 5 and 6.

First, the quadruple verification method is to divide the training data into quadrants, make an autocorrelation regression model using the data of the third portion, and then repeat the method of verifying the model using the remaining data in different combinations. This will perform a total of four verifications. Among these, the third data used to create the autocorrelation regression model is called learning data, and the first data used to verify the created regression model is called testing data. Each verification step is called a run. Therefore, the quad verification method performs four runs.

For each run, we use the Square Sum of Residuals (SSR) as an indicator of the superiority of the regression model. In the autocorrelation regression model, since the input signal and the output signal are configured identically, the residual becomes zero if the input signal and the output signal match exactly. In this case, the sum of squares of the residuals (SSR) is zero. Even if the data obtained from the regression model is different from the data used to generate the regression model, if the data obtained from the normal equipment is input, the residuals remain on average on average, and they fluctuate randomly with a standard deviation of a certain size. If this behavior follows a normal distribution, then even in this case, the sum of squares of the residuals (SSR) is on average zero. Inadequate regression models lead to residuals that do not follow a normal distribution or that the mean of the residuals deviates from zero. Therefore, the sum of squares of the residuals (SSR) is out of zero. Therefore, selecting the appropriate smoothing factor (λ) in the regression model will minimize the sum of squares of the residuals (SSR).

In the present invention, there is no particular requirement for a method of finding an optimal smoothing coefficient (λ) for minimizing the sum of squares of the residuals (SSR), but the practical point is highlighted, and the grid search method is performed. The lattice search method sets the smoothing coefficient λ to a moderately small value, for example, λmin = 0.01 and a moderately large value, for example, λmax = 100, and divides it by M to determine the smoothing coefficient λ. A quadrature verification method is performed for each smoothing coefficient [lambda], and the smoothing coefficient [lambda] showing the smallest sum of squares (SSR) is regarded as an optimized smoothing coefficient [lambda]. To find a more accurate smoothing coefficient (λ), set two consecutive smoothing coefficients (λ) with the smallest sums of squares (SSR) again as λmin and λmax, divide them by M, and for each smoothing coefficient (λ) Perform a four-factor verification method. The method of overlapping the grid search method is called a nested grid search method. The overlapping grid search method is illustrated in FIG. 6.

Where SSR _λ is the sum of squared residuals (SSR) of each run with respect to the smoothing coefficient (λ), k is the number of runs, X _ik is the input signal value at the kth run, and X ' _ik is at the kth run Is the output signal value of.

Is the mth smoothing coefficient in the superimposed lattice search, λmin is the smallest expected smoothing coefficient, λmax is the largest expected smoothing coefficient, and M is the number of lattice between λmin and λmax.

After finally finding the optimal smoothing coefficient λ, the learning mode ends (S19). The final learning mode result is an autocorrelation regression model using the training data and the optimal smoothing coefficient (λ). The frequency of performing the learning mode depends on the situation of the target system. This is usually done when there is a possibility that the correlations between variables in the system may have changed, such as through large plant replacements or large planned outages.

The execution mode is described step by step as follows.

In run mode, it is executed only when the data obtained under the operating condition is not exceeded the range of data prepared in the learning mode.If not, the user is alerted that the output result is not reliable or the calculation is automatically bypassed. Be sure to

First, the operation data used in the execution mode is input after undergoing a preprocessing step similar to basically preparing the training data (S21). The above operating data are obtained from the conditions under which the industrial equipment is in normal operation.The external conditions (air temperature, pressure, humidity, sea water or precipitation temperature if cooling water is required), the characteristics of the fuel injected, the operating range and Prepare data in various combinations, such as aging or degradation. All or some of the variables that can be measured on or off-line through the measuring equipment installed in the industrial facilities to be monitored are treated as operational data.

The driving data removes the singularity in the same manner as removing the singularity from the training data (S22). If operation data were obtained from a database where basic signal processing is being performed, it is assumed that the singularity has already been removed. If operational data is obtained directly from a database or sensor that is not performing sufficient signal processing, appropriate outlier removal methods should be used. Representative examples of singularity include problems caused by no data being entered at all, such as 'bad input', and cases where data is input like 'out of range' but is temporarily larger or smaller than well beyond the normal range. do. If such data are generated, the reliability of the operation data is increased by simultaneously removing the data of all variables acquired at that time under the following conditions.

if x ^L ≤x _i ≤x ^U then x _i OK, otherwise remove x _i

Where x _i is the i-th measured data value, x ^L is the normal lower limit of the variable x, and x ^U is the normal upper limit of the variable x.

Next, in the case of driving data, an average value is taken as in the learning data (S23). Considering the dynamic characteristics of the industrial equipment and the time interval of the system for acquiring the operating data, the interval to take the average value is determined. The driving data also applies the same interval as the average value interval used when preparing the training data.

Here, X _j is the j th value based on the sampling interval of the driving data, x _i is the i th value based on the data collection interval, and N is the number of data collection intervals within the sampling interval of the driving data.

Signal filtering may be performed on the driving data like the training data. However, the execution mode includes the step of finally presenting the result to the user in consideration of the uncertainty of each signal, so that even without filtering, it does not affect the reliability of the entire monitoring system. In addition, filtering requires data to accumulate over a period of time, which is often not appropriate in actual execution mode. Therefore, signal filtering is excluded in run mode.

Next, operation data is sampled using a dynamic sampling window used to prepare training data (S24). Data collected through the dynamic sampling window can eliminate the loss of correlation over time delay.

Subsequently, an autocorrelation regression model is applied to the signals classified by each group (S25). The process of applying the driving data to the regression model is the same as the method used to verify the training data. When the driving data is input to the regression model, the regression model performs autocorrelation regression to generate an output signal. Residuals are calculated for all variables in the group, finally providing information to the user about the process margin of the plant.

The process margin is determined by combining the confidence interval with the actual measured value. The confidence interval in the present invention is a combined form of two elements. The first is the residual calculated in the autoregressive regression model, and the second is all the uncertainty in the arrival of signal values from the sensor to the user interface. All of the uncertainties that occur within the system are included in some of the two components, but the first component, residuals, mainly contains bias and random uncertainties associated with the physical phenomena within the system. The second component, signal uncertainty, is the measurement system. It mainly contains biases and random uncertainties associated with

Where PMCI _ip is the Process Margin Confidence Interval at i-th point of variable p, e _ip is the residual at i-th point of variable p, and E _p is the signal uncertainty of variable p .

Where E _p is the signal uncertainty of the variable p, and u _j includes bias and random uncertainty as uncertainty factors that can occur in all factors associated with the variable p being measured.

Next, a process margin confidence interval (X _i ± PMCI _i ) is combined with the input signal and output to the user interface (S26). Some of the input signals may be signals linked to the stop set point of the industrial facility. The user can check the measured value and the value including the process margin for the important signal at the same time, so that the user can easily recognize the operating condition of the industrial equipment. Even if the signal has the same margin from the set value, the wider process margin means that the uncertainty is increased in the system operation state, and the narrow process margin means that the system operation state is clear without uncertainty. Can be. When the width of the process margin reaches the set point, a timely alarm is triggered to prepare the user for proper operation and maintenance.

1 is an operation flowchart of a dynamic monitoring and timely alarming method for a process margin of an industrial facility according to an embodiment of the present invention.

2 is a conceptual diagram of a dynamic sampling window of the dynamic monitoring and timely alarming method for the process margin of the industrial equipment according to an embodiment of the present invention.

3 is a diagram illustrating unsupervised grouping by a correlation between a dynamic monitoring and a timely alarming method for a process margin of an industrial facility according to an exemplary embodiment of the present invention.

4 is a conceptual diagram of an autocorrelation regression model of a dynamic monitoring and timely alarming method for the process margin of an industrial facility according to an embodiment of the present invention.

5 is a conceptual diagram of a quadruple verification method of the dynamic monitoring and timely alarming method for the process margin of the industrial equipment according to an embodiment of the present invention.

FIG. 6 is a diagram illustrating an optimal smoothing coefficient by a superimposed grid search method of a dynamic monitoring and timely warning method for a process margin of an industrial facility according to an exemplary embodiment of the present invention.

Claims (12)

'Learning mode', which is a process to perform statistical learning modeling and data to be used in 'execution mode', and data prepared in the 'learning mode' and modeling results and operation data which are input periodically, Including 'run mode', which is a process that performs dynamic monitoring and timely alarming, The learning mode described above, Inputting training data, Removing outliers, which are problematic in statistical processing, from the learning data; Averaging the training data; Filtering the training data; Cross-correlation, Applying a dynamic sampling window; Performing correlation analysis and grouping, Performing regression analysis and performing quadruple verification of training data for each group to create a predictive model for use in execution mode; Finding the optimum smoothing coefficient (λ), and finally the step of completing the learning mode after completing the training data modeling, characterized in that the dynamic monitoring and timely alarm method for the process margin of the industrial equipment. delete The method of claim 1, Dynamic singularity and timely alarm method for the process margin of the industrial equipment, characterized in that the removal of the singular point is performed by the following formula. if x ^L ≤x _i ≤x ^U then x _i OK, otherwise remove x _i Where x _i is the i-th measured data value, x ^L is the normal lower limit of the variable x, and x ^U is the normal upper limit of the variable x. The method of claim 1, Dynamic learning and timely alarm method for the process margin of the industrial equipment, characterized in that the averaging for the learning data is performed by the following equation.

Here, X _j is the j th value based on the sampling interval of the training data, x _i is the i th value based on the data collection interval, and N is the number of data collection intervals within the sampling interval of the training data. The method of claim 1, The cross-correlation method is a dynamic monitoring and timely alarm method for the process margin of the industrial equipment, characterized in that performed by the following formula.

Here, r _XY is the cross-correlation coefficient in the time delay d between the variables X and Y, Xi is the i-th value based on the sampling interval of the training data, Yi is the i-th value based on the sampling interval of the training data ( with the proviso that X and other variables), μ _X is the average of the variables X, μ _Y is the average of the variables Y, N is the number of times of the data collection interval in the sampling interval of the learning data, d is a time delay of 0, 1 , 2, ..., N-1, and i is 0, 1, 2, ..., N-1. The method of claim 1, The correlation analysis is a dynamic monitoring and timely alarm method for the process margin of the industrial equipment, characterized in that performed by the following formula.

Where ρ _XY is the correlation coefficient between variables X and Y, Xi is the i-th value based on the sampling interval of the training data, and Yi is the i-th value based on the sampling interval of the training data (variable from X) , μ _X is the mean for variable X, μ _Y is the mean for variable Y, σ _X is the standard deviation for variable X, σ _Y is the standard deviation for variable Y, and N is the sampling interval of the training data. The number of data collection intervals within. The method of claim 1, The regression analysis is a dynamic monitoring and timely alarm method for the process margin of the industrial equipment, characterized in that performed by the following formula.

Here, lambda is a value representing a balance between the residual sum of squares and the smoothness, and is a smoothing parameter (bandwidth). The method of claim 1, The quadruple verification method is a dynamic monitoring and timely alarm method for the process margin of the industrial equipment, characterized in that performed by the following formula.

Where SSR _λ is the sum of squared residuals (SSR) of each run with respect to the smoothing coefficient (λ), k is the number of runs, X _ik is the input signal value at the kth run, and X ' _ik is at the kth run Is the output signal value of.

Is the mth smoothing coefficient in the superimposed lattice search, λmin is the smallest expected smoothing coefficient, λmax is the largest expected smoothing coefficient, and M is the number of lattice between λmin and λmax. The method of claim 1, The execution mode described above is Inputting driving data; Removing the singular point which is highly problematic in statistical processing from the driving data; Taking an average value of the driving data, Sampling the operation data using a dynamic sampling window; Calculating a process margin confidence interval by applying an autocorrelation regression model to the signals classified by each group; And combining the process margin confidence interval with the input signal (X _i ± PMCI _i )) and outputting it to the user interface. The method of claim 9, Dynamic singularity and timely alarm method for the process margin of the industrial equipment, characterized in that the removal of the singular point is performed by the following formula. if x ^L ≤x _i ≤x ^U then x _i OK, otherwise remove x _i Where x _i is the i-th measured data value, x ^L is the normal lower limit of the variable x, and x ^U is the normal upper limit of the variable x. The method of claim 9, Dynamic learning and timely alarm method for the process margin of the industrial equipment, characterized in that the averaging for the learning data is performed by the following equation.

Here, X _j is the j th value based on the sampling interval of the driving data, x _i is the i th value based on the data collection interval, and N is the number of data collection intervals within the sampling interval of the driving data. The method of claim 9, The process margin confidence interval is a dynamic monitoring and timely alarm method for the process margin of the industrial equipment, characterized in that determined by the following equation.

Where PMCI _ip is the Process Margin Confidence Interval at i-th point of variable p, e _ip is the residual at i-th point of variable p, and E _p is the signal uncertainty of variable p .

Where E _p is the signal uncertainty of the variable p, and u _j includes bias and random uncertainty as uncertainty factors that can occur in all factors associated with the variable p being measured.

Images