EP2297623A1

EP2297623A1 - Method for monitoring an industrial plant

Info

Publication number: EP2297623A1
Application number: EP09772536A
Authority: EP
Inventors: Hajrudin Efendic; Gerald Hohenbichler; Stephan M. Winkler; Andreas Schrempf
Original assignee: Siemens VAI Metals Technologies GmbH Austria
Current assignee: Primetals Technologies Austria GmbH
Priority date: 2008-07-04
Filing date: 2009-07-03
Publication date: 2011-03-23
Also published as: US20110106289A1; AT507019B1; AT507019A1; WO2010000836A1

Abstract

The invention relates to a method for monitoring plants, in particular complex plants in the iron and steel industry, comprising the steps of recording at least two channels of measurement data of a plant, optionally saving the measurement data, defining a target channel from the measurement data, pre-processing the measurement data, creating at least one model of the target channel based on the measurement data and using the model so generated and the currently determined measurement data for detecting erroneous states of the plant. The object of the invention is to provide a method for monitoring industrial plants which improves upon the quality of the recorded measurement data of the plant and which can significantly reduce the amount of measurement data without resulting in a significant loss of information. This object is accomplished in a method in which the measurement data are, in the measurement data pre-processing step, subjected to the processing steps of 1) detection and elimination of "null channels", 2) detection and elimination of outliers, 3) filtering, 4) downsampling.

Description

Method for monitoring an industrial plant

The present invention relates to a method for the pretreatment of measurement data for monitoring an industrial plant.

Specifically, the invention relates to a method for monitoring systems, in particular complex systems of the iron and steel industry, with the steps of recording at least two channels of measurement data of a system, if necessary

Storage of the measurement data, definition of a target channel from the measurement data, pretreatment of the measurement data, measurement data based formation of at least one model of the target channel and use of the model thus generated and currently determined measurement data for the detection of malfunction of the plant.

Computer-aided methods for monitoring installations or processes are known to the person skilled in the art under the terms Fault Diagnosis or Fault Detection (hereinafter FD for short), see e.g. WO 02/086726 Al. These methods for the detection of fault conditions in complex industrial plants include the following method steps:

a) measuring data acquisition of at least two channels of a plant, b) optionally storing the measurement data, c) defining a target channel from the acquired channels of the measurement data, d) optionally pretreatment of the measurement data, e) measuring data based formation of at least one model of the

Target channel, f) use of currently determined measurement data and of the model formed to calculate a simulated value of the target channel, g) detection of false states by comparing between the current and the simulated value of the target channel. Modern industrial plants, for example blast furnaces or rolling mills, typically with a large number of coupled individual plants, are highly complex technical systems. Today, hundreds of thousands of measuring sensors are permanently and real-time monitored (eg with a sampling time of 1 ms) to monitor these systems or processes, and special relevant data is displayed by a measuring data acquisition system. Due to the high demands on the operating personnel, however, FD is increasingly used in complex systems, since these methods allow computer-aided errors in the computer system

Detect equipment and, if necessary, identify an affected plant area. FD is also in the condition monitoring, engl. Condition monitoring, used by industrial equipment. In this case too, a system is monitored by means of a large number of measuring sensors and automatically generates a warning or error message when the system behavior changes, for example as a result of wear of components. The measured data are typically various plant or process conditions, which are determined by sensors and usually in digital form of a

Recorded data acquisition system. A channel of measured data is a series of measured values recorded by a sensor; A destination channel is a channel of recorded video

Measurement data that contains or may contain relevant information about the behavior of the plant. In the simplest form of the FD, a model for the target channel based on one or more channels (excluding one target channel) of plant measurement data is created; By means of this model and depending on current measurement data, the values of a simulated target channel are calculated by a process computer and compared with current measured values of the target channel; if there are significant deviations between the simulated and the measured target channel, an error message is generated.

In the publication H. Efendic, L. Del Re and G. Frizberg. Iterative Multi-Step Diagnosis Process for Engine Systems. Presented at the SAE world congress, April 11-14, 2005. Detroit, Mich. USA.

is focused on the importance of pre-treatment of measurement data. Data preprocessing, pointed out in the course of the FD. Specifically, it is proposed to filter the measurement data to eliminate disturbances in the measurement data and thus to improve the performance of FD.

Further measures for the pretreatment of measured data are not indicated.

The object of the invention is to provide a method for monitoring of industrial plants, with which the quality of the recorded measurement data of the system improved and the size of the measurement data can be greatly reduced, without resulting in a significant loss of information. Below, if necessary online - ie. By a process computer associated with the system with possibly low performance, based on these compressed measurement data, very compact models for the detection of fault conditions of the system (FD) are generated, which are also suitable for use on low-cost process computers with relatively low performance and yet high quality. have the error detection.

This object is achieved by a method in which the measured data in the step pre-treatment of the measured data the method steps

1) detection and elimination of "zero channels",

2) detection and elimination of outliers, 3) filtering,

4) Downsampling be subjected. Since so-called "zero channels", ie channels of measurement data which are constant at all sampling times, do not contain any information, these channels are deleted, thus reducing the number of relevant channels for the subsequent steps.

In the detection and elimination of outliers, for example with regard to the process control, unjustified, outliers are detected in the measurement data and then eliminated. The elimination of an outlier occurs by replacing an outlier with an average of the affected channel.

During filtering, the measurement data channels are smoothed, for example by the use of median filters, resulting in e.g. Measurement noise is reduced and, as in the detection and elimination of outliers, the quality of the measurement data is increased.

When downsampling the information content of a

Data channel before downsampling determined and compared with the information content of a changed. With respect to the sampling time data channel. This results from downsampling, ie. by a reduction of the sampling frequency, no significant change in the information content, the sampling frequency is reduced, whereby a large reduction of the amount of data (a reduction of the sampling frequency by 50% reduces the amount of data also by 50%) is possible.

Advantageously, in the pretreatment of the measurement data, the measurement data are subjected to the process steps in the order of detection and elimination of "zero channels", detection and elimination of outliers, filtering and downsampling.This order results in a high quality of the measurement data and in a high efficiency of the inventive method. In an advantageous embodiment, the measurement data, preferably after downsampling, are subjected to detection of stationary areas and elimination of non-stationary areas, which further reduces the size of the measurement data and the formation of simple, static

Process models for a subsequent FD is enabled.

A further advantageous embodiment is that, for different target channels, the steps of defining a target channel from the measured data, pretreatment of the measured data and measurement data based formation of at least one model of the target channel per target channel are each performed at least once, and models formed thereby in the detection of malfunction of the system be used. As a result, a particularly comprehensive monitoring of the system is achieved.

In a further advantageous embodiment, the steps of defining a target channel from the measured data, pretreatment of the measured data and measurement-data-based formation of at least one model of the target channel on at least one process computer are performed in parallel for different target channels. This is especially in online operation, ie. in the implementation of the method on one of the system associated process computer, given an early model availability. The parallelization can be carried out either by a plurality of tasks or threads on a process computer, and / or by division into several process computers.

In an advantageous embodiment of the method according to the invention, the detection and elimination of outliers contains a univariate and a multivariate step. The univariate method step is particularly suitable for detecting and eliminating comparatively large outliers in one channel independently of other channels. On the other hand, in the multivariate step, the distance of the measured values of all channels at a time becomes the total distribution determined, which also difficult to detect outliers can be detected and eliminated.

In a further advantageous embodiment, the measured data are subjected to a median filtering. Median filters are known in the art and allow a very efficient smoothing of signals.

In a further advantageous embodiment, the downsampling of the measured data takes place taking into account the

Transinformation, engl. Auto mutual information, between a channel before and after downsampling. This makes it possible to reduce the sampling frequency as a function of the loss of information due to the downsampling and thus set an optimized downsampling rate.

A further advantageous embodiment is that the detection of stationary areas and elimination of non-stationary areas taking into account statistical variables for the variability.

By means of this measure, stationary areas can be detected simply and reliably, which leads to a high quality of the models.

Advantageously, the measured data, preferably after the

Downsampling or the detection of stationary areas and elimination of non-stationary areas, a detection and elimination of redundant channels subjected. By means of this step, the number of relevant channels is further reduced; In this case, complete redundancies and possibly also redundancies resulting from a derivation or integration of a signal can be taken into account.

Further advantages and features of the present invention will become apparent from the following description of non-limiting embodiments, reference being made to the following figures, which show the following: Fig. 1 to Fig. 4 Plots of the channels Xi to X20 of the unchanged measurement data of a system

Fig. 5 Plot of a section of the target channel xio before and after the step of detection and elimination of outliers

6 Plot of a section of the target channel xio before and after the step filtering

Fig. 7 Plot of a section of the channel X20 before and after

Step Filtering Fig. 8 Schematic representation of the downsampling

Fig. 9 plot of the target channel xio before and after the step

downsampling

10 shows representations of the auto-mutual information and the scaled auto-mutual information as a function of the downsampling rate

Fig. 11 Plot of the target channel xio and the redundant

Channels x _u , X13 and X14

Fig. 12 Plot of the target channel xio and the redundant channel X13

Fig. 13 Plot of the target channel xio before and after the step Detection of stationary areas and elimination of non-stationary areas and the channels xio, X20 and x. ₇

14 is a flow chart of the most important method steps

In the following, for reasons of clarity, 20 data channels (number of channels n = 20) and 19498 measured values per

Channel (number of measured values m = 19498) assumed; The data were recorded by a data acquisition system in a rolling mill (real measured data are usually much larger, eg n = 400, m = 1500000). MD symbolically shows a matrix of the measurement data with dimension 19498 x 20. A measured value X ₁ , - represents the i-th measured value of the j-th data channel of the measured data.

xu n, i9 ^Λ l, 20

X ₉ 2.1x. 2.2 X 2.19 ^X 2.20

MD =

^• ^ 19497,1 ^• ^ 19497,2 ^• ^ 19497,19 - ^ 19497,2O • ^ 19498,1 ^• ^ 19498,2 ^• ^ 19498,19 - ^ 19498,2O The measurement data are, for example, signals from pressure, force, displacement, velocity, acceleration or temperature sensors, which were recorded for subsequent use in an FD method.

The channel xio was selected as the target channel (tenth column of MD). In FIGS. 1 to 4, the channels Xi to X20 are shown graphically above the measured data index.

1) Detection and elimination of "zero channels"

In the detection and elimination of "zero channels", those channels are eliminated from the measurement data where all measured values are 0. This reduces the number of channels.

A "null channel" exists if, for a channel, ie for a column vector of MD, X ₁ = O with i = 1 ... m

By the method according to the invention, the channels 9 and 17 were identified as zero channels and eliminated. The dimension of the measurement data matrix after the step of detection and elimination of "zero channels" is 19498 x 18, ie the amount of data has been reduced to 90%.

Subsequently, a univariate and possibly a multivariate detection and elimination of outliers is performed.

2a) Univariate detection and elimination of outliers

First, the individual channels, ie. the individual columns of the measurement data matrix, a univariate, ie. based on only one channel, detection and elimination of outliers subject. This eliminates "big" outliers. The detection of outliers can be done in two ways:

i) Global detection and elimination of outliers

Since the execution of the method step on a channel of the measurement data matrix MD can not be shown clearly, the following vector x of a channel of measurement data with m = 10 elements is used x = [l 2 3 10 5 4 3 2 1 θf. The average and the empirical standard deviation of the vector x are

1 ms = - £ (X ₁ -x) ² = 2.86 xx

Now the test criterion G = s is calculated for each measured value X ₁ . The one-sided test is an outlier,

two-sided test is that

• Below, for example, a one-sided test, with α = 0.05, t _{α /} m, m-2 = to. ooos, 8 = 3, 35 the upper critical value of the t-distribution at a significance threshold of α / m and m-2 degrees of freedom (see Tab IX, Appendix El "Tables" in H. Rinne, Taschenbuch der Statistik, 4th Edition, 2008).

Since for the value X ₄ = 10 the one-sided test criterion G = 2.41 is greater than the limit of 2.18, there is an outlier; consequently, x _{4 is} replaced by the mean x = 3,1.

This procedure is now performed for each channel until there are no more outliers.

ii) Local detection and elimination of outliers Since global data collection and elimination is too costly for large data sets, a local method is advantageously used. In this case, a channel by means of a viewing window, engl. Sliding window, with NAUS elements examined for outliers and replaced found outliers by a local average x _locally within the viewing window.

x = • 1 2 3 10 5 4 3 2 1 0-

View window with NA US elements X local ^~ -MS _local = 2.86

The check for outliers is analogous to the global procedure, but is limited to the viewing window.

After performing the detection and elimination of

Outliers, in turn, have a measurement data matrix which is free of univariate outliers but still has the original dimension.

In Fig. 5, the target channel xio before and a section of the target channel after the local detection and elimination of univariate outliers (NAUS = IO, α = 0.05) is shown. The detection and elimination of outliers is applied to all channels.

2b) Multivariate detection and elimination of outliers

This process step can be performed following univariate detection and elimination of outliers. In this case, any outliers due to the so-called Mahalanobis distance or the distance due to the so-called principal component analysis of a measured value vector x (a row vector of the measured data matrix) are detected by the overall distribution. Any outliers found are marked and replaced by local averages.

The Mahalanobis distance as well as the principal component analysis are e.g. known from

Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang, "A Novel Anomaly Detection Scheme Based on Principal Component Classifier," Proceedings of the IEEE Foundations and New Directions of Data Mining Workshop, in conjunction with the Third IEEE International Conference on Data Mining (ICDM'03), pp. 172-179, November 19-22, 2003, Melbourne, Florida, USA.

The calculation of the Mahalanobis distance d is still out

Chapter D3.1 "Distance Measurements" from Rinne The square of the Mahalanobis distance (distance of a measured value vector x from the center of the distribution) is defined as where y, = e _t (xx) ^τ p Rank of the covariance matrix x measured value vector (row vector of the measured data matrix) x average of all measured values e _x i-th eigenvector of the covariance matrix

X ₁ i-th eigenvalue of the covariance matrix

Subsequently, the indices i of the N = γ-m become largest

Distances d ² and each measured value X ₁ , -, with n≥j≥l replaced by x ~ \ ~ x a local mean X _1} = - '- ^ -. In this case, γ is a freely selectable parameter, for example γ = 0.005.

Subsequently, a principal component analysis of the measured data is carried out. In principal component analysis, the main components of the measurement data matrix, ie. the eigenvalues and eigenvectors, either via an eigenvalue analysis or a SVD (Singular value decomposition) of the covariance matrix is calculated.

A metric x is an outlier, if, where j _; = e _; (xx) ^r and q <pp Rank of the covariance matrix x measured value vector (row vector of the measured data matrix) x average of all measured values e _x i-th eigenvector of the covariance matrix λ _± i-th eigenvalue of the covariance matrix A ₁ > A ₂ >...> λ _p > 0 χ _q ² Upper critical value of the Chi-square distribution with significance value α and q degrees of freedom.

As with the Mahalanobi distance, outliers are replaced by local averages.

3) filtering

For example, when filtering the measured data

Noise or other interfering signals are removed from individual channels of the measuring signals. In this case, for example, a median filter with a viewing window of size N (referred to as filter order) is used.

During filtering, each measured value x _{k of} a channel is given by _ _ (mean value (x _k _ _(N _ _{1) / 2} , ..., x _{i + (7V} _ _{1) / 2} ) for odd N {mean value (x _k _ _{N / 2} , ..., x _{k + N / 2} ) for even N replaced.

For example, assuming a second-order filter and a vector x = [l 2 3 3.1 5 4 3 2 1 θ] ^r , the filtered vector will be x _Püt = [l 2 2.7 3.7 4 4 3 2 1 θf. FIG. 6 shows a section of the target channel xio and FIG. 7 a section of the channel X20 before and after filtering with N = IO.

4) Downsampling

The downsampling of the measured data is carried out taking into account the transinformation (auto-mutual information, see chapter 3.3.1.3 "Entropy-oriented measures" from the gutter) AMI (τ) between a channel before the downsampling and the same channel after the downsampling. The transinformation AMI (τ) is calculated from k

H (A) = - ^ p (α _t ) • log ₂ p (α _t ) Entropy of channel A before downsampling

H (A) = - ^ p (u _j) • log ₂ p (u _j) the entropy of the channel A to Down Sampling

7 = 1

~ ^{k ι} H (A, A) = - ^ ^ P (Ci ₁ , S _j ) - log ₂ P (Ci ₁ , α _} ) Entropy of the common distribution

: = 1 J = I

AMI (τ) = H (A) + H (A) - H (A, A) Transinformation between A and A

FIG. 8 schematically shows the effect of downsampling; a downsampling rate of τ = i means that only every i-th value of the original signal is used.

For each downsampling rate τ = i with ie [r _mn , T _{n ^ x} ], the associated transinformation AMI (τ) is calculated. The optimal downsampling rate, ie. that τ, where as little information as possible is lost with the greatest possible data reduction, is the first local minimum of AMI (τ), ie. AMI (τ -1)> AMI (τ) <AMI (τ + 1). If such a minimum can not be found, then one uses a scaled one

Transinformation AMI (τ) = AMI (τ) + e ^TF ° ^clor -1, so that a local minimum can be found. The factor τ _Fa k _to r is a scaling factor and can be selected appropriately. In Fig. 9, the target channel is shown before and after downsampling, but of course, downsampling is applied to all channels. The downsampling rate used is τ = 13, ie. the data volume of the measured data was reduced to 1/13. FIG. 10 also shows the transinformation AMI (τ) and the scaled transinformation AMI (τ) ^* over τ

(used parameter τ _factor = 500).

Downsampling reduced the metric matrix to 6, 9% of the original dataset.

5) Detection and elimination of redundant channels

By means of this method step, completely redundant channels in the measurement data are identified and subsequently eliminated, whereby the number of channels is reduced. For this purpose, a viewing window with size NRED is used, which is applied in pairs to the target channel X = xio and another channel Y different from the target channel. Within the viewing window, a redundancy signal (cross-correlation between the target channel X and Y within the viewing window) is calculated, with RED (J) = corr (X _local J _local ).

^■ X- local ^{~ X} ι-NREDI2 1 + NREDI2 * local ⁼ yι-NREDI2 1 + NREDI2

If the redundancy signal for all indices is i, RED {ϊ)> RED _bound , with, for example, RED _bound = 0, 95 or 0.98, channel Y is redundant with respect to target channel X and channel Y can be eliminated.

FIG. 11 shows the channels X ₁₁ , Xi ₃ and Xi ₄ redundant to the target channel Xi ₀ ; in Fig. 12 of the target channel and the channel Xio X ₁₃ are shown in a plot (Parameters used RED _barrier = 0, 98, NRED = 150). The redundant channels are deleted from the measurement data matrix. However, it should be noted that a redundant channel to a destination channel already too represents a simple model of the target channel that can be used for error detection. Due to this process step, the measurement data matrix was reduced to 5.89% of the original data volume.

In the detection and elimination of redundancies resulting from the derivation of a target channel, one proceeds essentially as above, but a numerical derivative or integration of the channel Y is performed before that of the redundancy signal RED (i) is calculated becomes.

6) Detection of stationary areas and elimination of non-stationary areas

By means of this method step, substantially constant operating points are to be identified from the measured data. For this purpose, a viewing window with size NVAR is used, which is applied to the target channel. By means of this window, a variability signal VAR (i) is calculated for an index i of the destination channel, with VAR (ι) = | max X _local -min X _local .

^■ * ^■ local ^{~ X} i-NVÄRI2 i + NVÄR / 2 * local ^~ yı-NVÄR / 2 ι + NVAR / 2

Is the variability at position i VAR (i) less than one

Part x of the maximum variability, ie. VAR (i) <VAR _SC with VAR _limit = x • max VAR (i), this range is considered stationary. Of course, other statistical parameters, such as the standard deviation, can also be used for the variability.

FIG. 13 shows the target channel xio before and after the detection of stationary areas and the elimination of non-stationary areas (parameter NVAR = 30, x = 0.2). Further, the channels are shown X20 and x. ₇ As shown, channel X20 has a similar dynamic range as the Target channel on and would be in the sense of a - not explained here model formation - an interesting candidate for a simple static model. The channel x ₇ in contrast has a completely different dynamics than xio and would therefore be less suitable for modeling.

This process step reduced the measurement data matrix to 4.4% of the original data volume.

The resulting measurement data matrix is the basis for a subsequent formation of at least one model for the target channel. A process computer assigned to the installation calculates a simulated target channel, which is used for the comparison between the simulated and the measured target channel, on the basis of currently determined measurement data of the installation and the generated model. If there are significant deviations between these two channels, a warning or an error message is generated. The detection of faulty states of a system can be carried out particularly comprehensively if not only one model of a destination channel is created, but for each of the different destination channels at least one model for the respective destination channel is formed and models formed thereby are used in the FD. Further steps, such as the identification or the isolation of errors, e.g. the publication

H. Efendic, A. Schrempf and L. del Re. Data based fault isolation in complex measurement Systems using mode on demand. Presented at the Safeprocess Conference, June, 2003. Washington, DC, USA.

be removed. 14 shows in a flow chart the most important method steps in the pretreatment of the measured data. In this case, at least two channels of measured data (eg from different sensors, such as pressure, temperature, speed or force sensors) of a system of the iron or steel industry are recorded and stored by means of a measurement data acquisition system. After the definition of a target channel from the channels of measurement data, the originally present measurement data (1) in the pretreatment of the measurement data (7) are successively the method steps

- Detection and elimination of "zero channels" (2) Detection and elimination of outliers (3)

- Filtering (4) Downsampling (5) subjected, whereby the data volume of the resulting measured data (6), ie. the number of channels of the measured data and / or the number of measured values or measuring points per channel is greatly reduced, which in turn has a positive effect on the subsequent steps of the FD measuring data based formation of at least one model of the target channel

- Use of the model thus generated and currently determined measurement data for the detection of malfunction of the plant effects.

Claims

claims

1. A method for monitoring installations, in particular complex installations of the iron and steel industry, comprising the steps of recording at least two channels of measured data of a plant, optionally storing the measured data, definition of a target channel from the measured data, pretreatment of the measured data, measurement data based formation of at least one Model of the target channel and use of the model thus generated and currently determined measurement data for detecting malfunction of the system, characterized in that the measurement data in the pretreatment of the measured data the method steps

- Detection and elimination of "zero channels" - Detection and elimination of outliers;

- filtering;

- downsampling; be subjected.

2. The method according to claim 1, characterized in that the measurement data in the pretreatment of the measured data are subjected to the method steps in the sequence of detection and elimination of "zero channels", detection and elimination of outliers, filtering and downsampling.

3. The method according to any one of the preceding claims, characterized in that the measured data, preferably after the downsampling, a detection of stationary areas and elimination of non-stationary areas are subjected.

4. The method according to claim 1, wherein for different target channels, the steps of defining a target channel from the measured data, pretreatment of the measured data and measurement-data-based formation of at least one model of the target channel per target channel are carried out at least once each time and models formed during the Detection of malfunction of the system can be used.

5. The method according to any one of the preceding claims, characterized in that for different target channels, the steps of defining a target channel from the measured data, pretreatment of the measured data and measuring data based formation of at least one model of the target channel are performed in parallel on at least one process computer.

6. The method according to any one of the preceding claims, characterized in that the detection and elimination of outliers contains a univariate and optionally a multivariate step.

7. The method according to any one of the preceding claims, characterized in that the filtering of the measured data is performed by a median filter.

8. The method according to any one of the preceding claims, characterized in that the downsampling of the measured data taking into account the trans information between a channel before and after the downsampling takes place.

9. The method according to any one of the preceding claims, characterized in that the detection of stationary areas and elimination of non-stationary areas taking into account statistical variables for the variability.

10. The method according to any one of the preceding claims, characterized in that the measurement data, preferably after the downsampling or the detection of stationary areas and elimination of non-stationary areas, a detection and elimination of redundant channels are subjected.