ANOMALY DETECTION IN A PNEUMATIC SYSTEM

Abstract

An error detection and localisation in a pneumatic system and in particular an error detection module includes a read-in interface for reading-in digital signals from the automation plant, a first processor unit designed to execute a detection algorithm for calculating an anomaly score for the automation plant on the basis of the set of read-in signals, a second processor unit which is designedin the event that the anomaly score calculated with the first processor unit indicates an anomalyto perform a machine localisation method for localising the error, wherein the machine localisation method has been trained in a training phase in order to calculate and as a result provide, on the basis of a detected circuit diagram of the automation plant with respect to the calculated anomaly score, probabilities of possible causes of error in relation to individual components of the automation plant.

Claims

1. An error detection module for detecting and evaluating anomalies in automation plants, in particular in a pneumatic automation plant, comprising: a read-in interface for reading-in digital signals from the automation plant; a first processor unit which is designed to execute a detection algorithm for calculating an anomaly score for the automation plant on the basis of the set of read-in signals; and a second processor unit which is designedin the event that the anomaly score calculated with the first processor unit indicates an anomalyto perform a machine localisation method for localising the error, wherein the machine localisation method has been trained in a training phase in order to calculate and provide as a result, on the basis of a detected circuit diagram of the automation plant with respect to the calculated anomaly score, probabilities of possible causes of error in relation to individual components of the automation plant.

2. The error detection module as claimed in claim 1, wherein the first processor unit is implemented on a different device than the second processor unit and in particular on a control unit.

3. The error detection module as claimed in claim 1, wherein the second processor unit or a further processor unit which is designed to generate a model comprises a circuit diagram read-in interface for reading-in a circuit diagram for the automation plant in digital form.

4. The error detection module as claimed in claim 1, further comprising a configuration interface as a front-end for configuring and/or training the model.

5. The error detection module as claimed in claim 1, wherein the automation plant comprises a pneumatic system having a pneumatic drive, wherein a plurality of drives and/or actuators are connected to a valve and a plurality of valves are arranged on a valve cluster and a plurality of valve clusters are connected to a supply unit.

6. A method for detecting and evaluating anomalies in an automation plant, in particular in a pneumatic automation plant, the method comprising the method steps of: reading-in digital signals of the automation plant via a read-in interface; executing a detection algorithm for calculating an anomaly score for the automation plant on the basis of the set of read-in signals; in the event that the calculated anomaly score indicates an anomaly: triggering a machine localisation method for localising the error, wherein the machine localisation method has been trained in a training phase in order to calculate and as a result provide, on the basis of a detected circuit diagram of the automation plant with respect to the calculated anomaly score, probabilities of possible causes of error in relation to individual components of the automation plant.

7. The method as claimed in claim 6, wherein the detection algorithm for calculating the anomaly score is a pattern recognition algorithm or is effected by accessing a memory, in which a trained detection model is stored.

8. The method as claimed in claim 6, wherein the machine localisation method calculates probabilities of possible causes of error in relation to individual sub-components of a component.

9. The method as claimed in claim 6, wherein the signals come from at least two different digital sensors and a switching command and represent points in time of two final position switches on a cylinder of the pneumatic system and a valve switching point in time and/or wherein four time intervals are calculated from the three digital signals: reaction time during extension of the cylinder; travel time during retraction of the cylinder; reaction time during retraction of the cylinder; and travel time during retraction of the cylinder.

10. The method as claimed in claim 6, wherein the signals of two final position switches are read-in, comprise a valve switching point in time signal and/or a pressure signal and/or a flow signal.

11. The method as claimed in claim 6, wherein, after calculating the reaction time and travel time during extension and retraction of the cylinder, the detection algorithm performs the processing steps of: feature extraction; Z-score normalisation; principal component analysis; classification, in particular using K-means; logistical function; and/or smoothing.

12. The method as claimed in claim 6, wherein the detection algorithm comprises as a result an anomaly score and a sensor relevance value.

13. The method as claimed in claim 6, wherein the machine localisation method comprises a decision tree method, and wherein the decision tree is calculated on the basis of the detected circuit diagram or comprises a Bayesian network method.

14. The method as claimed in claim 6, wherein the machine localisation method extracts, from the detected circuit diagram, data relations between data sets which are based upon read-in signals.

15. The method as claimed in claim 6, wherein the result of the machine localisation method comprises an error probability value for all components and/or all sub-components of the components and wherein the method further performs the processing steps of: aggregating all error probability values; accessing a memory, in which a system of rules is stored for localising the error in relation to individual components and/or sub-components of the automation plant.

16. An error detection system for detecting and evaluating anomalies in automation plants, in particular in a pneumatic system, comprising: an error detection module as claimed in any one of the claims directed to the error detection module; a gateway; and a cloud-based server which is connected to the error detection module via a web interface.

17. A computer program with computer program code for carrying out all method steps of the method as claimed in claim 6 when the computer program is executed on a computer.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The disclosure will now be described with reference to the drawings wherein:

[0070] FIG. 1 shows an overview of the inventive error detection system comprising an error detection module;

[0071] FIG. 2 shows an exemplary embodiment of an error detection module which is an alternative to the illustration in FIG. 1;

[0072] FIG. 3 shows a further schematic view of an error detection module comprising a cloud-based server and further component parts;

[0073] FIG. 4 shows an alternative, schematically illustrated design of the error detection module;

[0074] FIG. 5 shows a flow diagram of method steps of an error detection method according to an exemplary embodiment of the disclosure; and

[0075] FIG. 6 shows a schematic illustration of an error detection system comprising further component parts according to an exemplary embodiment of the disclosure.

DESCRIPTION OF EXEMPLARY EMBODIMENTS

[0076] In the following detailed description of the figures, exemplary embodiments, which are to be understood to be non-limiting, together with the features and further advantages thereof will be discussed with the aid of the drawing.

[0077] The disclosure serves to technically monitor a pneumatic system as an example of an automation system or plant comprising various field devices (hereinafter also referred to as components) which are controlled via a control device (e.g., PLC). In particular, errors are to be recognised in good time and typically at a point in time before the respective component fails or causes an error in the plant. To this end, an error detection module, explained in greater detail hereinafter in relation to FIG. 1, is to be used.

[0078] The disclosure has the advantage that early error detection for complex, multiple-componenttypically pneumaticautomation plants becomes possible although only very little measurement data are available and which can be operated quasi with a minimal sensor system. In particular, it is possible to provide a result with error localisation although only two digital sensors and one switching command are used, in particular for detecting the points in time of two final position sensors on one cylinder and one sensor for detecting the valve switching point in time. This has the advantage that anomaly detection also becomes possible in such plants, in which only the actuator is equipped with a sensor system (e.g., final position sensors). The method presented here was based upon a model, in which at least these signals are taken into consideration. Optionally, still further signals, such as pressure signals and/or flow signals or other signals of sensors internal to the valve are taken into consideration which are detected in the pressure supply and/or in the valve. With the aid of the detection algorithm, deviations or changes from the correct or typical reaction behaviour of the pneumatic plant are now detected automatically and in real time, such as e.g. the time between valve switching and leaving final position 1 and travel time (final position 1 to final position 2). Moreover, in principle the time between sending the control command to the physical switching of the valve is measured and learned. In one advantageous development, an additional valve-internal sensor can be formed which detects if the valve has switched. The same applies for the return movement of the valve. The measurement variables and the patterns resulting therefrom are learned during the good operation (i.e., during error-free operation). Error images show characteristic patterns which are used in accordance with the disclosure for anomaly detection and for error localisation. Moreover, the circuit diagram of the pneumatic system is available in a digital pneumatic circuit diagram which is read-in, e.g., from a Fluid Draw or Eplan or Automation ML file, and is used for constructing decision logic. If, by means of the detection algorithm, a deviation from the GOOD pattern is detected, error localisation can be provided in a second step by applying a machine localisation method. To this end, a logic circuit comprising implemented decision logic can be used, e.g., using a decision tree or Bayesian networks or other machine learning methods.

[0079] The background of the solution proposed in this case is that the time behaviour of a tensioning or clamping system (e.g., automobile manufacture, vehicle body manufacture)consisting of a valve, hose system and clamping fixtureschanges as wear increases. A test arrangement is created in order to identify whether and how manipulations performed on the pneumatic system affect the time behaviour. Variations and manipulations have been performed on the pneumatic system in a targeted manner. This comprises friction and leakage at the clamping fixture and at the valve and changes in the length of the lever arm, the hose length between the valve and clamping fixture and a variation in the supply pressure. The closing time and the delay time have been recorded as the cylinder is opening and closing. As a result of the tests conducted by the applicant, it can be stated that a change in friction, leakage and supply pressure of the clamping fixture affect the delay and closing times which can be derived from the final position switch signals. The results from the test arrangement influence the configuration of the error localisation model, in which in a first stage the error is localised in relation to individual components of the plant and in a second stage the error is localised in relation to individual sub-components of the component. It is possible to unequivocally identify which type of malfunction is present. Therefore, it is possible to contain and in particular localise the error on the basis of the (three) digital signals.

[0080] FIG. 1 schematically illustrates the error detection module FM. It comprises on the side of the automation plant AA thetypically pneumaticcomponents K, e.g., valve clusters or valve disks, wherein a valve cluster comprises, in turn, a plurality of valves having clamping fixture/cylinder units and/or further pneumatic actuators (e.g., pneumatic drives etc.) and sensors as well as a pressure supply. Furthermore, a controller is provided which can be designed as a programmable logic controller which can also be designated as PLC. The components K are designed having sensors S which serve to detect digital signals or switching commands to a valve. A first component K1 comprises at least one sensor unit S1 for detecting three digital signals, a second component K2 comprises, in turn, a sensor unit S2 for detecting at least three signals etc.

[0081] As shown in FIG. 1, further sensors S3 can also send signals (e.g., pressure signals) to the PLC. The controller PLC receives the digital signals via a read-in interface I1 and furthermore is designed having a first processor unit P1 which serves to execute a detection algorithm on the basis of the detected or read-in signals. The detection algorithm serves to calculate an anomaly score for the automation plant AA on the basis of the set of detected or read-in signals. The calculated anomaly scores can be transferred to an IoT gateway GW via a data interface (e.g., OPC-UA). The calculated anomaly scores and/or the detected signals are communicated via a second interface I2 to a second processor unit P2 whichif the anomaly score calculated with the first processor unit indicates an anomalycan be designed to perform a machine localisation method S34 (which is described in greater detail below with reference to FIG. 5) to localise the error in order, with respect to the anomaly score, to calculate and provide as a result probabilities of possible causes of error in relation to individual components K of the automation plant AA.

[0082] In the example illustrated in FIG. 1, the first processor unit P1 is implemented on a different device than the second processor unit P2. The first processor unit P1 can be formed on the control unit PLC and the second processor unit P2 can be formed, e.g., on a gateway node GW (or gateway for short). In order to perform the machine localisation method, the second processor unit P2 accesses a memory MEM, in which a trained model is stored. The second processor unit P2 receives a circuit diagram of the pneumatic plant AA via a circuit diagram read-in interface 13. The circuit diagram is provided in digital form and contains information relating to the structure of the plant AA and relating to the functionality (in particular switching points in time of the valves etc.).

[0083] In the exemplary embodiment shown in FIG. 1, a separate gateway GW is provided which serves as an intermediary between, on the one hand, the plant AA with the components K and with the programmable logic controller PLC and, on the other hand, the server SV. The gateway GW can be implemented, e.g., in a superordinate management system of the plant AA and/or can be allocated to the plant AA (e.g. in the same security domain as the plant). A third processor unit P3 can be formed on the server SV in order to be able to perform e.g. the machine localisation method on a cloud-based server.

[0084] As schematically indicated in FIG. 1, it is fundamentally possible for the first processor unit P1 to send the locally calculated anomaly scores quasi as an intermediate result to the second processor unit P2 (solid arrow). Alternatively or cumulatively, the detected signals can also be communicated to the second processor unit P2. This can be effected either directly from the sensor S and/or from the component K (both are illustrated in FIG. 1 by a dashed line) and/or from the controller PLC.

[0085] FIG. 2 shows an alternative exemplary embodiment, in which the gateway GW comprises both the second and the first processor units P2, P1. The components K send their three digital signals to the controller PLC which has then sent the signals via the network connection (second interface 12) to the second processor unit P2. Alternatively, the components can send the locally detected signals directly to the second processor unit P2 (without the bypass via the PLC). It is even feasible that the sensors themselves can be designed having a further network interface in order to communicate the data.

[0086] FIG. 3 shows an exemplary embodiment using a cloud-based server SV. The sensor data are then detected on the components K of the pneumatic plant AA. The first processor unit P1 can now be formed either locally in the controller PLC or on one of the IoT gateway nodes GW which is allocated to the plant and can be designed as an edge computer. The gateway GW exchanges data via an Internet protocol-based data connection (e.g., https etc.) with the server SV, on which the second processor unit P2 is formed which is designed to perform the machine localisation method. The learned model can be held in the memory MEM of the server SV. Therefore, it becomes possible to use the higher computing resources (and memory resources) of the server for localising the error and for calculating the result.

[0087] As the above examples are intended to show, the functionality of the error detection module FM can also be effected in a distributed manner with the following two aspects: detection algorithm S2 and machine localisation method S34.

[0088] In other words, the first processor unit P1 and the second processor unit P2 can be implemented on different computer-based entities. It is also possible to design a further processor unit which serves to configure the model or to train the localisation method on the basis of training data. The training data can comprise patterns of signal combinations in GOOD cases (error-free operation of the plant).

[0089] As illustrated in FIG. 4, it is typical that the detection algorithm S2 is executed as locally as possible, in the proximity of the generated data, typically in the controller PLC and the machine localisation method S34 can be performed on an entity which provides sufficient resources, typically performed on the server SV. Only one client for model checking for the machine localisation method S34 can then be installed on the gateway GW and so the computationally intensive processes can be performed on the server SV and only the result is output to configurable entities, in particular to the gateway GW and optionally to the components K of the plant AA and/or to the controller PLC. The outputting can be effected via an output interface AS.

[0090] FIG. 5 shows a flow diagram of an error detection method. After the start, the digital signals are read-in in step S1. In step S2, the detection algorithm is executed on the or with the read-in signals. It calculates an anomaly score and a sensor relevant value as an intermediate result. The intermediate result thus represents whether an anomaly is present in the plant AA or not. Depending upon the result, the method branches to different calculation cases, as can be seen in

[0091] FIG. 5. If there is no anomaly present, the plant appears to function as alwaysi.e., in an error-free manner. The method can be ended or restarted with an EXIT. Otherwise (when an anomaly or deviation is detected), in step 34 a machine localisation method is performed which has been trained in a training phase in order to calculate probabilities of possible causes of error on the basis of a digitally or manually detected circuit diagram of the automation plant AA with respect to the respectively calculated anomaly score. The machine localisation method can comprise two stages. In the first step S3, localisation of the error is calculated at component level (e.g., error in clamping fixture X or valve Y) and in the second step S4 localisation of the error is calculated at sub-component level. In the second step S4, it is analysed where the error within the component identified as defective can be localised. The machine localisation method can be implemented as an algorithm which is executed taking into consideration the information of the detected circuit diagram (design, architecture and structure of the circuit and switching points in time). As shown above, the functionality of the algorithm can also be implemented on other devices or servers SV.

[0092] FIG. 6 is a further structural architectural image of an error detection system comprising a first processor unit P1 which in this case is implemented on the controller PLC, and comprising the second processor unit P2 which is implemented on the server SV which exchanges data with the gateway GW via a data connection. In addition, a configuration interface Config-UI can be provided, by means of which the machine localisation method and in particular the algorithms S3, S4 can be configured. The configuration interface Contis-UI is typically cloud-based or can also be provided locally as a computer program. The configuration interface Config-UI can comprise user interface elements, such as dashboards. In this case, a version of the learned model (e.g., constructed decision tree) can also be installed having a training master as an application for configuring the learning phase for the model or for generating the decision tree and having a scoring master as an application for calculating the anomaly score according to a further option. A suite of applications for error detection and localisation can be installed on the server SV (e.g., industry PC). In particular, a runtime environment (e.g., Java Runtime environment) of the trained model is implemented which is synchronised with the configuration interface Config-UI and interacts with the gateway GW typically via https/REST-Upload requests. The read-in signals are then sent via the gateway GW to the server for the purpose of error detection and localisation.

[0093] In one exemplary embodiment, a further processor unit which in FIG. 6 is defined as the third processor unit P3 can be provided and serves to generate the model for the machine localisation method S34. The user has the option of adjusting settings via the configuration interface Config-UI. The functionality for generating the model can also be implemented on the server SV.

[0094] In this exemplary embodiment, the IoT gateway node GW can be designed having a client for the machine localisation method. The client/gateway can be positioned in the field in the vicinity of the plant. The gateway GW can have a browsing functionality which can be used for paging through and inspecting the anomaly scores communicated by the controller PLC. Furthermore, the gateway node GW can have a proxy for the algorithm provided thereon which can be operated in the cloud (e.g. on the server SV) and a proxy for an automation suite with further applications and programs as a PC application. The functionality of the automation suite is the same as the functionality of the cloud. Furthermore, the gateway GW can have a circular buffer for intermediate storage of the data formed thereon, as well as a lite-version of the trained model (for performing the machine localisation method) for the purposes of persistence, configuration, license management and further functionalities in conjunction with the machine localisation method. Fundamentally, depending upon the configuration the gateway GW can have still further programs installed thereon which, inter alia, can also run in the background and can provide specific services. User interactions take place typically only indirectly, e.g. via signals, pipes and above all (network) sockets.

[0095] In one test, 6 pneumatic clamping fixtures were operated continuously for a runtime extended in comparison with normal operation, or for cycle time reduced in comparison with normal operation, over a long period of time until wear occurs. Indicators of wear could be seen in the data in all clamping fixtures 2 weeks prior to failure. Failures and induced error cases can be detected in accordance with the disclosure by means of the machine localisation method or trained model and automated process monitoring is possible.

[0096] Finally, it is noted that the description of the disclosure and the exemplary embodiments are fundamentally to be understood to be non-limiting with respect to a specific physical implementation of the disclosure. All features explained and illustrated in conjunction with individual embodiments of the disclosure can be provided in a different combination in the subject matter in accordance with the disclosure in order to achieve the advantageous effects thereof at the same time.

[0097] The scope of protection of the present disclosure is set by the following claims and is not limited by the features explained in the description or shown in the figures.

[0098] For a person skilled in the art, it is in particular obvious that the disclosure can be used not just for pneumatic plants but also for other hydraulic plants or other fluid-technology systems or electrical spindles. Furthermore, the component parts of the error detection module can be distributed over a plurality of physical products.