Methods and apparatus for monitoring the status of components of process facilities

Abstract

This invention relates to a method and apparatus for monitoring the state of components in a process facility, the process facility having multiple components, the state of which is determined based on process variables from at least one sensor associated with the component, and a corresponding dataset is recorded and stored for each sensor. A graph is generated from a digital flow diagram of the facility, the digital flow diagram including the structure of the facility and its components, as well as the functions of the components and their functional relationships as supplementary information. The graph includes components as nodes and functional relationships between components as lines of action according to the flow direction of the flow diagram. Nodes in the graph are selected as target nodes. Each target node corresponds to a component whose state should be monitored using a machine learning (ML) algorithm. The determination of component state is significantly improved by using an ML algorithm trained in this way.

Description

Technical Field

[0001] This invention relates to a method and a corresponding apparatus for monitoring the status of components in a process technology facility. Furthermore, this invention relates to a corresponding computer program and computer program product. Background Technology

[0002] The concept of "condition monitoring" means periodically or permanently detecting the technical condition of a machine or machine parts using sensors, and analyzing the resulting sensor data for further use. The most convincing measurements from sensors are mostly physical parameters, such as temperature, pressure, vibration, humidity, and liquid level.

[0003] Condition monitoring is particularly useful for process technology components in complex process technology facilities in the process industry. In this context, the term "process technology" component includes various machines or units within the facility, such as vessels, pumps, valves, heat exchangers, and their connections to each other, especially piping. All components are used to transport, separate, temperature, react, or otherwise alter the composition, type, or properties of certain media (e.g., whether they are liquids, gases, or solids). Sensing devices within the process technology facility are designed to detect process variables such as temperature, pressure, and level.

[0004] The simplest way to monitor status is to look at the measurements of each sensor associated with the device and analyze anomalies; that is, check if the current value is within preset boundary values, which means the machine is normal. If the current value is outside the boundary values, the device is abnormal and an alarm is sent.

[0005] To identify anomalies in sensor data, machine learning-based methods are frequently used. In machine learning (ML), a computer-implemented system learns patterns and associations from data. In this way, after completing a learning phase based on instances from historical data, it is able to generalize associations, identify outliers, and recognize unknown patterns. For this purpose, the algorithm builds a statistical model based on training or learning data. If a machine learning algorithm (ML algorithm) is trained to, for example, monitor or predict the behavior of process technology components in an industrial facility, a training dataset must be provided to the algorithm beforehand.

[0006] In large and complex process facilities, sensor data from different sources are often superimposed. In many cases, the training dataset for machine learning models includes sensor data that is irrelevant to the behavior of the part being monitored; furthermore, the values ​​recorded by sensors are often noisy. This makes it difficult to train algorithms or models, resulting in patterns and correlations in the sensor data often being difficult to identify and leading to erroneous descriptions of the parts being monitored.

[0007] Improvements in state monitoring through machine learning can be achieved by using extremely large datasets, particularly through training improvements. However, such large datasets are not always available, and training with such datasets is correspondingly time-consuming and complex. The quality of training (and evaluation) can be improved by manually selecting datasets of sensors not needed for monitoring components within a technical facility. This is disadvantageously associated with high workload. A common problem is that determining the relevant sensor datasets often requires domain-specific knowledge of the technical facility. Without a sufficient overview of the technological processes within the facility, specialized data analysts in this situation rely on collaboration with those possessing domain knowledge, which, in addition to the high workload, represents a significant additional cost for many and is often impractical for many projects. Summary of the Invention

[0008] The objective of this invention is to improve the condition monitoring of components in process technology facilities by means of ML algorithms, and in particular to reduce the high workload in selecting training datasets for the ML algorithms used.

[0009] This invention is based on the understanding that if the structure of the facility, pre-defined by the flowchart, is considered when learning the ML model, the ML algorithm for monitoring the state of components of the technical facility provides significantly better results. This is contingent solely on the existence of a digital flowchart.

[0010] According to the present invention, a diagram is generated from a digital flow diagram of a process facility. This digital flow diagram includes the structure of the facility along with its components and, as additional information, the functions and functional relationships of the components. In this diagram, components of the facility are designated as nodes, and the functional relationships between components are defined as lines of action according to the flow direction of the flow diagram, based on the additional information and relationships preset by the flow diagram. In the diagram, nodes are selected as target nodes. Each target node corresponds to a component whose state should be monitored using a machine learning (ML) algorithm. Subsequently, input nodes for the previously selected target nodes and sensor datasets related to the monitored components are derived based on rules and / or metrics. The datasets of related sensors for the input nodes are used as input data to train the ML algorithm for the target nodes. The state of the monitored components is determined using the output data and results of the ML algorithm.

[0011] The main advantage of this invention lies in its significant improvement in the training quality of the ML algorithm and its remarkable improvement in the use of the ML algorithm for monitoring the state of components of a technical facility by specifically pre-setting an input dataset for the device-specific structure. No domain-specific knowledge of the technical facility is required. The tedious manual selection of relevant sensor datasets is eliminated through pre-set rules and / or metrics. This method can operate largely automatically after selecting the component to be monitored (represented as a target node in a graph representation), and provides the output data of the ML algorithm quickly and reliably. Typically, the result of the ML algorithm will be at least one process variable characterizing the state of a component of the technical facility.

[0012] In a particularly advantageous implementation variant, the user can define, pre-set, or select rules and / or metrics. This allows for high flexibility in the application. For example, one could imagine the user selecting specific metrics for the relationship between the target node and the input node via a menu in the graphical user interface. Alternatively, the user could input process variables detected by specific sensors for the input node via an input menu. This implementation variant allows the user to optimize the method because different sets of metrics or different rules can be input to evaluate the computational results of the ML algorithm.

[0013] In another advantageous embodiment of the method according to the invention, the dataset for each sensor corresponds either to a measured value or a simulated value of the process variable. The dataset of real-time recordings of the measured values ​​of the process variables allows for monitoring of the condition of components during the continuous operation of the facility. If simulated sensor datasets are used, the method according to the invention can be used, for example, before commissioning of the facility. In particular, a digital twin of the facility can provide datasets for monitoring the equipment, enabling the learning of ML algorithms outside the operation of the facility using these simulated datasets.

[0014] In another advantageous embodiment of the method according to the invention, the ML algorithm is trained beforehand or during the continuous operation of the technical facility. Extracting the method steps of the training allows for improved state monitoring because the actual state determination is first achieved using the trained algorithm, which leads to better computational results.

[0015] A particularly advantageous aspect here is that the training timeframe is configurable. This allows users to independently determine when the ML algorithm delivers optimal results.

[0016] The described improvements relate not only to the method according to the invention but also to the apparatus.

[0017] Furthermore, the implementation of the present invention or the improved solution described herein can be implemented by a computer program, particularly a software application, and when the computer program is implemented on a computer, the method is implemented using program code instructions that can be implemented by a computer.

[0018] The present invention and / or each of the described improvements can also be implemented by a computer program product having a storage medium on which a computer program is stored, the computer program implementing the present invention and / or improvements.

[0019] The computer program product can be advantageously transferred to the working memory of the computing unit, and can be implemented therefrom by means of at least one CPU. The computer program product can advantageously be stored on data storage, such as a USB flash drive, hard disk, or CD-ROM / DVD-ROM, and can be retrieved or installed on the computing unit from there. Attached Figure Description

[0020] The invention will now be described and explained in more detail with reference to the embodiments shown in the accompanying drawings.

[0021] The following are simplified diagrams:

[0022] Figure 1 An exemplary cutout showing a flow diagram of a process technology facility.

[0023] Figure 2 The diagram is consistent with the first exemplary embodiment of the present invention.

[0024] Figure 3 This is a diagram consistent with the second exemplary embodiment of the present invention.

[0025] Figure 4 Two examples are shown of time-varying curves of the estimated values ​​of process variables of monitored components of a facility, obtained using ML algorithms.

[0026] Figure 5 An embodiment of an apparatus for performing a method according to the invention for monitoring the status of components of a process facility is shown. Detailed Implementation

[0027] The premise of this invention is the existence of a digital flow diagram. A flow diagram is a technical diagrammatic aid in process technology. A flow diagram of a process facility includes the structure of the facility and its components, and includes the functions and functional relationships of the components as additional information. One implementation of this flow diagram is a piping and instrumentation diagram (also known as a R&I flow diagram, P&ID: piping and instrumentation diagram). It shows the connections, flow directions, and all pipe fittings. Digital flow diagrams can be generated using facility planning tools (such as Siemens' COMOS).

[0028] Whenever technical facilities and flowcharts are mentioned in this application, they can also be understood as individual or interconnected (or to be connected) parts of facilities or flowcharts.

[0029] Figure 1 This is an exemplary cutaway of a process flow diagram of a facility. The cutaway FSA includes graphical process objects representing different components of the facility, such as pumps, containers, butterfly valves, or valves, and is interconnected according to the facility's piping. The process objects typically also contain technical information about parameters of the component they represent, such as pipe diameter or container filling volume (not shown here). The FSA cutaway shown here illustrates container B, on which the liquid level is measured by means of sensor S1. The inlet and outlet of container B are visible through the lines of action with arrows. The container outlet branches into two branches. Pumps P1 and P2, along with associated valves V1 and V2, and sensors S3 and S2 for measuring the flow rate at the pump outlets are present in the corresponding piping sections. Butterfly valves A1 and A2 are arranged downstream of pumps P1 and P2 in the flow direction. Additionally, pressure sensor S4 and another valve V3 are arranged in the flow direction of a common drain pipe. In addition to the sensing devices represented here by sensors S1, S2, S3, and S4, the actuators of the facility (represented by motor M for driving the parts) with corresponding technical information (not shown here) can also be derived from the process diagram. Process variables, such as pressure, temperature, level, or flow rate, are measured using sensors. Each sensor records a corresponding dataset, which can, for example, consist of time-series data of the measured process variables, and the dataset is stored in an intermediate memory or database.

[0030] As can be clearly seen from the truncated FSA, the state of components (e.g., container B, pumps P1 and P2, or butterfly valves A1 and A2) can be determined using multiple sensors. The component of pump P1 is in direct action relationship with sensors S1, S3, and S4. The line of action or action relationship can be restricted in one direction or bidirectional, i.e., acting in both directions. These action relationships can now be simplified and recorded in a graphical representation.

[0031] exist Figure 2 and Figure 3 The diagram illustrates an implementation method represented by a graph. The graph in the context of this invention includes each of the following schematic diagrams: in which components containing technical facilities are designated as nodes K based on additional information and associations preset by the flow diagram, and functional associations between components are included as action lines W having a directional description according to the flow direction of the flow diagram, or these action lines are at least partially bidirectional. Therefore, the action line according to the flow direction of the flow diagram can have one direction (for medium flowing through a pipe in one direction) or two directions (in the case of a pressure-balanced pipe). Here, this could be, for example, a knowledge graph. However, the invention is not limited to knowledge graphs. Different types of graphs or ontologies can be used, as long as the object type and connection between any two objects can be retrieved.

[0032] In this embodiment, the facility planning tool supports outputting digital R&I diagrams in the so-called DEXPI format. The abbreviation DEXPI stands for "Data Exchange in the Process Industry." The DEXPI format corresponds to a neutral data format, which enables the exchange of neutral process information between software products from different manufacturers of engineering tools. The digital flow diagram in DEXPI format can be used as a diagram or converted to other formats. It includes all devices and sensors and their connections, with orientation in a simplified diagrammatic representation. The diagram is based on a common data model. When transferring P&ID to the diagram, data is transferred, which should include graphics, symbols, topology, all technical attributes, enumerations, selection lists, etc., to enable seamless continuation of P&ID functionality in the target system.

[0033] If a digital R&I diagram exists, such as in the form of a knowledge graph or a similar ontology, it becomes possible to determine which sensors are relevant to the monitored components of the equipment or process facility in the relevant application context, and thus are particularly relevant for training ML algorithms.

[0034] Therefore, if it is necessary to monitor the status of components of a specific device or process facility using an ML algorithm, then, for example, a user of the method according to the present invention (which can be implemented as a computer program product in an embodiment) selects a node in a graph representation as a target node in one step. The target node in the graph representation corresponds to a component whose status will be monitored. The selection of the target node can be performed automatically and / or preset by software. For the target node, the ML algorithm should be trained.

[0035] In the next step, a sensor dataset should be determined, which is functionally associated with the components of the target node, and thus the state of the components to be monitored at the target node should be determined. The components and sensors associated with the monitoring of the components to be monitored at the target node are referred to as input nodes in the graphical representation. Here, an input node can be either a node connected upstream of the target node in the flow direction or a node connected downstream of the target node in the flow direction. Which input nodes (components) are relevant to the state determination of the components at the target node is determined from the corresponding application scenario. According to the invention, it is also possible to predefine specific input nodes based on metrics or rules.

[0036] In the context of this invention, metric is understood as a distance function or any distance measurement, as it is mathematically defined between two points in space. This can be, for example, the length of the shortest path, the drag distance, or the Euclidean distance. In principle, the metric should determine how "close" the (active) connection between the sensor and the device to be monitored must be for the sensor to be applicable. In the simplest case, the nearest neighboring node with its sensor dataset can be selected as the input node. However, more complex situations often arise due to the application. If, for example, the temperature of a component at a target node needs to be determined in 5 minutes, and the liquid transported in the input pipe will take approximately 1 ± 0.3 minutes to reach the next component, it makes sense to select only the upstream nodes located between 4 and 6 nodes from the target node as input nodes. In this case, the input nodes are thus selected according to a rule. Another example of such a rule is to select all pressure sensors located at a distance of at most one node downstream as input nodes, and all temperature sensors located at nodes located 2 to 3 nodes upstream from the target node.

[0037] In a particularly advantageous embodiment of the invention, the user (data analyst or domain expert) can pre-define which nodes represented by the graphs should be considered relevant input nodes for the corresponding application scenario, defining the rules and metrics. The rules and metrics for determining which nodes are relevant to a specific problem can be established once and subsequently applied to all sufficiently similar equipment / target nodes (e.g., all similar pumps in a facility). Therefore, during continuous operation, all sufficiently similar equipment / target nodes, and thus the entire facility, can be monitored without significant additional cost.

[0038] Since each node in the diagram represents the connection of the process facility to different sensors (see also...) Figure 1The sensor dataset related to the state determination of the target node's components is obtained from the determination of the input node. This sensor dataset is used as input data to train the ML algorithm for the target node, wherein the state of the target node is determined by means of the output data of the ML algorithm.

[0039] exist Figure 2 A first embodiment illustrating the method according to the present invention is given herein. Figure 2 In the graphical representation of a process technology facility-based R&I flow diagram, specific components required for the operation of the process technology facility are represented as nodes K, and the pipes extending between components are represented as lines of action. In a highly simplified truncated portion of the graphically illustrated R&I flow diagram, the flow direction extends along the arrow direction, typically with diagonal connections to the nearest neighbor from top to bottom, where some connections are missing. In this case, the components of a node represent temperature measurement points within the piping system. Nodes correspondingly correspond to pipe sections where the temperature of the delivered liquid is measured by means of sensors (temperature sensors). Arrows between nodes indicate the flow direction of the liquid between pipe sections with sensors. Therefore, components and sensors are considered the same in this embodiment. Typically, each component of a node represented graphically is connected to more than one sensor.

[0040] In this embodiment, an ML algorithm should now be trained for the target node K43. This ML algorithm should monitor the state of the pipe at the measurement point by estimating the temperature in the pipe segment using the ML algorithm. The measurement point in the pipe diagram is represented by the target node K43. For determining the state of the pipe segment at the target node, due to the conditions of this application, only the datasets of upstream temperature sensors are relevant, as only these datasets contain information about the liquid flowing into the pipe segment represented by the target node. Furthermore, temperature sensors located far upstream can be ignored because their information is reflected in the datasets of subsequent sensors and is strongly mixed with information from irrelevant sensors before reaching the target sensor. Therefore, in this example, the maximum distance between the target node and the two nodes derived as relevant input nodes is selected. Thus, the metric here is the length of the shortest path. Figure 2In the dark gray shaded area, the input nodes K32, K22, K33, K23, and K24 of the target node K43 are visible. Therefore, only the dataset of the sensors input to nodes K32, K22, K33, K23, and K24 is used to train the ML algorithm for the target node. The dataset of the sensors further away from the targets (nodes K00 to K16) in the light gray shaded area is not needed when training the ML algorithm and is negligible for the state of the pipe segment represented by the target node. The dataset associated with the downstream nodes located behind the target node is irrelevant to the state determination of the target node.

[0041] exist Figure 3 The diagram illustrates a second embodiment of the method according to the invention. In this diagrammatic representation, nodes represent sections within a pipe system where, in addition to using one sensor (temperature sensor) to measure the temperature of the transported liquid, another sensor is used to measure the pressure within the pipe section. The arrows between nodes also indicate the direction of liquid flow between pipe sections with sensors. Thus, in this embodiment, each component (i.e., pipe section) is connected to two sensors (here, a temperature sensor and a pressure sensor).

[0042] In this embodiment, node K33 is selected as the target node, and an ML algorithm is trained to monitor the pipe condition in this pipe segment. Here, input nodes are selected according to rules. The ML algorithm should only consider the pressure sensor datasets of adjacent nodes (i.e., nodes at most one node away from the target node) and the temperature sensor datasets of upstream nodes at a distance of 2 to 3 nodes. Therefore, nodes K22, K23, K24, K43, and K44 are determined as input nodes, where only the pressure sensor datasets of these nodes are relevant to determining the condition of the pipe segment at target node K33. Figure 3 The nodes in the diagram are stored in the light gray region P. Additionally, nodes in the dark gray region T are also identified as input nodes, where only the temperature sensor dataset is considered.

[0043] Selecting only specific, relevant datasets as input datasets for ML algorithms according to the method of the present invention significantly improves the output parameters computed by the algorithm. This should be based on... Figure 4 The two diagrams in the figure are used to illustrate this. These two diagrams show the use of Example 1 (see [reference]). Figure 2 The time-varying curve of the output parameter of the ML algorithm for the process value of the component of the target node K43 in the target node. Figure 2Node K43 in the diagram represents a temperature sensor in a pipe section of the piping system. The ML algorithm allows for the estimation of temperature measurements based on the input dataset, such as the dataset of the input nodes. In two graphs, D1 and D2, the output parameter A of the ML algorithm used is plotted with respect to time. Additionally, the target or nominal value of the output parameter S is plotted relative to time. The time period TR, which was used to train the algorithm, is also marked in both graphs.

[0044] As can be seen in Figure D1, during the training phase (TR), the two curves of the output parameter—the nominal value and the temperature value calculated using the ML algorithm—are approximately consistent. Conversely, a significant deviation appears in D1 during the evaluation phase. This is because the training dataset for the ML algorithm in Figure D1 also includes irrelevant input datasets, while in Figure D2, only the relevant training dataset according to the present invention is used for calculating the estimated temperature value. In Figure D1, the temperature value has already been calculated using... Figure 2 The graph shown represents the dataset used to train the ML algorithm for all nodes (60 temperature sensors). In graph D2, only the dataset used... Figure 2 The diagram shown illustrates the training of the ML algorithm on the relevant datasets for nodes K22, K23, K24, K32, and K33. This approach enables a significant improvement in the monitoring of the observed pipe sections' condition.

[0045] The present invention and the described improvements are preferably implemented in software and in hardware, for example, using specialized circuitry, or in a combination of software and hardware, for example, in a device for data processing.

[0046] exist Figure 5 The diagram illustrates an embodiment of a device CM for monitoring the status of components of a process facility. In this embodiment, the device CM has at least one interface S1 for receiving a digital flow diagram FS of the facility. This flow or piping diagram (R&I diagram), which may consist of multiple sub-images, is typically created using a facility planning tool (APW) in a known manner, for example, by linking graphical process objects represented by components of the facility through a facility planner or technology. The device CM also has at least one additional interface S2 for receiving a dataset D1 to DN (N = a natural number) or parameters derived therefrom from a sensor S1 to SN connected to at least one component of the process facility. The dataset containing process variables detected by the sensors can be cached in a database DB. Alternatively, the device CM can also have a single interface designed to receive arbitrary data and / or files.

[0047] Furthermore, the apparatus CM has at least one data processing device DV, which implements the method according to the invention based on the transmitted digital flowchart and the transmitted sensor dataset. The apparatus CM may also include, for example, a library or other arbitrary storage unit SP or database storing flowcharts or diagrams. The data processing device DV has at least one processor P or CPU and working memory M. The invention, implemented as a computer program, can be maintained in or loaded into the working memory M, and implemented from there by means of at least one processor P.

[0048] The device may also have a display unit A or be connected to a display unit designed to monitor components of the technical facility on a user interface (GUI). Through the graphical user interface, the user can freely interact with the data processing device DV of the device CM.

Claims

1. A method for monitoring the status of components of a process technology facility, The process facility has multiple components that are interconnected in terms of process technology. in, The state of the component is determined based on process variables from at least one component-related sensor, and a corresponding dataset is recorded and stored for each sensor. Its features are, A diagram is generated from the digital flow diagram of the technical facility. The digital flow diagram includes the structure of the technical facility, its components, and the functions of the components as additional information, including functional relationships between the components. Based on the additional information and relationships preset by the flow diagram, the components of the technical facility are included in the diagram as nodes, and the functional relationships between the components are included as lines of action according to the flow direction of the flow diagram. The node corresponding to the component is selected as the target node, and the status of the component is monitored using the machine learning algorithm. The input nodes and sensor datasets of the input nodes, which are associated with the component to be monitored, are determined based on rules and / or metrics for the previously selected target node. The relevant sensor dataset is selected as input data for training the ML algorithm for the target node, and The state of the component to be monitored is determined by using the output data of the ML algorithm.

2. The method according to claim 1, Its features are, The rules and / or metrics are preset or selected by the user.

3. The method according to any one of the preceding claims, Its features are, The dataset of measurements from each sensor corresponds to the process variable or the simulated value of the process variable.

4. The method according to any one of the preceding claims, Its features are, The training of the ML algorithm is performed either in advance or during the continuous operation of the technical facility.

5. The method according to any one of the preceding claims, Its features are, The training time range can be set.

6. A device (CM) for monitoring the status of components of a process facility, the process facility having multiple components interconnected in terms of process technology, wherein, Each component is connected to multiple sensors (S1, ...SN) for detecting a dataset (D1, ...DN) of process variables, and the device has: An interface (S1) for receiving a digital flow diagram (FS) of the technical facility. Another interface (S2) is used to receive the dataset (D1, ...DN) of the sensors (S1, ...SN) of the components of the technical facility, or parameters derived from the dataset; as well as A data processing device (DV) designed to perform the method according to any one of claims 1 to 5.

7. The apparatus (CM) according to claim 6. It also includes a display unit (A) designed for selecting at least one target node via a user interface (GUI), the target node being associated with a component to be monitored, and any interaction between the user and the data processing device (DV) can be performed.

8. A computer program having program code instructions executable by a computer, the program code instructions being configured to perform the method according to any one of claims 1 to 5 when the computer program is executed on a computer.

9. The computer program according to claim 8, wherein, The computer program is a software application.

10. A computer program product having a computer program according to claim 8 that can be implemented by a computer.

11. The computer program product according to claim 10, wherein, The computer program product is a data carrier or storage medium.

Images