DE102021116393A1 - Computer-implemented method for executing a controller with anomaly detection - Google Patents

Abstract

Computer-implemented method with anomaly detection using a training method (T), using a cloud(7) or a central computer for executing better implemented ML/AI training algorithms, and using the execution of the training algorithms to create a parameter file/matrix is determined, while there is also provision for carrying out an execution method (A) in which, in the event that the data is outside the predetermined range, this state is recognized as an unknown state, and the unknown state is transmitted to a control authority which communicates to the unknown State assigns one of the following two state variables: special or unknown operating state (BZ), in the case of the actually unknown operating state, the training method (T) is started with the data and a parameter file / matrix (M) is generated, while again in the case of a special operating state a message will be issued.

Description

The invention relates to a computer-implemented method for executing a control with an execution phase and a training phase.

It is known from the state of the art that, especially in connection with measures to establish so-called Industry 4.0 standards, more and more electronics are being used, artificial intelligence (abbreviation: KI or AI) or machine learning (abbreviation: ML ) uses. These algorithms ensure that machine commands or machine parameters are derived from the behavior of the user or from process-typical parameters with the help of algorithms and are thus taught in artificially.

The object of the invention is to propose a computer-implemented method for executing a control, with which an improved implementation of methods of artificial intelligence or machine learning is made possible. The object is achieved by the features of claim 1, based on a method of the type mentioned at the outset. Advantageous embodiments and developments of the invention are possible as a result of the measures specified in the dependent claims.

Accordingly, the method according to the invention is designed to carry out a control process in the form of a controller for a machine or for a device. The method usually requires a computer for control and processing, which can also be integrated into the machine using processors so that the corresponding algorithms can be executed. The control process can also include regulation processes in which sensor data is used in order to carry out the control or to generate the appropriate control command. Other embedded devices, in particular any field devices, can also manage with the use of microcontrollers, FPGAs or the like. The computer implementation according to the invention relates in particular to the use of a cloud comprising a computer in which more complex training methods for learning and applying AI methods to generate result algorithms can be carried out, for example using a support running in the cloud -Vector Machine (SVM) or an SVC learning algorithm. The learning algorithm can be an AI network (usually a neural network) or an ML algorithm (e.g. SVM, nearest neighbor, ...). SVC is a sub-algorithm of SVM for performing cluster analysis.

Methods of ML (machine learning) or AI (artificial intelligence) according to the invention can thus be used in an advantageous manner. As a rule, AI methods require more computing power, but more information can be stored than with ML. Conventional industrial systems consist of components such as a controller and field devices that are connected via a fieldbus and sensors/actuators that send data to the field devices or correspondingly physically convert control signals.

According to the current state of the art, a permanently programmed algorithm (unchangeable) runs in the conventional field device and carries out control and monitoring tasks.

According to the invention, an AI and/or ML algorithm is now executed in the field device, which can learn the algorithm in the cloud and/or in the central computer.

In general, the control device accesses field devices that include sensors or at least interact with sensors whose measurement data are in turn required for the control. This data can be collected and transmitted to the control device. Field devices are generally devices outside of the control cabinet of a plant or outside of the actual control system, which include sensors in particular. For example, it can be actuators such as actuators, moveable valves or the like. In one embodiment, the field devices can be connected to the control device via a bus, in particular a field bus.

The method according to the invention also uses algorithms with two different phases, a training phase and an execution phase.

The execution phase corresponds to an evaluation and control logic. The methods of artificial intelligence and machine learning are basically not carried out in the execution phase, but their result algorithms are used, since they are limited to the resulting logic circuit. When processing in the context of the execution phase, there is a fixed number of inputs and outputs as well as predetermined evaluation and processing processes, which are based, for example, on a predefined logic circuit. The execution phase is usually carried out on the field devices, possibly also on the control device. However, the field devices usually do not have high computing power because they are often not equipped with high-performance processors. The processing of a Logic circuit does not require high computing power. Sometimes the evaluation can also be done with small microcontrollers.

In the training phase or the training method, as a rule, very large amounts of data have to be processed in order to carry out an evaluation, or more precisely: a learning phase or a training of the network, with the aid of artificial intelligence. With machine learning, a general rule of action is derived from given individual examples. Massive computing power (multi-core processors, GPU, mainframe, etc.) is required to execute these algorithms and process a correspondingly large amount of data in order to carry out this process in a reasonable time that is acceptable in industrial operation.

A cloud is provided for executing the training method. It is therefore not necessary to make this corresponding computing power available on site at the plant or machine that is to be controlled. On the contrary, it can advantageously also be outsourced to third-party providers who make a corresponding computer system available via the cloud. The AI or ML training process can be carried out via the cloud with high computing power. The result of the training process is made available as a parameter file or matrix on the field device for execution.

One exemplary embodiment is distinguished, among other things, by the fact that a data memory is used in which the data is initially collected and saved. The data can be transferred from the data storage to an external cloud.

The central collection of the data or measurement data in a data memory makes it possible for not all data to be sent directly to the cloud. This caching or collection can relieve the amount of data transfer when only a very limited bandwidth is available for transfer to the cloud.

According to one embodiment of the invention, however, the cloud can directly request data from the data storage device. The data memory is used for backup and temporary buffering at the same time. For security reasons, it can be advantageous not to immediately transfer all data to the cloud. Therefore, only the data required to process the AI or ML algorithm is moved to the cloud. In addition, processing time can be saved, depending on how quickly the transfer to the cloud can take place, because rapid central intermediate storage can take place in the (local) data memory. A lower bandwidth for transmission to the cloud can thus be bridged, because not all existing data has to be transmitted immediately.

Since data is constantly being measured and transmitted by the field devices or sensors during operation, it is generally not possible to store unlimited amounts of data. Therefore, in a development of the invention, the (local) data memory can be designed as a rolling memory, i.e. the currently measured data is transmitted to the data memory and stored, but deleted again after a certain time when there is no more free memory space. For this purpose, the data in the data memory is provided with time information.

The field devices carry out an execution phase or an execution method. For this purpose, the field device can have a logic circuit that is programmed into a microcontroller, an FPGA or the like, for example. The number of inputs is fixed. The data or measured variables that are present at the inputs are also specified in certain areas. The output variables or processing values at the outputs of the respective field device are therefore also fixed and are also in predefined ranges. In summary, the field devices send data to the controller on the one hand and evaluate the AI and/or ML algorithm on the other and use them to control the actuator (e.g. a valve or the like).

1. 1. Gerät arbeitet in einem erwarteten Zustand (Normalbetrieb)
2. 2. Gerät erkennt einen Fehler- oder Signalzustand (Gestörter Betrieb, bzw. Warn- oder Steuerfunktion wird eingeschalten)
3. 3. Gerät arbeitet in einem unbekannten Zustand, wenn z.B. bei der Auswertung unbekannte Werte oder unerwartete Abfolgen von Werten auftreten.

The AI/ML execution algorithm present on the field device evaluates the sensor information with regard to the values in relation to one another and also in relation to time. A general distinction can be made between 3 states:

1. 1. Device is working in an expected state (normal operation)
2. 2. Device detects an error or signal status (faulty operation or warning or control function is switched on)
3. 3. The device is working in an unknown state, for example if unknown values or unexpected sequences of values occur during evaluation.

A control authority decides whether a special operating condition, eg a faulty condition, or an actually unknown condition is present. In the case of a special operating status (eg an error), a corresponding message (eg an error message) is output. In the case of an unknown state, the data is passed to the training procedure so that the Learning process in the cloud or the central computer adapts the parameter file or matrix and is sent to the field device.

example

• 1: eine schematische Darstellung eines computerimplementierten Verfahrens zur Ausführung einer Steuerung mit Anomalie-Erkennung gem. der Erfindung, hier mit rollierendem Datenspeicher,
• 2: eine schematische Darstellung des Zusammenwirkens von Ausführungs- und Trainingsverfahren gem. der Erfindung, sowie
• 3 eine schematische Darstellung eines computerimplementierten Verfahrens zur Ausführung einer Steuerung mit Anomalie-Erkennung gem. der Erfindung mit direkter Cloud-Anbindung.

An exemplary embodiment is shown in the drawings and is explained in more detail below, giving further details and advantages. Show in detail:

• 1 : a schematic representation of a computer-implemented method for executing a control with anomaly detection according to the invention, here with rolling data memory,
• 2 : a schematic representation of the interaction of execution and training methods according to the invention, and
• 3 a schematic representation of a computer-implemented method for executing a control with anomaly detection according to the invention with direct cloud connection.

In 1 a system 1 for executing a computer-implemented method 12 for executing a control with a training phase T and with an execution phase A (cf. also 2 this) and presented with an anomaly detection according to the invention. In an industrial process that can operate according to one of the current Industry 4.0 standards, a system is used that uses a control device 2 that is coupled to field devices 4 via a fieldbus 3 . The field devices 4 in turn have individual sensors 5, which is also 1 is indicated schematically. Depending on the process, the field devices 4 with sensors 5 can be various external components outside of the actual control device 2, which monitor the industrial process or take on other functions in it, such as actuators (actuators, valves, etc.) that report back , which position, which opening state, etc. is currently present, or other sensors for determining a temperature, a speed, a speed, an electrical variable, etc.

Training phase T:

Carrying out a learning process using methods of artificial intelligence or machine learning regularly requires a great deal of computing power, especially since in-depth knowledge of a system is required in order to be able to derive generally valid algorithms that the system can use. This knowledge is available in the form of individual examples or a large number of sensor data. The more data that can be evaluated, the more precise the prediction using a developed algorithm. On the other hand, only a limited time is available to carry out this learning process with AI in order to make an application in industry possible.

The required computing power is made available in a so-called cloud 7 (or a central computer). Sometimes the computing effort can justify the use of a GPU that is not always available in a company, and certainly not on every individual production machine. With the cloud 7, a central computer system is used, which may also process and evaluate the evaluation of several independent data sets from different systems. The Cloud 7 takes over the training process T with AI or ML. In addition, an even larger storage capacity can be made available by means of a cloud 7 than is possible in practice with a rolling data memory 6 in the area of the production plant. The data is trained within the training phase T using an AI or ML learning algorithm.

Execution phase A:

An evaluation circuit using a microcontroller MC is integrated in the field devices 4, with which the sensor data is evaluated and converted and transferred into an evaluation value or into another data form (e.g. a Boolean variable), a request to the controller or a command. These converted data are then transmitted to the control device 2 via the bus 3, for example. The evaluation in the microcontroller MC of the field device 4 takes place via a logic circuit which was created with the aid of artificial intelligence or with the aid of a machine learning method. The data generated in this way are transmitted to the rolling data memory 6 via the field bus 3 and are collected and stored there. Time information is assigned to the data in each case, which depends on when and in which order the data were generated or arrived in the data memory 6 . This further processing of the data originating from the sensors 5 in the field devices 4 is part of the execution phase A.

As already described in connection with execution method A, the data is collected and buffered in a central data memory 6 . The cloud 7 then retrieves only the data from the data store 6 that it needs for the training T that is currently being carried out. The data in the data memory 6 have been stored on a rolling basis, ie time information has been assigned to them in each case. Is the previously specified memory cher has been filled with data, the oldest data in the data buffer 6 is overwritten. This intermediate buffering allows the data to be passed on to the cloud 7 in portions, depending on the retrieval by the cloud 7, so that a low bandwidth on the transmission path to the cloud 7 is sufficient.

Coupling and interaction of execution phase A and training phase T:

The cloud 7 reports the parameter files/matrices M it has generated, which were generated by the learning process in the cloud 7, to the control device 2 and/or to the field devices 4 via the connection structure 10 (e.g. a network or another wireless or wired data line ), where these parameter files or matrices M are implemented by the processors of the control device 2 or in the microcontrollers of the field devices 4 as a logic circuit in the existing algorithm to be executed as part of the execution phase A.

2 shows once again schematically the interaction of training methods and execution methods A. The cloud 7 reports the logic newly learned in the training phase T to the control device 2 or the field devices 4 via the connection structure 10 and thus changes the course of the execution phase A.

• - Es handelt sich um einen speziellen Betriebszustand, z.B. um einen Fehler, etwa bedingt durch einen Mess- oder einen sonstigen Sensorfehler, bzw.
• - es handelt sich nicht um einen Fehler oder einen anderen bislang bekannten Betriebszustand, sondern um einen (Betriebs-)Zustand, der im Einlernprozess bislang noch nicht aufgetreten ist. Dieser neue Zustand soll dann vom Training T miterfasst werden, damit die verwendeten Ergebnisalgorithmen darauf zugreifen und diesen Zustand berücksichtigen können. Ein neuer künstlicher Lernprozess setzt sein.

In the execution phase A, the following must now be expected due to the predefined logic: since the system 1 is in a learning process, certain states may not yet be recorded by the learning and training process T, for example. The data, for example the measured sensor data, can therefore lie outside of a range that can be processed by the logic circuit. In the case of other design variants, the time behavior or the history of the data must also be taken into account during the evaluation. The system 1 recognizes a previously unknown state. The control device 2 then reports this state; because this state must be checked by a supervisory authority 11 . In general there are two options:

• - It is a special operating state, e.g. an error, for example caused by a measurement error or another sensor error, or
• - It is not an error or another previously known operating state, but an (operating) state that has not yet occurred in the teach-in process. This new state should then also be recorded by the training T so that the result algorithms used can access it and take this state into account. A new artificial learning process begins.

After a decision has been made by the supervisory authority 11, the control device 2 is informed whether there is an error which may also require intervention in or the process to be aborted, or whether the training T needs to be expanded. From the execution phase A, the data is collected in the data memory 6 and transmitted to the cloud 7 for the implementation or extension of the training method T, in that it is read out via the computer connection 9 if necessary. The result algorithms are transmitted and integrated via the connection structure 10 to the embedded devices, in particular the field devices 4 or the control device 2 . From then on, execution phase A changes accordingly due to the addition of other learned and trained states.

3 again shows a system 1 without rolling data memory. A size is measured in a machine by a sensor 5 . In the embedded field device 4, this is pre-processed here in the microcontroller MC (eg via a logic circuit), then the data are processed using the algorithm entered in the field device 4. An operating state BZ is determined and output. This method, which the field device 4 executes with the aid of the microcontroller MC or another logic circuit or a simple processor, is referred to as execution method A. This execution method A is carried out by executing an algorithm using specified parameters or a matrix M.

The pre-processed data (e.g. a Fast Fourier Transformation FFT may have been performed on it in the field device 41) are transmitted via a broadband network 10 to a cloud 7 in which the training phase T takes place. The data is processed with a support vector machine or an SVC learning algorithm in a mainframe computer in the cloud, so that a new parameter file / matrix M can be output, which in turn is transferred via line 9 to field device 4 via data transfer and stored there in the algorithm is included in execution method A.

If a previously unknown state is also recognized here, it can be transmitted to the cloud 7 via the connection 10 and included in the training process T, so that this state can be included in the parameter file/matrix M.

• - spezieller Betriebszustand oder
• - tatsächlich unbekannter Zustand,

wobei im Falle des tatsächlich unbekannten Zustands das Trainingsverfahren T mit den Daten des tatsächlich unbekannten Zustands gestartet und eine neue Parameterdatei / Matrix M erzeugt wird, während wiederum im Falle eines speziellen Betriebszustands eine Meldung ausgeben wird.All exemplary embodiments and developments of the invention are distinguished by the implementation of a training method T provided as part of the computer-implemented method 12, with a cloud 7 for the execution of Machine learning and/or AI training algorithms are used using the data, and a parameter file/matrix M is determined by executing the training algorithms, while further performing an execution method A, in which, in the case when the data outside are within the predetermined range, this state BZ is recognized as an unknown state, and the unknown state is transmitted to a control authority, which assigns one of the two following state variables to the unknown state:

• - special operating condition or
• - actually unknown condition,

where, in the case of the actually unknown state, the training method T is started with the data of the actually unknown state and a new parameter file/matrix M is generated, while again a message is output in the case of a special operating state.

Reference List

1

System for carrying out the procedure

2

control device

3

fieldbus

4

field device

5

Sensor in the field device

6

rolling data storage

7

cloud

8th

SVM or SVC learning algorithm

9

Computer connection from data storage to cloud

10

Connection structure between cloud and embedded device

11

supervisory authority

12

computer-implemented method with anomaly detection

A

execution procedure

BZ

operating condition

M

AI/ML Matrix

MC

microcontroller

T

training procedures

Claims (6)

Computer-implemented method (12) for executing a control with anomaly detection, comprising: • providing a control device (2), • providing at least one sensor (5), data being recorded by the sensor (5), • providing at least one field device (4 ), Which interacts with the at least one sensor (5) and/or comprises the sensor (5), the sensor (5) transmitting the data to the field device (4) for executing an execution method (A), in which the at least one Field device (4) evaluates the data according to a logic circuit, which assigns a processing value to the data if they are in a predetermined range and thus in a known state (BZ), • the field device (4) interacting with the control device (2), in particular receives and/or sends and/or exchanges data and/or commands, • providing a network (3) for communication between the control device (2) and the at least one a field device (4), • provision of a cloud (7) and/or a central computer for carrying out a training method (T), • transmission of the data to the cloud (7) and/or the central computer, at least via the network (3 ), • where: i. the cloud (7) and/or the central computer is used as a computer system for executing machine learning and/or AI training algorithms, in particular with higher computing power than required for executing the execution method (A), using the data and ii. a parameter file and/or matrix (M) is determined by executing the training algorithms, • carrying out the execution method (A), in which, if the data lie outside the predetermined range, this state is recognized as an unknown state, and i . the unknown status of a control authority (11) is transmitted, ii. which one of the following two status variables is assigned to the unknown status: o special operating status (BZ) or o unknown operating status, • in the case of the actually unknown status, the training method (T) is started and/or supplemented with the data of the actually unknown status and one Parameter file and / or matrix (M) is generated, and • wherein in the case of a special operating state (BZ) a corresponding message is issued, and wherein • the parameter file and / or the matrix (M) to the control device (2) and / or forwarded to the field device (4) or the at least one field device (4) whose data relate to the parameter file and/or the matrix (M), and • the parameter file and/or the matrix (M) being stored on the field device (4 ) and/or in the controller device (2) for carrying out the execution method (A) is applied. Method (12) according to claim 1 , characterized in that the cloud and / or the central computer is also used as a processing memory for the data. Method (12) according to one of the preceding claims, characterized by providing a data memory (6) connected to the network for the central collection of the data provided by the at least one field device via the network, and the data memory being used as an intermediate memory in order to To transmit data in portions to the cloud and/or the central computer and/or to compensate for a low bandwidth during data transmission to the cloud and/or to the central computer. Method (12) according to one of the preceding claims, characterized in that a rolling data memory is used as the data memory (6), in which the data read out from the at least one field device (4) is provided with time information until the data memory ( 6) and/or stored there until the corresponding data is transferred to the cloud (7) and/or the central computer and when the data memory is full (6) and/or after the data has been transferred to the cloud (7) and/or the central computer, these data are deleted depending on the time information, in particular starting with the oldest data. Method (12) according to one of the preceding claims, characterized in that the cloud (7) and/or the central computer receives the data from the network (3) and/or the data belonging to the training method (T) just carried out and/or to the actually unknown state. or requests from the data store (6) for their transfer to the cloud (7) and/or the central computer. Method (12) according to one of the preceding claims, characterized in that the field devices (4) use a microcontroller (MC) to evaluate the parameter file and/or the matrix (M) and/or to carry out the execution method (A).

Images