DE102019215016A1 - Measuring arrangement, method for setting up a measuring arrangement and method for operating a measuring arrangement - Google Patents

Abstract

According to the invention, the measuring arrangement has an input module for acquiring data from several data sources of the machine or system - the data representing status information and / or controlled variables and / or manipulated variables and / or other measured variables of the machine or system - and for creating an observation vector for an acquisition time Items that represent data from the data sources at collection time. The observation vector can be in the form of a simple data record, with data fields that contain values of the sizes mentioned from the data sources. There is a first processing module in which an artificial neural network is set up, the artificial neural network being set up to generate at least one output date on the basis of the observation vector supplied to it, which characterizes an operating state of the machine or system, and there is an evaluation module which evaluates the operating status based on the date of issue and outputs the evaluation to a maintenance system and / or to a control of the machine or system. A special feature is in particular that in the first processing module at least one first element of the observation vector is assigned to a first one-dimensional Kohonen map, by which a value of said first element is mapped to an output node of said first Kohonen map and thus a first hit output node is obtained and furthermore a hit deviation is obtained as a measure of the deviation of the value of said first element nts from the first hit output node, the first hit output node and the hit deviation being output data of the first processing module.

Description

The present invention relates to a measuring arrangement that can use machine learning to recognize operating states of a machine or system.

Such measuring arrangements are generally known. This is how the scientific publication "Krogerus, T .; Vilenius, J .; Liimatainen, J .; Koskinen, KT 2006. Self-organizing maps with unsupervised learning for condition monitoring of fluid power systems. Fluid Power for Mobile, In-Plant, Field and Manufacturing. SAE SP-2054 pp. 43-51 " proposed to use two-dimensional Kohonen maps - also known as "self organizing maps" - to identify normal and abnormal operating conditions of a hydraulic system. Attempts are made to use the data from known error states and normal operating states to generate areas on the Kohonen map that represent these states.

However, this proves to be difficult and often difficult to understand, since each network node of such a Kohonen map represents a mixture of measured variables as a multidimensional state. With "supervised machine learning", in which - as described above - known error states are used for training the Kohonen map, the training is very complex since measurement data must be generated for a large number of error states. In addition, the risk of incorrect diagnoses is high if previously unknown error states occur.

In so-called "unsupervised learning", in which only the normal state or good state is taught, there is a simple automatic detection of an error, but the error can hardly be identified.

With the present invention, the effort for training a neural network, which is part of the measuring arrangement, is to be reduced and nevertheless a better identification of an error should be possible.

This is achieved according to the present invention with a measuring arrangement with the features of claim 1, with a method for setting up a measuring arrangement according to the features of claim 11 and with a method for operating a measuring arrangement with the features of claim 16.

According to the invention, the measuring arrangement has an input module for acquiring data from several data sources of the machine or system - the data representing status information and / or controlled variables and / or manipulated variables and / or other measured variables of the machine or system - and for creating an observation vector for an acquisition time Items that represent data from the data sources at collection time. The observation vector can be in the form of a simple data record, with data fields that contain values of the sizes mentioned from the data sources. There is a first processing module in which an artificial neural network is set up, the artificial neural network being set up to generate at least one output date on the basis of the observation vector supplied to it, which characterizes an operating state of the machine or system, and there is an evaluation module which evaluates the operating state based on the date of issue and outputs the evaluation to a maintenance system and / or to a control of the machine or system.

A special feature is in particular that in the first processing module at least one first element of the observation vector is assigned to a first one-dimensional Kohonen map, by means of which a value of said first element is mapped to an output node of said first Kohonen map and thus a first hit output node is obtained and wherein a hit deviation is also obtained as a measure of the deviation of the value of said first element from the first hit output node, the first hit output node and the hit deviation being output data of the first processing module.

With such a one-dimensional Kohonen map, which is assigned to a single element of the observation vector and thus to a single size of a data source, a deviation from learned value ranges of the element can be clearly recognized at least for this size. A fault in a machine or system can also be localized more precisely because it is assigned to this measured variable. Even without further steps, error states can be clearly recognized that generate a measured variable outside the taught-in range or that generate an assignment error for the value of the measured variable to a hit output node of the assigned one-dimensional Kohonen map above a threshold.

Of course, several selected or all available sizes of a system can be recorded and the recorded sizes can be evaluated with their own Kohonen map assigned to the size. Then you can see exactly the size of the deviation and draw conclusions about the type of error.

One or more one-dimensional Kohonen maps can be created with little computing effort train, the training can be done in batch mode, for example, and only the faultless condition of the machine or system needs to be trained.

The Kohonen maps represent value ranges permitted for the faultless condition of the machine or system. A large number of different operating states, that is to say states in an operating cycle of the system, can therefore be used for the training, and a robust representation of the error-free state of the machine or system in the Kohonen maps is obtained. Errors in the correct assignment of operating states are therefore minimized.

Advantageous developments of the present invention are the subject of the dependent claims.

Several hit output nodes of the Kohonen cards can be combined into one hit output node vector. According to a particularly advantageous development, a statistical model is provided in a second processing module, which maps predetermined, known operating states of the machine or system, the second processing module being arranged to receive the hit output node vector from the first processing module, and the second processing module being assigned means which are configured to assign the received hit output node vector to a known operating state with the aid of the statistical model, and which are further configured to form a measure of the assignability of the received hit output node vector to a known operating state, wherein the second processing module is set up in such a way that said measure is supplied to the evaluation module as the date of issue.

The statistical model thus evaluates whether a certain combination of hit output nodes - i.e. a combination of values of different sizes - corresponds to an error-free, error-free state of the machine or system that also occurred during setup or training. This type of two-stage detection of deviations, in which, in the first step, a deviation from individual sizes outside of taught-in value ranges is recognized, and in the second step, a deviation from the combinations of values of different sizes that occurred during setup or training is recognized, enables a very reliable detection of abnormal states of the machine or system, while at the same time the effort for training is low.

On the one hand, as I said, only one-dimensional Kohonen maps are trained for different measurements or other sizes. On the other hand, the statistical model does not need to be trained, but can be calculated using the measurement data.

For this, e.g. the data recorded in the training mode are fed back to the already trained Kohonen maps and the statistical model is used to determine with what probability which vector occurs from hit output nodes.

A statistical model which comprises a Bayesian network is particularly preferred. There the hit output nodes from Kohonen maps can be assigned to rows of network nodes. Weights preferably only network nodes of different rows and indicate the frequency of a joint occurrence of the hit output nodes assigned to these network nodes in the data recorded in the training mode. If, in normal operation, a hit-output node vector cannot be mapped to an existing network of network nodes in the Bayesian network because the network nodes are not present, or because there is no network of this combination of network accounts due to sufficiently strong weights, then a combination of values of the sizes that did not occur in training operation is reliably recognized.

The present invention is described in more detail below with reference to the figures:

• 1 eine hydraulische Anlage 3 mit einer elektrohydraulischen Maschine 24, 26, wobei die Anlage mit einer erfindungsgemäßen Messanordnung 1 ausgestattet ist,
• 2 wesentliche Bestandteile der erfindungsgemäßen Messanordnung in einem Konzeptdiagramm,
• 3 eine detailliertere Darstellung der Verknüpfung von Kohonenkarten und einem Bayes'schen Netz in der erfindungsgemäßen Messanordnung,
• 4 verschiedene Beispieldaten aus einem Trainingsbetrieb der hydraulischen Anlage 3 dargestellt als Druck-Drehzahl Cluster Diagramm,
• 5 Schritte zum Einrichten der erfindungsgemäßen Messanordnung, und
• 6 Schritte für das Betreiben der erfindungsgemäßen Messanordnung zur Erkennung eines anomalen Betriebszustandes.

Show it:

• 1 a hydraulic system 3rd with an electro-hydraulic machine 24th , 26 , the system with a measuring arrangement according to the invention 1 Is provided,
• 2nd essential components of the measuring arrangement according to the invention in a concept diagram,
• 3rd 1 shows a more detailed representation of the linking of Kohonen maps and a Bayesian network in the measurement arrangement according to the invention,
• 4th various sample data from a training session of the hydraulic system 3rd represented as a pressure-speed cluster diagram,
• 5 Steps for setting up the measuring arrangement according to the invention, and
• 6 Steps for operating the measuring arrangement according to the invention for detecting an abnormal operating state.

According to 1 is an electro-hydraulic system 3rd shown schematically with a Plant control 13 which, for example with the help of a control loop, the position or speed or the force of a hydraulic cylinder 18th that is a plant element 16 moves, regulates. For this are in the system control 13 Corresponding target curves for the position / speed of the system element 16 and / or for the pressure in the cylinder 18th available or are fed to it by another superordinate control. At least an actual position 41 ' a displacement sensor 41 of the cylinder 18th is fed to the system control.

The system control 13 controls a subordinate hydraulic control 20 on, e.g. with set pressure 45 , Target position 46 , and / or target speed 47 . The hydraulic control 20 generates a target speed, for example 44 for an electric drive with converter 22 and electric motor 24th . The electric motor 24th drives the hydraulic pump 26 on. Pressure medium, which the hydraulic pump 26 is promoted via hydraulic controls 28 - This can be a valve block with directional valves - the cylinder 18th fed to move it or to exert a force.

An actual pressure value 40 ' is from a pressure sensor 40 recorded and to the hydraulic control 20 as well as an actual speed 43 ' obtained from an angle sensor 43 on the electric motor 24th . These measurement and control variables are from the hydraulic control 20 to the system control 13 passed on. In addition, other measured variables such as a temperature measurement can also be used 42 ' of the hydraulic fluid through a temperature sensor 42 are recorded. In addition, lie in the system control 13 Operating data 48 that indicate, for example, in which part of an operating cycle the system is currently, whether it is in a standby phase, a positioning run or a power-holding phase, etc.

The essential parts of the measuring arrangement according to the invention can be used in the system control 13 be arranged. These parts include an input module 5 , to which the required measured variables, controlled variables, manipulated variables, and / or operating variables are fed. This could be the actual pressure value, for example 40 ' , Actual position 41 ' or as their derivation the speed of the cylinder 18th , the manipulated variable for the speed 44 or the actual speed value 43 ' etc. The input module 5 summarizes the values of the required quantities recorded at an acquisition time in an observation vector 30th together. The observation vector 30th is a data structure that contains the individual values for the sizes as elements 31 to 34 (please refer 2nd ) contains.

Next are a first processing module 7 and a second processing module 8th and an assessment module 9 present in the plant control system based on the observation vector 30th a rating 70 to be created, which evaluates the operating status at the time of the recording and classifies it as abnormal or normal if necessary. This rating 70 is connected to a maintenance system 11 passed on. The maintenance system 11 can mainly be used to notify you of errors in plant operation. It can also be notified of wear conditions and the plant operator with statements about an impending failure of components of the plant 3rd , e.g. the pump 26 , help with preventive or planned maintenance.

In addition, the evaluation 70 already in the system control 13 can be used, for example, to inform the plant operator about error or wear conditions or to control the plant 13 to control other operating modes that may be more gentle on the material.

Due to the modular structure of the measuring arrangement, the individual modules can also be used outside of the system control 13 be arranged. The input module can e.g. B. as a gateway 5 ' be formed, to which the variables described above are supplied, and which the data set of the observation vector 30th over a data communication network 99 to a server 10th sends. On the server 10th are then the processing modules 7 and 8th as well as the assessment module 9 represented by software. The assessment module 9 on the server 10th gives the evaluation in this case 70 again via the data communication network 99 to the maintenance system 11 out. Other scenarios regarding the division of the modules are easily conceivable.

Of course, the modules 7 , 8th and if necessary 9 can also be combined into a single module.

Based on 2nd become essential components of the measuring arrangement 1 explained further. The first processing module 7 has access to several one-dimensional Kohonen maps 81 , 82 , 83 , 84 that are in memory 15 the system control 13 are stored or in a database with the system control 13 connected is. One element each 31 , 32 , 33 , 34 of the observation vector is a separate Kohonen map 81 , 82 , 83 , 84 directly assigned. The processing module uses the respective Kohonen map, for example 81 µm, for the corresponding element, for example 31 of the observation vector a hit output node 81 ' from the output nodes 81.1 , 82.2 , 81.3 etc. of the Kohonen map. The element contains 31 a measured value of a single variable, for example the actual pressure value 40 ' at a certain acquisition time. Be analog for the other element 32 , 33 , 34 of the observation vector 30th Hit exit nodes from the Kohonen maps 82 , 83 , 84 determined and together with the previously mentioned hit output node 81 ' as a hit Exit node vector 50 summarized. In addition, the extent to which the individual values of the elements deviate from the determined hit output nodes is determined. This deviation is used individually or as a collective value as a further issue date 50 ' of the processing module 7 spent.

The second processing module 8th has access to a statistical model 60 which is also in memory 15 or is stored in the database. The second processing module 8th is the hit exit node vector 50 fed. It evaluates using the statistical model 60 the hit exit node vector 50 and determines in particular whether the combination of the hit output nodes from the different Kohonen maps matches a combination that has already occurred in a setup mode or training mode of the measurement arrangement. The date of issue is used as a measure of this correspondence 50 " generated.

The assessment module 9 are at least the output data 50 ' and 50 " of the first processing module 7 and the second processing module 8th fed, i.e. a hit deviation 50 ' in one or more of the Kohonen map 81 , 82 , 83 , 84 and the measure of conformity 50 " from the statistical model. Exceeds the hit deviation 50 ' a predefined first threshold, it is recognized that a single variable is already outside the value ranges learned in set-up mode or training mode and an abnormal operating state is identified in the evaluation 70 signals. In addition, the size in which the deviation has occurred can be signaled, thus making fault diagnosis easier.

Falls below the conformity measure 50 " another predetermined threshold, it is recognized that the combination of values of the detected quantities 40 ' , 41 ' , 42 ' , 43 ' , 44 , 48 , etc. in setup mode or training mode of the measuring arrangement in the normal, error-free operating states did not occur at all or only with a very low frequency. Then there is also an abnormal operating condition in the evaluation 70 signals. The assessment module 9 can also only signal the abnormal operating state if a combination of values of the detected variables occurs more frequently than would be expected according to the set-up mode.

The assessment module 9 can with the help of the processing module 8th Check which size in the combination of values of the recorded sizes deviates the most from combinations that occurred in set-up mode. This size can help facilitate troubleshooting in the evaluation 70 be signaled.

More details on the relationship between the elements 31 , 32 , 33 , 34 of the observation vector 30th , the Kohonen cards 81 , 82 , 83 , etc. and a statistical model implemented as a Bayesian network 60 show the 3rd .

Every element of the observation vector 30th is one, in particular exactly one, of the one-dimensional Kohonen maps 81 , 82 , 83 etc. assigned. Using the Kohonen map 81 becomes a value of the element 31 to one of the output nodes 81.1 , 81.2 , 81.3 , 81.4 etc. mapped. It is the output node whose value is the least of the value of the element 31 deviates, here for example the output node 81.4 . This is the hit output node 81 ' . The same goes for the other elements 32 , 33 , 34 etc. proceed and you get several hit output nodes 81 ' , 82 ' , 83 ' etc., one for each Kohonen map 81 , 82 , 83 , etc. These hit output nodes 81 ' , 82 ' , 83 ' etc. are in the hit output node vector 50 summarized.

For every Kohonen card 81 , 82 , 83 , etc. is a Bayesian network 60 a row 61 ; 62 ; 63 etc. from network nodes 61 ' , 61 ' , 61 '"; 62 ' , 62 ' , 62 '"; 63 ' , 63 ' , 63 '" available. Essentially (more detailed later under setup) is for each output node 81.1 , 81.2 , 81.3 , 81.4 ; 82.1 , 82.2 , 82.3 ; 83.1 , 83.2 , etc. exactly one network node 61 ' , 61 ' , 61 "'; 62 ' , 62 ' , 62 '"; 63 ' , 63 ' , 63 '" available. Weights 65 , 65 ' , 65 " or weight factors network the network nodes 61 ' , 61 ' , 61 '"; 62 ' , 62 ' , 62 '"; 63 ' , 63 ' , 63 '" from different rows with each other but not the network nodes from the same row.

• Zunächst wird die Anlage 3 mit Hilfe der Anlagensteuerung 13 möglichst im Neuzustand oder nach einer Wartung in typischen Betriebszuständen betrieben, siehe Schritt s10 in 5. Es werden z.B. Betriebszustände aus einem zyklischen Betriebsablauf angesteuert. Idealerweise haben diese Betriebszustände eine statische Phase, bei der sich die Soll- und Istwerte der Regelgrößen Druck 40', oder eine Geschwindigkeit des Zylinders 41', 47 nicht oder nur unwesentlich ändern. In solchen statischen Phasen werden mit Hilfe des Eingangsmoduls 5 bzw. 5' z.B. die Werte für den Druck-Istwert und die Stellgröße der für das Halten des Drucks notwendigen Drehzahl 44 erfasst, siehe Schritt s20 in 5. Nur zur Veranschaulichung sind in 4 für verschiedene Betriebszustände 90, 91, 92 und 93 solche erfassten Werte für die Größen Druck 40' und Drehzahl 44 in Cluster Diagrammen dargestellt, wobei jeder Punkt 95 ein Druck-Drehzahl-Wertepaar für eine bestimmte Erfassungszeit darstellt.

Setup operation of the measuring arrangement:

• First, the facility 3rd with the help of the system control 13 if possible operated in new condition or after maintenance in typical operating conditions, see step s10 in 5 . For example, operating states are controlled from a cyclical operating sequence. Ideally, these operating states have a static phase in which the setpoints and actual values of the controlled variables pressure 40 ' , or a speed of the cylinder 41 ' , 47 do not change or change only insignificantly. In such static phases with the help of the input module 5 or. 5 ' eg the values for the actual pressure value and the manipulated variable of the speed required to maintain the pressure 44 recorded, see step p20 in 5 . For illustration purposes only 4th for different operating conditions 90 , 91 , 92 and 93 such captured values for the quantities of pressure 40 ' and speed 44 represented in cluster diagrams, with each point 95 represents a pressure-speed value pair for a specific acquisition time.

The recorded data for the pressure and the speed are used to create a one-dimensional Kohonen map, for example the Kohonen map 81 and the Kohonen map 82 to train for speed and pressure, according to step p30 . The Kohonen maps are initialized with the same weights or equidistant distances between the output nodes and span a range of values that ideally corresponds to the possible range of values of the measured variables pressure and speed measured. The training is carried out until a quantification error or possibly a moving average of the quantification error, that is to say a deviation of a newly added value from output nodes in the Kohonen map, falls below a training threshold value. Then the training of the respective Kohonen map is according to step p40 completed.

The training can be carried out in a so-called batch mode with a previously recorded data record, which contains a large number of measuring points at different acquisition times in static operating states, possibly from a cyclical operating sequence. However, the respective Kohonen map can also be used immediately after the acquisition of the values of the acquired quantities 81 , 82 Be trained sequentially step by step.

The trained Kohonen cards 81 , 82 then contain output nodes 81.1 , 81.2 , 81.3 , 81.4 or. 82.1 , 82.2 , 82.3 which, on the one hand, cluster around values of pressure or speed that are common in training. In addition, the output nodes indicate a frequency of occurrence of these values in the training data.

In step p50 then becomes the statistical model 60 created. No training needs to be carried out for this. Rather, the static model is a Bayesian network with a certain structure that can be easily calculated based on the existing or additionally recorded training data.

The output nodes in the trained Kohonen maps are used for the training data of the recorded variables 81 , 82 as hit output node 81 ' , 82 ' determined and for each acquisition time as a hit output node vector 50 filed in a list. A network node is then created in the Bayesian network for each hit output node in the list. As a simplification, network nodes could only be created for hit output nodes that exceed a certain frequency of occurrence, for example 5%. The hit output nodes and the network nodes created for this purpose are assigned to one another, that is to say represent the same value of the detected variable. The structure of the Bayesian network is designed so that network nodes 61 ' , 61 " , 61 '" who have favourited Output Nodes 81.1 , 81.2 , 81.3 etc. from the same Kohonen map 81 correspond, in rows 61 are arranged. Each network node is assigned a frequency of occurrence of the corresponding output node in the list of hit output nodes generated with the training data.

Between the individual rows 61 , 62 and 63 are the network nodes in pairs by weights 65 , 65 ' , 65 " networked. Network nodes from a row 61 are not themselves networked. The weights are vectors from the hit output nodes contained in the list 50 calculated so that it a frequency of a common occurrence of both output nodes assigned to these network nodes in the two different Kohonen maps 81 , 82 specify, ie this corresponds to a frequency with which the training data the two assigned output nodes 81 ' 82 ' have assigned the two Kohonen maps as hit output nodes at the same time.

The Bayesian network created in this way thus represents a statistic of possible combinations of values of the measured quantities, which are recorded here during the training and occur at the same acquisition time, here pressure and speed.

Of course, instead of or in addition to pressure 40 ' and speed 44 other or different sizes recorded and for training your own one-dimensional Kohonen maps and creating the statistical model 60 to be used.

Measuring operation of the measuring arrangement, see 6 and 3rd .

In normal operation of the plant 3rd the sizes are recorded in a stationary operating phase, see step s100 and an observation vector for each acquisition time from the acquired variables 30th created, see step s110 .

As said, the observation vector contains the recorded values of the quantities as individual elements 31 , 32 , 33 , 34 , e.g. pressure 40 ' and speed 44 .

The values 31 , 32 , etc. from the observation vector of the respective Kohonen map assigned to the sizes 81 , 82 etc. supplied, see step s130 . This makes hit output nodes 81 ' , 82 ' determined that come closest to the values. The determined hit output nodes 81 ' , 82 ' are called hit output node vector 50 to the statistical model 60 passed on, see step p135 . At the same time there will be a hit deviation 50 ' determined that directly the assessment module 9 is fed.

Already based on a hit deviation 50 ' which lies above a predetermined threshold value, for example a threshold value which corresponds to the quantification error in Training, an abnormal operating condition can be detected, see step s140 . For example, it can be recognized whether the actual pressure value 40 ' or the speed control value 44 lies outside a range that was recognized in the training as the range of values for the normal state.

The Hit Output Node Vector 50 with its hit output nodes 81 ' , 82 ' several Kohonen maps 81 , 82 is from the second processing module 8th based on the statistical model 60 checks to see if the combination of values that the hit output node 81 ' , 82 ' represent, have occurred with a certain probability in the training of the normal state. To do this, the hit output node 81 ' , 82 ' assigned network nodes e.g. 61 '" and 62 '" identified, see step s150 . The weight 65 " that the network node 61 '" and 62 '" linked gives the frequency of occurrence of these hit output nodes 81 ' , 82 ' or the assigned network node 61 '" and 62 '" for a normal state, see step p155 . If more than two network nodes in more than two rows 61 , 62 , 63 The weights are identified 65 " and 65 ' offset against each other, e.g. multiplied to determine the total frequency of occurrence.

If it is not possible to identify network nodes because no network node was formed for a hit output node in setup mode or training mode, the frequency of occurrence determined here is zero. The frequency of occurrence is called the issue date 50 " again subjected to an assessment, see also step s140 . If the frequency of occurrence of the detected combination of values of different sizes exceeds a predetermined threshold, the normal state is recognized and as an evaluation 70 spent. Otherwise, that is, if the specified threshold is undershot, an abnormal operating state is used as an evaluation 70 reported.

This can be used, for example, to detect when in a pressure control mode with a pressure regulator in the system control 13 a print 40 ' as a controlled variable keeps constant at a certain value, eg an unusually high value for the manipulated variable speed 44 occurs. Even if the higher speed value is within a valid normal range in other operating states, the deviation of the speed is combined with a predetermined specific pressure actual value 40 ' detected and reported as an abnormal operating condition. The higher speed may be necessary due to leaks, for example, in order to achieve the pressure value specified by the controller despite the leak.

Reference list

1

Measuring arrangement

3rd

investment

5, 5 '

Input module in machine control or gateway

7

First processing module

8th

Second processing module

9

Assessment module

10th

server

11

Maintenance system

13

Plant control

15

Memory or database

16

Movable system element

18th

Hydraulic consumer, cylinder

20

Subordinate control hydraulic system, short hydraulic control

22

frequency converter

24th

Electric motor

26

Hydraulic pump

28

Hydraulic controls

30th

Observation vector with elements 31 , 32 , 33 , 34

40, 40 '

Pressure sensor, pressure reading

41, 41 '

Position sensor, actual position

42, 42 '

Temperature sensor, temperature reading

43, 43 '

Angle sensor, actual speed

44

Target speed

45

Target pressure

46

Target position

47

Target speed

48

Operating status data

50

Output data of the first processing module, e.g. hit Output node, vector from hit output node

50 '

Output data of the first processing module, hit deviation

50 "

Output data second processing module

60

Bayesian network

61

First row of network nodes 61 ' , 61 ' , 61 '"

62

Second row of network nodes 62 ' , 62 ' , 62 '"

63

Third row of network nodes 63 ' , 63 ' , 63 '"

65, 65 ', 65 "

Weights of the Bayesian network

70

rating

81

One-dimensional Kohonen map with output nodes 81.1 , 81.2 , 81.3 , 81.4 etc.

82

One-dimensional Kohonen map with output nodes 82.1 , 82.2 , 82.3 etc.

83

One-dimensional Kohonen map with output nodes 83.1 , 83.2 etc.

84

One-dimensional Kohonen map

81 ', 82', 83 ',

Hit output node

84 '90

Pressure-speed cluster Diagram of actual pressure values and actual speed values of a first stationary operating state

91

Pressure-speed cluster Diagram of actual pressure values and actual speed values of a second stationary operating state

92

Pressure-speed cluster Diagram of actual pressure values and actual speed values of a third stationary operating state

93

Pressure-speed cluster Diagram of actual pressure values and actual speed values of a fourth steady state

95

(p, n) measuring points from the actual pressure value and the actual speed value recorded at the same acquisition time

99

Data communication network

s10

Operate step

p20

Step acquisition of training data

p30

Exercise step

p40

Completing the step

p50

Create a statistical model

s100

Acquisition of data in measurement mode

s110

Create an observation vector

s130

Assign in a Kohonen map

p135

Output a hit output node

s140

Assess an operational status

s150

Assign in a statistical model

p155

Form a measure of assignability

QUOTES INCLUDE IN THE DESCRIPTION

This list of documents listed by the applicant has been generated automatically and is only included for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent literature cited

• "Krogerus, T .; Vilenius, J .; Liimatainen, J .; Koskinen, K.T. 2006. Self-organizing maps with unsupervised learning for condition monitoring of fluid power systems. Fluid Power for Mobile, In-Plant, Field and Manufacturing. SAE SP-2054 pp. 43-51 "[0002]

Claims (22)

Measuring arrangement (1) for the automated detection of operating states of a machine or system (3), in particular for the automatic detection of abnormal operating states, with an input module (5, 5 ') for acquiring data from several data sources (40, 41, 42, 43, 13) of the machine or system (3) - the data containing status information and / or controlled variables and / or manipulated variables and / or represent other measured variables of the machine or system (3) - and for creating an observation vector (30) for a detection time with elements (31, 32, 33, 34), the data from the data sources (40, 41, 42, 43, 13) at the time of acquisition, with a first processing module (7), to which the observation vector (30) is fed and in which an artificial neural network is created, the artificial neural network being set up to use the observation vector (30) supplied to it to set at least one output date ( 50, 50 ', 50 "), which characterizes an operating state of the machine or system (3), and with an evaluation module (9) which evaluates the operating state on the basis of the date of issue (50, 50 ', 50 ") and evaluates (70) to a maintenance system (11) and / or to a controller (13) of the machine or system ( 3) outputs, where in the first processing module (7) at least one first element (31) of the observation vector (30) is assigned to a first one-dimensional Kohonen map (81), by means of which a value of said first element (31) is assigned to an output node (81.1 - 81.3) of said first Kohonen map (81) is mapped and thus a first hit output node (81 ') is obtained and furthermore a hit deviation (50') is obtained as a measure of the deviation of the value of said first element (31) from the first hit Output node (81 '), the first hit output node (81') and the hit deviation (50 ') being output data (50, 50') of the first processing module (7). Measurement arrangement according to Claim 1 , characterized in that the evaluation module (9) is set up to output an abnormal operating state as evaluation (70) if the hit deviation (50 ') exceeds a predetermined first threshold value. Measurement arrangement according to Claim 1 or 2nd , characterized in that the first processing module (7) is set up in such a way that a second element (32) of the observation vector is assigned to its own second one-dimensional Kohonen map (82), by means of which a value of the second element (32) is assigned to a second hit Output node (82 ') is shown, and in particular the first processing module (7) creates a hit output node vector (50) which comprises at least the first hit output node (81') and the second hit output node (82 ') . Measurement arrangement according to Claim 3 , characterized in that a second processing module (8) is provided, which comprises a statistical model (60), which depicts predetermined, known operating states of the machine or system (3), the second processing module (8) being arranged for this purpose by the first processing module (7) receive the hit output node vector (50), and wherein the second processing module (8) comprises assignment means which are set up to receive the received hit output node vector (50) using the statistical model (60) of a known one Assign operating state, and which are further set up to form a measure for an assignability of the received hit output node vector (50) to a known operating state, the second processing module (8) being set up in such a way that said measure is used as the output date (50 ") is fed to the evaluation module (9). Measurement arrangement according to Claim 4 , characterized in that the statistical model (60) comprises a stored table of hit output node vectors of predetermined known operating states of the machine or system (3), each of these hit output node vectors in the table being a measure of a frequency of occurrence with respect to a is assigned to the predetermined known operating states. Measurement arrangement according to Claim 4 or 5 , characterized in that the statistical model comprises a Bayesian network (60). Measurement arrangement according to Claim 6 , characterized in that the Bayesian network comprises a multiplicity of event nodes (61 '- 63 "') which indicate a frequency of occurrence of values from the data sources (40, 41, 42, 43, 13), in particular weights (65) being provided with reference to at least two of the event nodes (61 '- 63'"), which a common frequency of occurrence of the two event nodes (61 '- 63'"" ) represent represented values, and that the assignment means are set up to assign hit output nodes (81 ', 82', 83 ', 84') of the received hit output node vector (50) to the event nodes (61 '- 63'") . Measurement arrangement according to Claim 7 , characterized in that in the Bayesian network (60) only event nodes (61 '- 63'") are created that exceed a specified frequency of occurrence. Measuring arrangement according to one of the Claims 6 to 8th , characterized in that one event node (61 '- 63'") of the Bayesian network (60) is directly assigned to one output node (81.1 - 83.2) of the Kohonen map (81, 82, 83, 84). Measurement arrangement according to Claim 9 , characterized in that in the Bayesian network (60) event nodes (61 '- 63 "') are arranged in rows and columns, the output nodes (81.1, 81.2, 81.3) of a Kohonen card (81) of a series of event nodes ( 63 '- 63'") are assigned, and the weights (65) only interconnect the event nodes (61 '- 63'") of different rows. Method for setting up a measuring arrangement according to one of the Claims 1 to 10th , with the following steps - operating (s10) the machine or system (3) in a predetermined normal operating state, - collecting (s20) data from several data sources (40, 41, 42, 43, 13) of the machine or system (3 ) during the specified normal operating state at several acquisition times, - training (s30) the first one-dimensional Kohonen map (81) with the acquired data from a first data source (40) of the plurality of data sources (40, 41, 42, 43, 13), - completing (s40) of the training if a quantization error of data supplied during the training of the first Kohonen card (81) falls below a second threshold value and / or if a predetermined number of data has been supplied to the first Kohonen card (81). Setup procedure according to Claim 11 , the step training (s30) comprising training at least one second one-dimensional Kohonen map (82) with the acquired data from a second data source (43) of the plurality of data sources (40, 41, 42, 43, 13). Setup procedure according to Claim 11 or 12th , in connection with a measuring arrangement according to at least Claim 4 , with the additional step of creating (s50) the statistical model (60) - in particular a Bayesian network - with the aid of the output nodes (81.1 - 83.2) obtained from the training in the Kohonen maps (81, 82, 83, 84) and the data already recorded or further data recorded in the predetermined normal operating state. Setup procedure according to Claim 13 , wherein in the step of creating (s50) the statistical model (60) a Bayesian network is created, wherein for each output node in the Kohonen maps (81, 82, 83, 84), which has an occurrence frequency that lies above a third threshold value , an event node (61 '- 63 "') is created in the Bayesian network (60). Setup procedure according to one of the Claims 12 to 14 , in the step of creating (s50) the statistical model (60) for each Kohonen map (81, 82, 83, 84) a series (61, 62, 63) of event nodes (61 '-61''', 62 '- 62 "`, 63 '-63'") is created in the Bayesian network (60) to which the output nodes (81.1, 81.2, 81.3, 82.1, 82.2, 82.3, 83.1, 83.2) of the respective Kohonen map (81, 82, 83) are assigned, with weights (65, 65 ', 65 ") being provided between event nodes of different rows, a weight (65, 65', 65") indicating a common frequency of occurrence for obtaining output nodes from different Kohonen cards (81, 82 , 83, 84) as hit output nodes (81 ', 82', 83 ', 84`) for the specified normal operating state. Method for operating a measuring arrangement according to one of the Claims 1 to 10th which according to one of the procedures Claim 11 to 15 is set up, with the steps - acquiring (s100) data from several data sources (40, 41, 42, 43, 13) of the machine or system (3) during an operating state to be recognized at an acquisition time and creating (s110) an observation vector ( 30) with the data of this acquisition time, - assigning (s130) a first element (31) of the observation vector (30) to an output node (81.1 -81.3) of the first Kohonen map (81), whereby a first hit output node (81 ') is obtained and wherein a hit deviation (50 ') is obtained as a measure of the deviation of the value of said first element (31) from the first hit output node (81'), - output (s135) of the first hit output node (81 ') and the hit deviation (50 ') as output data (50, 50') - evaluating (s140) the operating state to be recognized with the aid of the output data (50, 50 '). Procedure for operating according to Claim 16 , wherein in step evaluating (s140) the operating state to be recognized is output as an abnormal operating state if the hit deviation (50 ') exceeds the predetermined first threshold value. Procedure for operating according to Claim 16 or 17th , in the step of assigning (s130) to a second element (32) of the observation vector a separate second one-dimensional Kohonen map (82) is assigned, whereby a value of the second element (32) is mapped to a second hit output node (82 '), and being a hit Output node vector (50) is created, which comprises at least the first hit output node (81 ') and the second hit output node (82'). Procedure for operating according to Claim 18 , with the further steps - assignment (s150) of the hit output node vector (50) to a known operating state with the aid of a statistical model (60), which depicts predefined known operating states of the machine or system (3), and - forming (s155 ) a measure of the assignability of the hit output node vector (50) to a known operating state, the said measure of assignability (50 ") and in particular the assigned operating state being used for evaluating (s140) the operating state to be recognized. Procedure for operating according to Claim 19 , wherein the statistical model comprises a Bayesian network (60) with a multiplicity of event nodes (61 '- 63 "'), which indicate a frequency of occurrence of values recorded in a set-up mode in known operating states of the machine or system (3) Data sources, wherein weights (65) are provided with reference to at least two of the event nodes (61 '- 63'"), which indicate a common frequency of occurrence of the values represented by the two event nodes (61 '- 63'"), and in which Step assignment (s150) hit output nodes (81 ', 82', 83 ', 84') of the hit output node vector (50) are assigned to the event nodes (61 '- 63'"). Computer, server or system control, which with the aid of a computer program product for executing the method according to at least one of the Claims 11 to 20 is set up. Computer program product which is used to carry out the method according to at least one of the Claims 11 to 20 is set up if it is executed on a system controller or on a server or on a computer.

Images