The invention relates to a method of monitoring and/or diagnosing the current operational state of a complicated dynamic system constituted by a machine, apparatus or an industrial process installation, in accordance with the pre¬ amble to claim 1.

Such a method, intended for monitoring machine vibrations while utilizing frequency analysis as well as sampling and pattern recognition techniques is known from EP-A-Θ4902732.1 , the reference class being frequency spectra and the pattern recog¬ nition and detection means being adapted such as to calculate the probability for each new frequency spectrum that it belongs to a class other than the reference class. An abnormal opera¬ tional state of the machine is determined when this probability exceeds a predetermined limit.

According to the present invention, it is not only a question of analysing occuring vibrations, but a large number of measurable or calculable physical and possibly even logical parameters or state variables, which to a greater or less extent may be con¬ sidered as effecting the operational state of a complicated system.

For example, in the operation and maintenance of complicated machines, e.g. jet engines, measurements of a large number of statistic and dynamic state variables are collected. Accordingly, being able to readily indicate an unfavourable combination of such variables is of the greatest importance here, not only with regard to limit values decided for each individual variable. The situation is similar for process installations, e.g. paper manufacture or energy conversion (such as in nuclear power stations).

There is further a desire of being able to calculate beforehand whether a proposed modification of a proven system will give acceptable results.

A parmount object of the present invention is thus to monitor and/or diagnose, possibly by calculation performed beforehand, and with the aid of mathematical, statistical methods, the existing operational state of a complicated machine, apparatus or process installation and in such a case either identify a specific operational state (corresponding to a reference class) or determine a deviating operational state by simultaneously identifying what state variables cause the deviation.

This object is achieved by the process of steps disclosed in the characterizing part of claim 1.

Statistic methods known per se are thus applied in accordance with the invention, these methods being used within so-called multivariate analysis, e.g. for rapidly discovering an undesirβd state in a machine and pointing out individual variables or groups of variables responsible for this state. The method may be applied either in real time, namely to incoming measurement values, or to data bases, where the measurement values have been stored.

The invention will know be explained in more detail below and with reference to the accompanying drawings, where

Fig. A is a schematic diagram over the measuring and data equip- ment, its structure being known per se, which is used in the method according to the invention;

Figs. 1 - 3 are flow diagrams corresponding to some application examples A, B and C;

Fig. la schematically illustrates a variant of the application example A; and

Fig. 3a is a variant of the flow diagram in Fig. 3.

The invention is founded on the following procedures:

The measurement values for different state variables Vi are obtained from suitable measuring sensors after amplification, filtering and digitizing (ββe Fig. A). For rotating machines both static and dynamic state variables may be used, e.g. the dimensions, taken from a coordinate table, of the machine com¬ ponents, clearance and play between given parts, throw and in- balance measured in a balancing machine, as well as different types of vibrations.

Measurement of M different state variables V ^* i is repeated at different instances N times. The results ( ifc, i = 1 , ... , M; k = 1, ... , N) , which form a matrix or table with the dimension M x N can be presented in a N-dimensional space as a cluster of N dots. The appearance of the cluster can be described, for example with the aid of a covariance matrix. If the dots have been collected during two different states of the system they will form, with the greatest probability, two clusters in the M space, provided that among the state variables there are pre¬ cisely those which differentiate the current states in the system. A possibility is that the major part of the measurement data, which are derived from one or more known states of the system (so-called learning groups), are combined with a plu- rality of current measurement data which are to be tested (a test group) .

For easy handling of this type of multivariate measurement data it is proposed in accordance with the invention that a co- ordinate transformation is made. For such a transformation to be utilizable it must:

- reduce the number of coordinate directions;

- maintain the variance of the dot clusters; and

- separate different dot clusters in the best way.

A transformation found to meet these conditions is the so-calle principal component method, which has been previously used in chemical data processing and in qualitative analysis of food¬ stuff. In this connection reference is made to the article "Pattern Recognition: Finding and using regularities in ulti- variate data" by S. Wold, C. Albano, W.J. Dunn, K. Esbensen, S. Hellberg, E. Johansson and M. Sjόstrόm in the publication "Food Research and Data Analysis" Applied Science Publishers, London and New York, 1983.

Applied to monitoring the state of machines, apparatus or industrial process installations, this principal component method means that multivariate measurement data Xi _k are approximated by linear projection to a new coordinate system with fewer coordinate axes (L; L < M) according to the relationship:

L X _ik = Ai + Σ B _ia T _ak + R _ik a=l where the new variables A, B and T describe the systematic part of the measurement data Xi _k , ^an where the residue matrix describes the random part of Xi _k . It should be pointed out that the transformation separates measurement data into parameters (A, B) describing the state variables, on the one hand, and parameters (T) describing the system state at the respective measurement (k), on the other hand. The parameters T _ak describe the location of the k-state in the M-dimensional dot cluster an are therefore used for discriminating state classes.

Adaptation of the principal components takes place in conse¬ cutive steps, one component at a time.

The procedure starts by normalizing the measurement data matrix _Xik with the standard deviation s^ of the respective state variable V^, so that all variables will be of appoximately the same order of magnitude, the mean value Ai being substracted, i.e. the matrix X^ _k - A^ is formed. At this stage the first principal component

PCI = Bii T _ik

is adapted while using the so-called NIPAL's algorithm (Appendix A in the above mentioned article by Word et al), which in short means that beginning with an attempt or qualified guess with respect to the vector B, the vector T is calculated and the convergence is checked, which is repeated until the desired accuracy has been achieved. The first residue matrix Rifci = ^x ik ^_A i~ ^{PC1 is} then formed, which is used as the input matrix for determining the next (second) principal component, etc.

The number of components (L) is determined by the character of the data matrix. The process of adaptation can be terminated when no systematic parts remain in the residue matrix R. _k . For practical reasons the process can also be terminated after PC2. The state can then be depicted on a plane with coordinate axes corresponding to PCI and PC2. A three-dimensional depiction can be constructed if the adaption is terminated after PC3. Thrθe- -dimensional depiction of measurement data are generally pre¬ ferred.

It is of the greatest interest to find out how much different state variables contribute to the variance along each principal coordinate. In order to separate different state classes it is required, as mentioned above, that measurement data really do contain the state variables which are relevant for describing the classes. To decide this, the so-called modelling capacity P^ _a of the state variable j_ along the respective coordinate axis a (a = 1, 2, ..., L) is defined:

^p ia = 1 ~ s _ia ⁽ R _{i ka} ) / β i (X _{i k} - A _± >

where Si _a is the standard deviation for the residue matrix Ri _a and Si is the standard deviation for the matrix (X-j. _k - Ai . P varies between 0 and i. If Pi _a is close to 0, the variable Vi does not affect the variance along the principal coordinate in question, and in principle can be ignored. The modelling capa¬ city can also be expressed as a percentage. In addition, the total modelling capacity P _it for PCI, PC2 and PC3 is defined according to the following:

Variables with a total modelling capacity Pi close to 1 are denoted "strong" variables. Strong variables are relevant for identifying different states, particularly fault states. Know¬ ledge of these variables helps the operator to identify faults in machines, for example, or may be used in coaction with a knowledge base for automatic fault identification.

Some practical examples A, B and C of the method in accordance with the invention are given below.

Example A. Analysis of a data base with motor parameters (se Fig.l)

In connection with the inspection of a jet engine, about a hundred different static and dynamic variables in the engine are measured. Thus, e.g. tolerance and throw in different parts of the rotor-stator system are measured accurately. The measure¬ ments are placed in a data base. When an engine has been approved after inspection, these measurement values are stored together with the engine documents. On the other hand, when an engine has been rejected, attempts have been made in different

ways to search among about a hundred measurement variables for the one or ones which have caused the rejection. This is an insolvable problem using elementary stastics as the sole aid. The use of conventional packages for advanced statistic analysis is unfortunately too time-consuming for enabling their opera¬ tional use.

However, in applying the inventive method, about ten data base posts from approved engines are used as a learning group. The post for the tested, rejected engine is added to the table

(matrix) . After principal coordinate transforming, the variables are ordered in rank with a modelling capacity according to size. In this way the one or ones of the state variables which are responsible with the greatest probability for rejection of the motor can be indicated. In combination with a knowledge base (a rule system) directions as to required technical measures can be worked out. If the method is applied before the final check in a test bench economies can be made in many cases, since the number of rejections after checking can be reduced.

A variation of this example is that the data vectors (the columns in the total data matrix) for the approved engines are divided into sections corresponding to different work stages j (see Fig. la) during assembly of the engine ( = 1, ....P) . In a corresponding way the data vector is divided for the tested engine and the analysis is made of each terminated work stage of a growing data matrix. This gives the possibility of breaking off the costly assembly process as soon as it is found that the tested engine will belong to the group of non-approved engines. At this stage the problem can be solved by using another selec¬ tion of parts or further adjustment.

Example B. Checking blade distribution in rotor discs (see Fig. 2)

Hitherto,the blades in a rotor disc are distributed according to simple rules decided by the manufacturer, with respect to the momentum weight of each blade. Computer programs calculating the distribution of blades around the rotor disc have also begun to come into use. Such a program distributes the blades with reference to their momentum weights so that inbalance is minimized. The method in accordance with the present invention enables the checking of a blade distribution arranged according to experience or calculation by comparison with previous, successful blade distributions, preferably from the same engine. Information in a data base containing such distributions of momentum weights which have given satisfactory dynamic properties are compared with a proposed distribution. It is here possible to obtain a measure of the quality of goodness of the distribution (distance to the group "successful distributions"), and an account of the blades causing possible negative dynamic properties in the rotor stage (strong variables) can thus be obtained. It is also possible to test several proposed distri¬ butions simultaneously.

Example C. Monitoring and advising with the engine running on a test bench (see Fig. 3).

When running an engine on a test bench, a large number M (usual¬ ly about a hundred) operational variables are measured and stored. In modern test bench equipments the positions of all manual controls handled by the operator are also read off.

Hitherto, the automatic monitoring is arranged such that maximum and minimum limits are given for certain essential variables, and that the system gives an alarm for certain programmed com¬ binations of discrete and logical variables. The system thus gives an alarm for previously known undesired situations.

The inventive method can here be applied according to two different variations for an engine running on a test bench:

Variation 1. A large learning group is formed containing 2 - 3 measurements for each permitted operational situation

(corresponding to a specific reference class). Each new measure¬ ment ("present state" in Fig. 3) is continuously added to the table or data matrix thus obtained. The new measurement (a dot in the M-dimentional space) is tested with respect to its distance to the learning group. If the new dot is outβidβ those of the learning group, a deviating stage is established and possible strong variables are accounted for. If measures taken by the operator give rise to strong variables the system can, with support of a knowledge base, give the operator advice and information about suitable measures in this situation. Possible operator errors can be discovered and dealt with in this way.

Variation 2. Several learning groups are formed, one for each interesting operational situation. Each learning group contains about ten measuring points coming from this situation. Certain undesired situations, e.g. faults or errors, can be stored among the learning groups. Each new measuring point, or several con¬ secutive incoming measuring points can then be tested againβt the learning groups. By searching for the best adaptation to a learning group, the operational situation can be identified with great certainty. A simultaneous account of strong variables can, in combination with a knowledge base, be used fox achieving a diagnosis and for giving advice with respect to possible measures .

Both variations can be used both in real time as well as 'subsequently for analysis of collected measurement data. Both variations can further be used on all types of machines, apparatus or industrial process installations, where a large number of operational variables make the situation difficult for a human being to comprehend.

This method can be made even more effective (see Fig. 3a) if the data vector in question ("present state") is transferred to an autoassociative neural network, e.g. of the type ART (Adaptive Resonance Theory Network), as deσribed in Carpenter, G.A., Grossberg, S., "A Massively Parallel Architecture for a Self-Organizing Neural Pattern Recognition Machine", Computer Vision, Graphics and Image Processing 37, pp 54-115, 1987. The network is trained with typical data from normal operational conditions. The remaining part of the data flow is only acti¬ vated on the discovery of a new category, for then carrying out the complete statistical analysis. In this way the statistical analysis does not need to be made during normal operational conditions .