Harmonizing parameter data for use in digital twins

Abstract

In a computer implementation method (400) for harmonizing parameter data, the computer receives a source parameter dataset (221-xx) representing technical parameters from a component (110-xx) of an industrial machine (410). The source parameter dataset (221-xx) has a source parameter identifier (231-xx) and a source parameter value (241-xx). The computer uses a first pre-trained subnetwork (330) to match the source parameter identifier (231-xx) with a target parameter identifier (232-xx) (420). The computer uses a second pre-trained subnetwork (340) to match the source parameter value (241-xx) with a target parameter value (242-xx) (430). The second subnetwork (340) is selected according to the target parameter identifier (232-xx). The computer transfers both the target parameter identifier (232-xx) and the target parameter value (242-xx) to an interface that shows a user interface element corresponding to the industrial machine component (110-xx) and visualizes the target parameter value (242-xx) (440).

Description

Technical Field

[0001] Generally, the present disclosure relates to industrial processes, and more particularly, the present disclosure relates to a computer system, method, and computer program product for visualizing harmonized parameter data within a user interface.

Background Art

[0002] From a high-level perspective, and speaking very generally, an industrial system is a technical system that performs an industrial process, such as processing materials for the purpose of obtaining a product. A parameter is a characteristic that is related to an industrial process (and / or an industrial system) and can affect the performance of the process (and / or the operation of the system). To give some examples, a parameter can be a physical phenomenon (such as temperature or pressure), an indicator of the quality and quantity of the material being processed, or any other characteristic.

[0003] An industrial system can be considered to have a plurality of components, and knowledge of the parameter values specific to each component is a condition for operating the system.

[0004] In a fictional example of an industrial system, a number of industrial machines are arranged at a manufacturing site. The system components should include a compressor for supplying air, a pipe network for distributing compressed air, and machines that use compressed air. To measure the air pressure, the components are associated with pressure gauges. The system operator considers the pressure values to control the system. Speaking very simply, the operator makes the air reach all the machines at an appropriate pressure.

[0005] Traditionally, meters are implemented as pressure gauges (or "manometers") with a mechanical pointer that rotates on a dial. Human operators occasionally visually inspect the meters. Since the machines (and their meters) do not necessarily have the same manufacturing origin, the pointers and dials may look different. For example, an operator might see the pressure of a first machine by a red pointer moving on an odd-numbered ("1, 3, 5, 7, 9...") white dial, or they might see the pressure of a second machine by a black pointer on an even-numbered ("2, 4, 6...") yellow dial.

[0006] The operator understands that a specific pressure value (i.e., meter reading) requires immediate action, but also understands that the critical value may differ between machines. For example, the operator might open a valve in machine 1 when the pressure reaches "7," but activate machine 2 when it reaches "10."

[0007] However, due to changes in the industry, more and more meters are being replaced by digital sensors. As used herein, digital sensors are: • Sensing the physical phenomenon which is a parameter, • Provides parameter values ​​in the form of digital data. It is a computing device.

[0008] In other words, data from a digital sensor represents one or more technical parameters. A digital pressure sensor is still a mechanical component mounted on a pipe, but it is also an electronic device equipped with analog-to-digital converters and the like to provide data. Often, digital sensors have a small computing function (a software-controlled microprocessor) to preprocess the data, adding metadata such as timestamps and machine / sensor identification.

[0009] In this method, the operator no longer needs to walk directly to the manufacturing floor to check the meters. Instead, the operator may sit in a control room. Part of the control room may be located remotely. When the operator no longer interacts with the components manually, the actuator receives control signals from the control room.

[0010] In the industrial sector, there is a tendency to virtually represent individual industrial systems and their components using so-called "digital twins." The twin metaphor represents the relationship between a real system and its virtual representation on a computer. Each industrial system has its "system twin," each component has its "component twin," and so on.

[0011] Representing industrial machinery with digital data (such as "twin") makes it easier to extend the control loop into the control room (for example, automatically opening a pressure valve via an actuator at "7"), and also allows the computer to predict parts of the operation. For example, the computer can decide to open (or close, or keep closed) a valve by taking other data into consideration. The computer can even simulate the operation of the machine, and depending on the results of the simulation, the computer may start (or stop) the operation.

[0012] There is a growing trend towards making industrial machinery operate autonomously (i.e., by limiting the interaction between the machine and the operator), and the provision of comprehensive data on the machinery is one way to enable such autonomy.

[0013] Pressure gauges all share the common characteristic (or property) of measuring pressure. However, when the meter's pointer and dial are replaced by the predefined data structure of a digital sensor, various sensor manufacturers may use different data protocols. In other words, communication between the system and the twin is far from standardized.

[0014] Various data conventions can lead to improper operation, and a computer processing incorrect data may generate control signals at the wrong time (or with the wrong intention). In this embodiment, the computer may receive pressure data from the first machine, but may process that pressure data according to the configuration of the second machine. The computer may open the valve in the first machine too late when the pressure reaches "10", but not too early when the pressure reaches "7". In other words, the "twin" should match the original, otherwise a malfunction may occur.

[0015] While the example of measuring pressure is convenient for explanation, the data is not limited to pressure data. All machines provide data for multiple parameters.

[0016] Using a sensor fleet is generally feasible in that all digital sensors provide data according to a standardized, identical data protocol.

[0017] Patent Document 1 discloses a computer platform for processing industrial data arising from various sources. The processing focuses on analysis and event prediction. [Overview of the Initiative]

[0018] According to embodiments of the present invention, data harmonization is performed by processing source data into target data using a pre-trained neural network. The neural network learns the relationship between the source data and the target data by processing historical data available in time-series form.

[0019] Such machine learning techniques may benefit from historical data summarizing previous harmonization activities. Since harmonizing data conventions can be error-prone when done manually, the process of training a neural network ignores accidental inconsistencies in historical data records. In other words, a human expert may have introduced some incorrect harmonization somewhere in the time series, but should have harmonized the data correctly for a large portion of the time series.

[0020] A computer implementation method is disclosed as a method for harmonizing parameter data.

[0021] In the receiving step, the computer receives a source parameter dataset representing technical parameters from the components of the industrial machine. The technical parameters belong to the components. The source parameter dataset has a source parameter identifier and a source parameter value.

[0022] In the first matching step, a computer uses a first subnetwork of a pre-trained neural network to match source parameter identifiers with target parameter identifiers. (The first pre-trained subnetwork is trained on multiple pairs of historical parameter identifiers, each having a historical source parameter identifier at its input and a historical target parameter identifier at its output.)

[0023] In the second matching step, the computer uses a second subnetwork of a pre-trained neural network. It matches source parameter values ​​with target parameter values. The second subnetwork of the pre-trained neural network is selected according to a target parameter identifier and is trained with multiple historical parameter value pairs, each having a historical source value at its input and a historical target value at its output.

[0024] In the transfer step, the computer transfers both the target parameter identifier and the target parameter value to a user interface that shows user interface elements corresponding to components of the industrial machine and visualizing the target parameter values.

[0025] Optionally, the source parameter dataset (representing the technical parameters belonging to the component) is either a source parameter dataset from a digital sensor or a source parameter dataset from a virtual sensor.

[0026] Optionally, in addition to being trained with a set of parameter identifier pairs, the first subnetwork is simultaneously trained with historical context data in its input.

[0027] Optionally, in addition to being trained with a set of parameter value pairs, the second subnetwork is simultaneously trained with historical context data in its input.

[0028] Optionally, in the receiving step, the computer further identifies the real-time context that the component is currently operating. The real-time context is a set of data points related to the data points within the source parameter dataset of the technical parameters.

[0029] Optionally, in the receiving step, the computer identifies the real-time context, thereby accessing the status data.

[0030] Optionally, in the transfer step, the computer uses the user interface to visualize the target parameter values by symbols.

[0031] Optionally, after receiving the source parameter dataset, the computer uses first and second subnetworks to determine whether the source parameter dataset conforms to the specifications of the target parameter dataset. If the determination is positive, the method continues the transfer, thereby skipping the matching step.

[0032] Optionally, the computer performing the method may, after receiving the source parameter dataset, use a context network to identify the context type and may select a specific first subnetwork according to the context type to match the source parameter identifier with the target parameter identifier.

[0033] Optionally, after the transfer, control signals are acquired to control the operation of the components of the industrial machine.

[0034] This disclosure also relates to the use of a method for harmonizing parameter data by a computer system that specifies control data and transmits control signals to one or more components of one or more industrial machines.

[0035] The computer system is adapted to implement a method for harmonizing parameter data from components of industrial machinery (the details of this step and any optional steps are summarized herein).

[0036] From a training perspective, a computer implementation method for training a neural network is disclosed. The neural network has a first subnetwork and a second subnetwork, and the neural network is trained for subsequent use in a computer implementation method for harmonizing parameter data from multiple source protocols to a single target protocol. The data flows from input to output through the first and second subnetworks.

[0037] Parameter data relates to the components of industrial machinery, and a parameter dataset represents the technical parameters belonging to the components of the machine. A parameter dataset has a parameter identifier and a parameter value.

[0038] A source specification includes a source parameter identifier and a source parameter value, while a single target specification includes a target parameter identifier and a target parameter value.

[0039] This method includes training a first subnetwork with a plurality of pairs of historical parameter identifiers, each having a historical source parameter identifier at the input of the first subnetwork and a historical target parameter identifier at the output of the first subnetwork.

[0040] The method further includes training a second subnetwork with a plurality of historical parameter value pairs, each having a historical source value at the input of the second subnetwork and a historical target value at the output of the second subnetwork.

[0041] A computer program product that, when loaded into the memory of a computer and executed by at least one processor of the computer, causes the computer to perform steps of a method for harmonizing or steps of a method for training. [Brief explanation of the drawing]

[0042] Embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0043] [Figure 1] This diagram shows industrial machinery, a user interface on a remote computer, and a machine operator.

[0044] [Figure 2] This figure shows machine data written as a multivariate time series with data points.

[0045] [Figure 3] This figure shows multiple system components having digital sensors, a conversion function for harmonizing parameter data, and multiple digital component equivalents for processing target data.

[0046] [Figure 4] Figure 3 is at least partially repeated and shows the transformation function in an implementation using a neural network.

[0047] [Figure 5] This diagram shows the functionality of a neural network that provides syntactic transformation.

[0048] [Figure 6] This diagram shows the functionality of a neural network for providing semantic translation.

[0049] [Figure 7A] This diagram shows a flowchart of a computer implementation method for harmonizing parameter data.

[0050] [Figure 7B] This diagram shows a flowchart of a computer implementation method for harmonizing parameter data by applying a context-dependent network selection cascade.

[0051] [Figure 8] This figure shows mechanical components with various sensor implementation configurations.

[0052] [Figure 9] This figure shows first, second, and third industrial machines, each having multiple components, that communicate with a digital equivalent or twin via a neural network.

[0053] [Figure 10]This is a diagram of a typical computer. [Modes for carrying out the invention]

[0054] overview Figure 1 shows industrial machine 100, and further industrial machines having reference numbers 100-f and 100-F are also shown herein. All machines together can be considered a fleet of machines. Machine 100 comprises components 110-xx and further components 110-yy. Further machines also comprise further components.

[0055] Most of the components of a machine are, • Measurement data (i.e., data acquired from components via sensors, etc.) • Quasi-measured data (i.e., data in the function of measurement data from a "virtual sensor," see Figure 9) • Control data (i.e., data provided to components, such as data for modifying parameters), • Environmental data (i.e., data describing the external environment of components that potentially affect the machine), • Production data (i.e., data describing materials, products, tools, etc.) These are associated with parameter data.

[0056] From a very high level perspective, data is processed by a computer, which is represented by a user interface 200. Machine 100 (and 100-f and 100-M as well) is connected to its computer in a communicative manner. Operator 290 is the user of this computer.

[0057] Computers process data selected for specific purposes, such as remotely controlling a machine, rather than processing all available data, which in some aspects is convenient for describing data. Some data processing involves the use of data-trained computer modules, such as neural networks. Data in space and time, i.e., spatial and temporal aspects.

[0058] Figure 2 shows machine data written as a multivariate time series with data points. This figure shows the first data matrix {{X}}, the second data matrix {{Y}}, and the passage of time (t1, t2, ..., tm, ..., tM). The index m represents the current time tm, with smaller index m representing past times and larger index m representing future times.

[0059] Data points X and Y have column indices (index m=1 to M) for each time point and row indices (index n=1 to N) for multiple variables. For simplicity, M and N in the time interval Δt between consecutive time points tm and t(m+1) must also be equal for {{X}} and {{Y}}. In implementations, data points can be identified in other ways, and assuming equal Δt is merely a simplification of the explanation.

[0060] Returning to Figure 1, in spatial terms, the data can be further distinguished according to the physical location associated with the origin of the data point. In the first embodiment, various components are always located in various physical positions within the machine 100, so the multivariate time series {{X}} associated with component 110-xx is different from the further multivariate time series {{Y}} (associated with component 110-yy). Figure 1 shows the spatial configuration of the first embodiment with horizontal lines. The granularity by which component data {{X}} is distinguished from component data {{Y}} is based on the structure of the machine having the components. In the second embodiment, when a sensor (or other data provider) is associated with various physical locations within component 110-xx, a single time series {X} for variable n within {{X}} of component 110-xx will have spatially different data for other variables in {{X}}. For simplicity of explanation, Figure 1 does not show the second embodiment, but the principle is the same, where a component has subcomponents, and again for simplicity, the subcomponents can be associated with individual variables.

[0061] The parameters can be distributed along the sub-components. For example, the temperature distribution may vary across the surface, so that the temperature value is obtained as a function of surface position (x, y) and time.

[0062] The temporal aspect refers to the relationship between the past, present, and future.

[0063] Time series (univariate or multivariate) can be divided into a real-time portion and a historical portion. The real-time portion includes data representing conditions that may affect the operation of the components (and the machine) at the present (index m) or future (index > m). For univariate time series, the portion may include data points with indices (mK), ..., (mk), (m-1), and m. K is the "viewback" counter. For multivariate time series, the counter K may differ for each variable. The history section represents situations that no longer affect the operation. In simple terms, it includes data points that are out of reach of counter K.

[0064] For simplicity, and without the need for a clear distinction in this specification, real-time data can be used for machine control, and historical data can be used for training networks, etc.

[0065] A database symbol with the index "hh" represents the availability of historical data for {{X}} and {{Y}} (ideally, all components, all machines). Index hh can, but does not necessarily, correspond to index xx. Context concept

[0066] Figure 2 also illustrates the concept of context by dashed rectangles 501, 502, or the “context window.” The context concept takes into account both spatial and temporal aspects. A data point Xmn has one or more contexts. As used herein, a context is the set of data points (indicated by dashed rectangles) associated with a data point Xmn.

[0067] The context (i.e., the set) may change, and in principle, each data point has its own context. In this embodiment, a data point Xmn (i.e., data about variable n at time tm) may relate to data in the same multivariate time series {{X}} for constituent elements 110-xx. Context 501 can be used, for example, as a set of data points X21, X31, and Xm1 for variable 1 and data points X22, X32, and Xm2 for variable 2. Context 501 can be thought of as a two-dimensional context having dimensions (or aspects) in time and variables.

[0068] The data point Xmn may also have a context 502, namely data points Y12, Y22, Y32, Ym2 (variable 2, further component 110-yy). Context 502 is a one-dimensional context (in time).

[0069] The context is, • Artificial rules that take into account the relationships between parameters, or • Machine learning rules based on observations (e.g., by a computer processing historical data) It can be defined by rules such as those listed above.

[0070] The context does not need to be known in advance, but it can be learned.

[0071] Contexts can be distinguished into real-time contexts and historical contexts, that is, As Figure 2 shows the time series in real time (time index m represents the current time, as explained), the parameters represented by the data within the dashed rectangles 501 and 502 directly affect Xmn at tm. Therefore, contexts 501 and 502 are real-time contexts. In the case of historical data (see index hh in the set in Figure 1), if a multivariate time series {{X}} or {Y}} is part of it, the data within dashed rectangles 501 and 502 influenced Xmn. This is a past real-time context and can be considered a historical context. The historical context can be applied during the training phase (see Figures 5-6).

[0072] Optionally, a context can be established by having a hierarchical structure of temporal and spatial aspects.

[0073] The temporal aspect may take precedence over the spatial aspect. For example, a temporal context window is provided by going back in time by a predetermined number of Δt (e.g., a backview counter K). Within that window, some variables contribute to the context, while others do not.

[0074] Spatial aspects may take precedence over temporal aspects. For example, some variables may be related, while others may not (the computer can learn this), and the computer can further learn how much of a time window Δt it can open for related variables.

[0075] Contexts can be further distinguished according to the level of granularity, and the description gives two extremes: namely, the “global context” can be defined by data points that apply to substantially all machines at a particular time of year in a particular industrial site. For example, the first global context can be defined by June (i.e., the time of June) for a location in Asia (implicitly assuming relatively high air humidity), and the second global context can be defined by December for a location in Europe (dry but cold). The global context can be further enhanced by subcontexts, including material consumption, failure rates, relative sharing of process conditions, etc.

[0076] "Micro-contexts" can be data-driven, as shown in Figure 2. Arranging contexts across granularity (from "large" to "small," or vice versa) and considering further hierarchy (i.e., subcontexts) can be computationally advantageous. Many context details ("small") may require more computation than few context details ("large"). reference concept

[0077] In Figure 1, the reference concept is also shown with vertical lines. The reference concept is complementary to the context concept.

[0078] Component 110 (in machine 100), component 110-xx-f (in a further machine 100-f), and component 110-xx-F (in a further machine 100-F) are similar components. One machine may have two or more similar components. As used herein, a component is, • Common structural elements, • Common parameters related to its structural elements and If they have [a certain characteristic], the components are similar.

[0079] Therefore, similar components can be used as references (for several purposes, such as harmonizing parameter data).

[0080] Although not described herein, industrial machinery can similarly have multiple components that are similar to one another. Component states and machine states as data derivatives

[0081] Figure 2 also shows a state that may change over time. The data points within {{X}} (and {{Y}}) can usually be processed according to predefined rules to obtain the state through aggregation. It is not necessary to process all the data; processing only some of the variables is usually sufficient.

[0082] Figure 2 illustrates states S0 and S1. Simply put, at any given time, machine 100 (or component 110-xx or 110-yy) can be in a specific state. States can be distinguished into machine states (states derived from {{X}} and {{Y}}) and component states (states derived from the {{X}} of the component).

[0083] A state can contribute to the reference and context (i.e., to the determination of the reference and context). A state can be related to the process performed by machine 100. In other words, contexts 501 and 502 are sets or data points, but further contexts are • A set of data points and time series having states (as the series S0, S1, S1, S1, S1 in Figure 2), and • Time series having states (as the sequence S0, S1, S1, S1, S1 in Figure 2), and sets of such time series It can be defined by: Examples

[0084] For example, component 110-xx could be a container (or tank) of industrial machine 100. Data point Xmn should represent the volume of material (such as a liquid) poured into the container. In an industrial process, component 110-xx should receive material at a constant volume / time ("material flow" as variable 1 in {{X}}) and a constant temperature (variable 2 in {{X}}). Some time points belong to the current performance of the process (see context 501), while others (e.g., t1) do not. Again, this is a simplification, but time points in backview counter K are "real-time", others are "historical".

[0085] In this embodiment, since the container is filled with material (during that time), the material flows into the container, and the temperature of the material at a particular past point in time gives context 1. Since the container is emptied from time to time (for cleaning or other maintenance), not all past points in time belong to the context.

[0086] Regarding similarity, component 110-yy may be a container, and the {Y} series of variable n should also represent the volume of the material (i.e., data point Ymn). Although Ymn has a different context (not context 501), due to similarity (both are containers, and so are variables 1, 2 and n), component 110-yy can function as a reference.

[0087] Regarding the state, S0 can be a state in which the volume of material does not change, and S1 can be a state in which the volume of material increases or decreases. In this embodiment, the rule is simple: a variable 1 in which the material enters (or leaves) the container and a variable n that increases over time describe state S1; otherwise, the state transitions to state S0.

[0088] As already mentioned, states (and state transitions) can belong to a context, but they do not have to.

[0089] The explanation will return to such simplified examples when describing network training. Data flow to the operator

[0090] As described above, the operator interacts with the machine remotely, and Figure 1 illustrates this method by showing the (computer) user interface 200 and operator 290. In principle, operator 290 can access {{X}} (and {{Y}}), see Figure 2. The data is enhanced by placing data points in contexts (contexts 501, 502, etc.), establishing inter-component references, and further aggregating the data into states (see S0, S1 references).

[0091] Such data enhancement can be seen as a means to assist the operator 290. For example, the operator 290 may see a digital component equivalent 210-xx in the user interface 200 that tells the operator that the liquid volume in component 201-xx has changed (i.e., aggregated state information, state S1), that component 110-xx is involved in a particular process step (i.e., that is also context), and that a particular relative volume (information obtained by referring to other components 110-xx-f or 110-xx-F acting as peers).

[0092] The bidirectional arrow DATA indicates that data from machine 100 goes to the computer, and control data goes to machine 100, at least to some extent.

[0093] The unidirectional arrow DATA indicates that the historical data hh is also available to operator 290.

[0094] DATA can include pre-processed data such as context, state, and references, but the physical location of the pre-processing is not important. Flow constraints

[0095] As mentioned above, various data conventions can lead to incorrect activity, and worse, they can further interfere with digital component equivalents 210-xx. They can also lead to incorrect determinations regarding context, state, references, etc.

[0096] In other words, similarity (of components, etc.) does not mean that data equivalence (of those components) exists. The data in {{X}} and {{Y}} are different.

[0097] The structural elements may differ in quantity (see the example in Figure 3, which has small and large containers and containers of various sizes), and the parameters, while common, may be represented by various data structures. solution

[0098] The constraints necessitate solutions. These are presented below by harmonizing the parameter data with network-aware transformation techniques. Optionally, the network takes into account (one or more) contexts.

[0099] In this specification, the term "conversion" may be used because, at both the semantic and syntactic levels, the computer transforms source data into target data, and in the case of multiple different sources, the computer provides the target data in a single format.

[0100] However, computers do not perform natural language processing. The data is not available in this form. Rather, computers perform transformations that can be likened to compiling source code into machine code (transformation between artificial languages, not between natural languages), but the transformation function needs to learn the details before applying them. In this specification, learning is described as training one or more networks. supplement

[0101] Commercial translation computers may use pre-trained networks to translate natural language text from a source language to a target language. The computer may be trained to recognize patterns. For example, the German phrase "ich bin gerade dabei, ein Buch zu..." means "I am trying to write a book." The ellipsis can be replaced with a verb such as "schreiben / write" (to write) or "lesen / read" (to read). The computer's training data appears to have far more statements for "write" than for "read." A relatively high frequency (in the training data) determines the selection of a particular pattern in the target. Therefore, the computer translates to the target "I am trying to write a book" even if it doesn't specify whether the source should be "write" or "read."

[0102] However, in the case of source and target data for industrial machinery, the frequency of a particular event is not important. Component 110-xx (if it is a container) may eventually overflow (even in the true sense of the word, the liquid will leak out of the container), but such an event (or even a failure condition) may occur relatively infrequently in the historical data (see hh). Nevertheless, confusion or assumptions to the opposite meaning must be avoided. Even if overflow events are rare, they must be correctly identified within the target.

[0103] Optionally, the solution takes context into account. This approach avoids "conversion" to majority-preferred target conventions in situations where minority occurrences apply. Components, transformations, and component equivalents

[0104] Figure 3 shows Multiple system configurations 110-xx, each having a digital sensor 120-xx that provides source data 221-xx. • A conversion function 300 (dotted line) for harmonizing parameter data from a digital sensor by implementing a computer implementation method (see method 400 in Figure 7A) that converts source data 221-xx to target data 222-xx, and • Multiple digital component equivalents 210-xx (or “digital component twins”) that process target data 222-xx (which, in the simplified embodiment of Figure 3, are linked to symbols 252-xx on the user interface). This indicates that.

[0105] Figure 3 also shows the operation of data conversions 430-xx and 440-xx, in which the conversion function 300 converts source data 221-xx to target data 222-xx. Reference numbers 430 and 440 are for the method (see Figure 7A).

[0106] The dashed line representing the conversion function is also shown in other figures. Such lines indicate the application of the method in use case scenarios (see Figure 9).

[0107] Multiple elements (components, operations, twins, etc.) are distinguished by a two-digit index xx = (11, 12, 21, 22, 23). The same xx index also indicates a relationship. Elements with the same xx are related, and the diagram shows such relationships horizontally (e.g., component 110-11, transformations 430-11 and 440-11, equivalent 210-11).

[0108] A digital equivalent 210-xx can be a model describing its components. In this specification, in this figure, the equivalent is simply “described” by a user interface symbol (e.g., symbol 252-xx in Figure 3).

[0109] Components 110-xx are similar in that they all have a specific parameter (e.g., level or volume), or a "common parameter"; see the introduction of this concept in Figure 1. Components 110-xx can belong to various technical systems, and they do not need to interact with each other. For example, components 110-11 and 110-12 belong to group 101, but they do not need to interact with each other, and components 110-11 and 110-12 have their own context, which can change over time. The parameters are common, but they do not need to have the same parameter values.

[0110] The figure is simplified in that it shows a specific component 110-xx having a single digital sensor 120-xx that provides (digital) source data 221-xx (which are converted to 430-xx and 440-xx). The digital sensor 120-xx provides data for at least common parameters. Of course, the component may have other sensors for other parameters not described herein. Some sensors may provide data for two or more parameters.

[0111] The conversion function 300 (performing steps 430-xx and 440-xx of method 400, Figures 7A and 7B) provides the (digital) target data 222-xx to a specific equivalent 210-xx. Each conversion 430-xx and 440-xx is represented by a rightward arrow.

[0112] As shown by the dashed rectangle, components 110-11 and 110-21 (and their digital sensors 120-11 and 120-21) form a first component group 101, and components 110-21, 110-22 and 110-23 (and their sensors 120-21, 120-22 and 120-23) form a second component group 102.

[0113] Both groups can belong to various physical locations. For example, a group can be located in two halls in the same place, or belong to two different machines (see Figure 1). As shown by the dashed lines, the digital equivalents 210-11, 210-12, 210-21, 210-22, and 210-23 are available to a single computer (receiving target data 222-11, 222-12, 222-21, 222-22, and 222-23). ​​Operator 290 views the symbols of the components on a screen or the like. The figure simplifies by showing one operator 290, but there may be multiple operators for different groups, for example. Operating components and / or industrial machines from various physical locations while supplying data to a single computer may impose or exacerbate constraints. Some of these constraints are described below.

[0114] Of course, an industrial machine has multiple components (see Figure 1 for xx and yy), and each component usually has two or more sensors. From another perspective, groups 101 and 102 can represent two industrial machines having further components and sensors.

[0115] Component 110-xx provides source data from many parameter-specific sensors, and the simplification to a single digital sensor 120-xx represents a representative of sensors that may be pressure sensors, temperature sensors, volume level sensors, power sensors, material color sensors, etc.

[0116] Component 110-xx has common parameters and properties. Since the figure focuses on data transformation, it should be assumed that source data 221-xx and target data 222-xx should relate to common parameters (common to all components 110-11 to 110-22). Parameter values ​​can be provided for all components 110-11 to 110-22 at any given point in time (e.g., at tm, or during a specific time interval, at a sensor reading time associated with tm). Those skilled in the art will understand that digital sensors 120-xx may not provide source data 121-xx at exactly the same point in time, but for the purposes of this explanation, it is not relevant whether, for example, sensor 120-11 provides data a few seconds later than sensor 120-12.

[0117] For the sake of simplicity, the diagram shows an embodiment that is easy to visualize. The components should be the aforementioned containers (or tanks, or reactors, etc.) that hold a specific volume of material. A common characteristic should be that the volume level parameter is a common parameter that applies to all containers and that must be handled by the digital equivalent 210-xx. In the diagram, this parameter is represented by bidirectional vertical arrows (inside the components).

[0118] Digital sensor 120-xx represents a parameter value (a level having a volume value, as specified herein), and digital sensor 120-xx should provide source data 221-xx as a parameter identifier in combination with the parameter value (further details also in Figures 7A and 7B). For example, in the xx notation (11,12,21,22,23), the parameter identifier is (In,In,Vol,Vol,Vol) to indicate that the volume of material in a container was measured (or otherwise identified, see Figure 9). Note that having common parameters does not mean having common identifiers. Parameter values ​​are, for example, (1.5,3.0,3.0,1.0,2.0). For simplicity, values ​​are given without measuring units in this specification. Note further that having common parameters does not mean using common values ​​(with respect to data format, with respect to units of measurement, etc.).

[0119] For the sake of simplicity, it can be assumed that operator 290 needs to have an overview of the relative material levels in each of the components 110-xx. In the diagram, the relative material levels correspond to the length of the vertical arrows. However, this does not necessarily correspond to the sensor readings.

[0120] The conversion can be defined manually by examining identifiers. The identifier "In" is derived from "in-take" (or the German word "Inhalt" meaning "content"), and "Vol" is derived from "volume". Since sensor 120-xx is a level sensor (in the examples herein), the first conversion rule can be defined for the syntax of converting "in" to "V" and "Vol" to "V".

[0121] Regarding the parameter values, a second transformation rule can be defined by taking into account the overall component capacity, and the rule is: • In the case of component 120-11, divide the value by 2, • If the components are 120-11, -21, -22, and -23, divide the value by 4. It can be defined as follows.

[0122] Such rules are thought to support semantic transformation, where absolute level values ​​are converted to relative level values.

[0123] Applying the rule, the data is transformed to (V 0.75, V 0.75, V 0.75, V 0.25, V 0.5), i.e., the target data 222-xx. This figure symbolically represents the target data 222-xx in the user interface by black dots on the center line (above, above, above, below the line). Operator 290 may serve as an indicator for drawing conclusions. For example, operator 290 may investigate why components 110-22 have unused space.

[0124] However, manually defining the conversion may not cover all possibilities. The following is an example: • Human definitions are prone to errors, and confusion may only be detected when the digital equivalent begins to return control signals. This is merely an overview of potential limitations. In this embodiment, the control signal to stop the injection of material may be too late (leading to overflow). (Since only three symbols are defined as "on the line," "on the line," and "below the line," see the black dots in symbol 252-xx in Figure 3.) User interface symbols within equivalent 210-xx can have only three states. However, a simple calculation using the division rules does not necessarily identify the position of the dot. For example, dividing the measured value 1.2 by 2 to obtain the calculated value 0.6 does not yet give the position of the dot. • Some elements of the conversion calculation may change over time. For example, repairs may upgrade component 110-11 from a "small" container to a "large" container, while sensor 120-11 remains the same. However, this simple change will affect the division rule (dividing by 4 instead of 2). There is no guarantee that the rule will be updated. The identification using "In" and "Vol" is provided herein for illustrative purposes only; however, in actual implementations, such short words may be sufficient to distinguish between multiple sensors. Sensor data 221-xx may not be encoded in a human-readable format. For example, "In" could be misinterpreted as "input". • When new components are added (or old components are removed), conversions must be adapted. The same applies to sensor replacements. (As a side note, in virtual reality, such as twin systems, components can be added or removed to simulate industrial systems, but in either case, conversions must be adapted.) • Conversion rules must be technically implemented. If the rules are implemented on a sensor (i.e., using electronics on the sensor device), appropriate software must be added to the sensor. However, certain sensors may be part of a control loop that must not be modified, otherwise the technical safety of the components can no longer be guaranteed. In other words, the original data convention, i.e., the control loop using the source data, may have been tested to comply with the technical standard, and therefore the original data convention must be used, otherwise compliance with the technical standard cannot be guaranteed. If the rules are implemented "right-hand" (i.e., by expression 210-xx), the updated rules may not be applicable to all components, and the updated rules cannot interfere with standard-compliant control loops. • Metadata may not correctly identify a particular sensor. Incorrectly assigning a specific ID to a particular sensor can occur in various situations, such as when a sensor is first registered in the database, when a sensor is repaired or replaced, or when the database is modified.

[0125] Placing components in physically different locations can lead to increased constraints.

[0126] The container examples are simplified because each container ultimately represents all possible parameter values ​​(from 0 to full). However, as a rule, some sensor readings may occur less frequently than others, and the conversion (or transformation) from source to target must be accurate and independent of frequency. Transformation function implemented by neural networks

[0127] The constraints can be alleviated by using machine learning transformation techniques. Machine learning relies on considering the context of historical data, optionally.

[0128] Figure 4 again shows the mechanical components 110-xx and digital equivalents 210-xx from Figure 3, but shows a conversion function 300 implemented by a neural network 300 for harmonizing parameter data 221-xx (source data) from sensor 120-xx into converted data 221-xx (target data). (Reference number 300 is used throughout the drawings for network-based functional and structural implementations).

[0129] As described below, the transformations 430-xx and 440-xx (in Figure 3) can be implemented by one or more pre-trained networks (or subnetworks 330 and 340). Network 300 is an overall element that can be implemented by one or more dedicated transformation networks. Network 300 is shown using multiplexers for its input (receiving source data) and output (providing target data), and multiplexing is just one example of a technique that represents the need to separate data belonging to various components.

[0130] This network approach changes from manually defining transformation rules to defining them using machine learning. This may be advantageous in that it can alleviate the constraints mentioned above.

[0131] Examples of source and target data are single data values ​​that apply to a specific point in time, but data may also be available in further dimensions (such as space and time). Network Functions

[0132] Figures 5 and 6 are shown below. • Subnetwork 330 provides syntax translation, • Subnetwork 340 provides semantic translation. The two functions of network 300 for various conversion purposes are shown above; please refer to Figure 3.

[0133] The figure shows network 300 during training (i.e., before the method is implemented). In the figure, the index changes from xx to hh simply to show that the training stages in Figures 5 and 6 are performed before the operational stages in Figures 3 and 4.

[0134] Data from component 110-hh can provide data from the same component 110-xx (i.e., hh=xx), but a component provides data for training, and the trained network transforms data from any component. A specific component (i.e., a component with a particular xx index may not yet be working), but training originates from components that have worked in the past (i.e., have historical data). It is advantageous that component 110-hh, which provides historical data, is similar to component 110-xx in that the transformation should be applied to it (see Figure 3, real time). Using a (training) context can prevent the network from being trained on incompatible data (e.g., due to dissimilarity).

[0135] This specification takes into account the separation into groups. For example, training within groups is preferable to training across groups because groups can evoke a variety of contexts.

[0136] The attributes "Syntax" and "Semantic" are provided herein solely for explanatory purposes. The attributes indicate an order in which parameters are identified first, and their values ​​are viewed second.

[0137] These may sometimes be referred to as "auxiliary networks." Such networks operate with different known characteristics than the networks shown in Figures 5 and 6.

[0138] In other words, the history source identifier 231-hh, the history target parameter identifier 232-hh, the history source value 241-hh, and the history target value 242-hh can be obtained from one or more further components 110-hh similar to the components 110-xx of the industrial machine 100. Training (First Subnetwork)

[0139] Figure 5 shows the neural network 330 being trained on historical data. The neural network 330 (i.e., a subnetwork of network 300) has the ability to provide syntactic translation (i.e., in the embodiment where parameter identifiers rather than parameter values ​​are harmonized).

[0140] In this embodiment, the historical data for training is a combination of parameter data (from the past) and artificial annotations "V" (as ground truth).

[0141] The diagram refers to a simplified embodiment of Figure 3. The network learns that "In" and "Vol" (in input 331) must be matched with "V" (in output 332 during training).

[0142] This figure also shows that random and accidental errors in the training data do not have an impact. Artificial annotations "P" ~ "In" do not lead to a rule simply because the number of such annotations is relatively small. Or, more precisely, annotation "P" is statistically not representative because, by its nature, it is an outlier or anomaly. Data processing techniques can be applied beforehand, thereby ignoring non-representatives for training. One embodiment of such a technique is the application of an autoencoder, which compresses the data (thus ignoring P) and expands the data (data without P).

[0143] Similarly, there may be data formats called "Inhalt" that lack sufficient annotation. In other words, some sensors, especially if they are corrupted (or miscalibrated), provide incomprehensible data, making it impossible to create annotations. However, such unannotated data is not used as training data.

[0144] Once training is complete, network 330 receives metadata of the same type (e.g., all pressure sensors) (from various data sensors) and provides harmonized identifiers.

[0145] Optionally, historical context data 261-hh is also provided to input 331. Data 261-hh is provided simultaneously to identifier pairs 231-hh and 232-hh, and the data corresponds temporally, for example, CONTEXT_A, where state 0 is the context applied to the four pairs on the left side of Figure 5, and the network in training processes the context as it processes these individual four pairs. The network receives other context data (CONTEXT_B) when processing the next pair, and so on.

[0146] at least In the first option, the computer processes a multivariate time series {{X}} with (virtually all) variables n=1 to N, see Figure 2. These are some of the options available. The figure shows context data synchronized with annotations and parameter identifiers, such as {X}1 representing the variable at t=t1. The time at index m progresses from left to right. In the second option, the computer processes {{X}} but only as a subset of data that is eligible as context (see Figure 2, showing contexts 501 and 502 within a dashed box). In the third option, the computer processes context descriptors such as "context 1" (t1-t4) and "context 2" (t5-t9), and also see Figure 2. In the fourth option, the computer processes state descriptors for embodiments having states S0 and S1; see Figure 2.

[0147] Providing context during training can be advantageous as it also addresses low-frequency constraints. In certain contexts, the majority of "In" and "vol" are annotated with "V," and the artificial annotation "P" for "In" does not lead to a rule as described above. However, in various contexts, the network learns this, and such annotations may point to events with relatively low frequency. Such events (e.g., overflows) must also be reported in the digital equivalent 210-xx, provided that the parameter data is properly harmonized. Training (Second Subnetwork)

[0148] Figure 6 also shows that the neural network 340 is trained on historical data. The neural network 340 has the function of providing semantic translation. • Sequence of measured values: 0.75, 0.75, 1.00, 0.00, 0.10, 0.50 (see Figure 242-hh), • Corresponding sequence of component sizes: 2, 2, 2, 2, 2, 2 (The figure shows only the values ​​for the "small" component 110-11 in Figure 3 of reference 241-hh) There are two inputs, as shown above.

[0149] The ground truth (output to the network during training) is the sequence of annotations 0.75, 0.75, 1.00, 0.00, 0.10, 0.50. In this example, the annotations have the following meanings: "three-quarters", "three-quarters", "full", "empty", "almost empty", and "half full".

[0150] The annotation can be matched, for example, with the dot symbol for the equivalent 210-xx in Figure 3. During training, the historical data will eventually include 1.2 in the first row, along with the annotation "half full".

[0151] As shown in Figure 5, network 340 (when being trained) can optionally be trained with historical context data 271-hh. In such a case, data 271-hh can be processed simultaneously with parameter value pairs 241-hh and 242-hh. The figure is simplified when writing the context, but the context can change over time (see the explanation of contexts 501 and 502 in Figure 2). Context 261-hh (in Figure 5) and context 271-hh (in Figure 6) can be the same data, but in most cases they are different (because the historical data may come from various points in the past).

[0152] Although the explanation described networks 330 and 340 separately, both networks can be implemented in a single network (sub-functions 330 and 340 are network modules). Training Overview

[0153] In other words, training can be described as a computer implementation method for training network 300 (which has a first subnetwork 330 and a second subnetwork 340). Network 300 is trained for subsequent use in a computer implementation method 400 (see Figures 7A and 7B) for harmonizing parameter data from multiple source protocols into a single target protocol. The data flows from input 331 / 341 to output 332 / 342 through the first and second subnetworks 330 and 340 (see Figures 5 and 6).

[0154] As described above, the parameter data pertains to component 110-xx of industrial machine 100. Parameter dataset 221-xx represents the technical parameters belonging to component 110-xx, and parameter dataset 221-xx has parameter identifier 231-xx and parameter value 241-xx. In the source specification, there is source parameter identifier 231-xx and source parameter value 241-xx. In the single target specification, there is target parameter identifier 232-xx and target parameter value 242-xx.

[0155] The method includes training the first subnetwork 330 with a plurality of historical parameter identifier pairs 231-hh, 232-hh, each having a historical source parameter identifier 231-hh at the input 331 of the first subnetwork 330 and a historical target parameter identifier 232-hh at the output 332 of the first subnetwork 330.

[0156] The method further includes training the second subnetwork 340 with a plurality of historical parameter value pairs 241-hh, 242-hh, having historical source values ​​241-hh at input 341 of the second subnetwork 340 and historical target values ​​242-hh at output 342 of the second subnetwork 340.

[0157] The overall method can be defined by a computer implementation for training the network and a computer implementation for harmonizing the parameter data (see Method 400 and further details described herein). training accuracy

[0158] Source and target data are represented by data elements in a predefined order of identifiers and values. This order can be changed.

[0159] However, such ordered data has some remote similarity to code in a programming language (such as source code), and similarity to text in a human language.

[0160] In the case of computer languages, a compiler or interpreter is established.

[0161] In the case of human language, training can be likened to training a language converter. The source is a sequence of words (time series) such as "vol" and "In," and the target is a sequence of standardized words ("V," etc.) (which is also a time series). Because artificial text inevitably exhibits errors (see history mismatch in Figure 5), language converters are robust against such errors.

[0162] The accuracy relates to the quality of the control data from the equivalent 210-xx. Continuing with the container embodiment, the computer can be programmed to stop the material flow in a context (or state) that constitutes "overflow." It does not matter whether such states are aggregated individually by components (from the source data) or aggregated by the computer (from the target data, e.g., when V reaches 1). However, in the second case, the conversion to the target data must be accurate. Incorrect state data will result in missing the stop signal.

[0163] Taking the context into account (already during training) and now (i.e., at point tm where the parameter V can reach 1) can reduce the risk of malfunction. Details of the implementation

[0164] The appropriate networks (330 / 340 and 300) are as follows: • RNN (Recurrent Neural Network) • LSTM (Long-Term Short-Term Memory Network) • CNN (Convolutional Neural Network) • GRU (Gated Recurrent Unit) GPT That is correct.

[0165] An overview of such a network can be found in the paper "A Novel Encoder-Decoder Model for Multivariate Time Series Forecasting" by Huihui Zhang et al., Comput Intell Neurosci. 2022;2022:5596676, published online on April 14, 2022, doi:10.1155 / 2022 / 5596676.

[0166] Time series forecasting using deep learning is described in the following research paper, Lim, B, Zohren S. 2021 Time series forecasting using deep learning, survey.Phil.Trans.R.Soc.A379:20200209. https: / / doi.org / 10.1098 / rsta.2020.0209.

[0167] Regarding RNNs, "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation" by Kyunghyun Cho et al., arXiv:1406.1078v3, is also useful.

[0168] Regarding context, Figures 5 and 6 are simplified by showing a single input 331, 332 for networks 330 and 340 to receive context (during training as shown in Figures 5 and 6, and during operation, see Figure 3).

[0169] A 2D context (such as context 501) can enter a neural network via a separate network layer. Techniques for receiving data in 2D vector arrays are known in this field, for example, in the paper "CROSSFORMER: TRANSFORMER UTILIZING CROSSDIMENSION DEPENDENCY FOR MULTIVARIATE TIME SERIES FORECASTING" by Yunhao Zhang and Junchi Yan (MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University and Shanghai AI Lab, presented as a conference paper at ICLR 2023). Transfer learning

[0170] Referring back to Figure 3, various groups (components 110-11 / 110-12 and 110-21 / 110-22 / 110-23) can have different operational histories. One group may have been operating for many years, while another group may belong to a new location. The same applies to a single component. A new component may function as a replacement for an old component, and there may not be enough time to manually replace all source-versus-target relationships.

[0171] Components without a history (whether grouped or not) do not provide historical data, but historical data from "long history" components can be stepped in to help the transformation learn.

[0172] Such transfer learning is further supported when descriptive data about “long history” components is added to the training data. The training data (see Figures 5 and 6) is not limited to history identifiers and values, but can also include component type identifiers (e.g., large container, small container). For example, if “small size” components (such as component 110-11) are added, the network applies transformations to such “small size” components.

[0173] Such identification is an optional example of a context that can support training (see 261-hh in Figure 5). Method flowchart

[0174] Figure 7A shows a flowchart of a computer implementation method 400 for harmonizing parameter data 221-xx from sensor 120-xx. This figure shows a classic flowchart with method steps 410, 420, 430 and 440 on the right, but also shows first and second subnetworks 330, 340 with inputs and outputs.

[0175] Computer implementation method 400 is a method for harmonizing parameter data.

[0176] In receiving step 410, the computer receives a source parameter dataset 221-xx representing technical parameters (from component 110-xx of the industrial machine 100, see Figure 3). Technical parameters belong to components. The source parameter dataset 221-xx has a source parameter identifier 231-xx and a source parameter value 241-xx.

[0177] The computer uses the first subnetwork 330 of the pre-trained network 300, and in matching step 420, the computer matches source parameter identifiers 231-xx with target parameter identifiers 232-xx. (As illustrated in Figure 5, the first pre-trained subnetwork 330 is trained with multiple historical parameter identifier pairs 231-hh, 232-hh, each having a historical source parameter identifier 231-hh in its input 331 and a historical target parameter identifier 232-hh in its output 332.)

[0178] In step 430, the computer uses a second subnetwork 340 of the pre-trained network 300 to match, matching source parameter values ​​241-xx with target parameter values ​​242-xx. As explained and illustrated by the arrow “Selection” in Figure 7A, the second subnetwork 340 (of the pre-trained network 300) is selected according to the target parameter identifier 232-xx. (As explained in Figure 6, it is trained with multiple historical parameter value pairs 241-hh, 242-hh, having historical source values ​​241-hh in its input 341 and historical target values ​​242-hh in its output 342).

[0179] In the transfer step 440, the computer transfers both the target parameter identifier 232-xx and the target parameter value 242-xx to the user interface 200 (see Figure 3), which displays a user interface element 210-xx that corresponds to the component 110-xx of the industrial machine 100 and visualizes the target parameter value 242-xx.

[0180] Source parameter dataset 221-xx represents technical parameters that optionally belong to component (110-xx) and are either source parameter datasets from a digital sensor or source parameter datasets from a virtual sensor (see Figures 8-9).

[0181] Optionally, the first subnetwork 330 is trained with the set of parameter identifier pairs 231-h, 232-h, in addition to being simultaneously trained with the historical context data 261-hh at its input 331 (see Figure 5). Optionally, following the same principle, the second subnetwork 340 is trained with the set of parameter value pairs 241-h, 241-h, in addition to being simultaneously trained with the historical context data 271-hh at its input 341 (see Figure 6).

[0182] Optionally, receiving step 410 further includes identifying the real-time context 501, 502 (see Figure 2) in which its component 110-xx is currently operating. The real-time context is a set of data points related to data points (Xmn) in the source parameter dataset 221-xx of technical parameters. Identifying the real-time context may include accessing state data (S0, S1), see Figure 2.

[0183] Optionally, in transfer step 440, the computer uses the user interface 200 to visualize the target parameter value 242-xx by the symbol 252-xx (see the example with dots on the line in Figures 2-3).

[0184] Optionally, after performing step 410 to receive (source parameter dataset 221-xx), the computer uses the first and second subnetworks 430,440 to determine whether the source parameter dataset 221-xx conforms to the rules of the target parameter dataset 222-xx. If the determination is positive, method 400 continues with transfer 450, thereby skipping steps 430,440 to perform the matching. In other words, if the transformation rules are available outside the network, the rules are applied. Missing identifiers and other circumstances are described in detail below.

[0185] Optionally, after transfer 400, control signals are acquired to control the operation of components 110-xx of the industrial machine 100. In other words, the computer becomes part of the control loop of the components, and data harmonization is one technique for acquiring the data to execute that loop. Using Context Types in Network Selection Cascading

[0186] Figure 7B, like Figure 7A, shows a flowchart of a computer implementation method for harmonizing parameter data, but also demonstrates the application of a context-dependent network selection cascade.

[0187] As already mentioned in the explanation of Figure 7A, it is possible to identify the real-time context (component 110-xx is currently operating), and as explained in Figures 5 and 6, the context can be supplied to the network as data (e.g., historical context data 261-hh in its input 331 in Figure 5 during training). The network (e.g., network 330 in Figure 5 and network 340 in Figure 6) then processes the context along with parameter identifiers and parameter values.

[0188] The arrow "Selection" in Figure 7A already indicates a selection, and output 232-xx (output 232-xx, which is the target parameter identifier for the first subnetwork 330) functions as a selector for a specific second subnetwork 340.

[0189] Figure 7B shows a method 400 similar to that in Figure 7A, with the additional option that a context (i.e., a context type, see context_A, context_B, state 0, state 1 in Figure 5) can also be used to select a particular first subnetwork 330.

[0190] In that sense, there exists a context network 315 that identifies contexts (this conceptually involves identifying context types, which are coarse-grained contexts, given in this specification as a result of an additional step 415). The identified context types are selectors for the first subnetwork 330. The number of trained subnetworks 330 corresponds to the number of context types. Figure 7B shows a dashed rectangle representing this number behind network 340. The number is not limited to 3 (as shown in the figure).

[0191] In that sense, Figure 7B is, • Determine the context type to select a specific first subnetwork 340 from multiple pre-trained networks, followed by step 420 as described in Figure 7A, and As explained using Figure 7A, the second subnetwork 340 is determined according to the target parameter identifier 232-xx. This shows a selection sequence having the following characteristics.

[0192] Such a selection sequence may also be called a cascade having a first decision step and a second decision step.

[0193] The context network 315, which identifies context types, can also be implemented using a neural network. It can be trained with context annotations (contexts 501 and 502 are merely examples) and context type annotations. In relation to subnetworks 330 and 340, the context network 315 is an auxiliary network.

[0194] To summarize this section, Method 400 is shown in one embodiment in which, after receiving a source parameter dataset 410, a context network 315 is used to identify a context type 415, and further, a specific first subnetwork 330 is selected according to the context type to match the source parameter identifier with the target parameter identifier 420. Sensor implementation configuration

[0195] Figure 8 shows components 110-V, 110-R, and 110-RV with various sensor implementation configurations. As described above, digital sensors are computing devices that sense physical phenomena (i.e., parameters) and provide parameter values ​​(and identifiers, if available) in the form of digital data.

[0196] Depending on the implementation of sensing, digital sensors can be distinguished into hardware sensors and software sensors. Similarly, this distinction can also be used to distinguish between a real sensor 120-R and a virtual sensor 120-V. Other synonyms are "physical sensor" and "non-physical sensor."

[0197] Component 110-R is a component having one or more sensors implemented as a real sensor 120-R, and component 110-V is a component having one or more sensors implemented as a virtual sensor 120-V.

[0198] Real-world sensors, in their broadest sense, have a physical element that comes into contact with that part of the mechanical component to which the parameter is applied. The following examples are useful for understanding. A pressure sensor may have a membrane that comes into contact with a gas or liquid, and the pressure value can be obtained by electronically sampling the bending of that membrane. A temperature sensor may have a temperature-sensing element (e.g., a resistor, transistor, thermistor, etc.), which is surrounded by a material whose temperature needs to be known. Measuring the current flowing through the element yields a temperature value. The temperature sensor can be implemented as an infrared thermometer (or an infrared camera that measures the temperature of the entire surface). The sensor is in "contact" via thermal radiation.

[0199] In contrast, a virtual sensor has no physical element that could have such contact (with that part of the mechanical component to which the parameter is applied). In other words, a virtual sensor can be the output or rule of a simulation model or machine learning model, since it provides a signal that is not measured by a physical instrument by definition. Techniques for implementing such virtual sensors are described in International Publication No. 2022 / 194871.

[0200] Those skilled in the art can select an appropriate sensor according to the structure of the components and the parameters. It may be impossible to install a real sensor because the components simply do not allow it. However, deriving parameter values ​​using a virtual sensor can be an alternative. For example, measuring the temperature of molten material in a furnace can be difficult because it is hard to place a temperature-responsive element (along with cables, etc.) inside the furnace. However, using a virtual temperature sensor can provide a temperature value (with accuracy suitable for most applications).

[0201] In both cases, parameter data is available through direct measurements (real-world sensors) and indirect measurements (virtual sensors).

[0202] As a result, the virtual sensor can obtain parameter values ​​for situations where it is impossible to use real-world sensors (i.e., hardware sensors).

[0203] Component 110-RV is a component having one or more sensors implemented as a combination of a virtual sensor and a real sensor. The figure shows a selection logic that outputs sensor signals. The logic may select data in real time according to the probability that the data actually corresponds to parameters, with data of higher probability being preferred. Those skilled in the art can implement such selection, for example, by using an auxiliary neural network (trained to determine probabilities), by applying the Viterbi algorithm, or by other means.

[0204] In all cases, the components 110-R, 110-V, or 110-RV have their digital sensors for providing actual data (i.e., data describing parameters).

[0205] The design choice for using 110-R, 110-V, or 110-RV depends on the technical constraints of the components and the availability of data. • Contact or access to parameters that are advantageous to real-world sensors, • Knowledge of the relationships between parameters (even if acquired through machine learning within the auxiliary network) enables the use of virtual sensors. These two examples are noteworthy.

[0206] For example, while the temperature and pressure inside a furnace (i.e., parameters within the system) may be difficult to measure with real-world sensors, a relationship between temperature and pressure and wall temperature can be established (either by applying formulas or by using a trained network). Since wall temperature can be measured (by real-world sensors), this relationship makes it possible to estimate temperature and pressure (i.e., virtual sensors). From parameter values ​​to system state

[0207] As already shown in Figure 2, a person skilled in the art can abstract parameter data (as source data or target data) into system state data. For example, the data "V 0.5" and "V 0.75" can be summarized as "normal" operation, and the data "V 1.0" can be summarized as "abnormal" operation. Other data can also be considered. For example, data showing the flow of liquid into a tank in combination with "V 0.75" may lead to the state "overflow expected".

[0208] Harmonization can also be applied to states, and states such as "overflow expected," "overflow imminent," "more than full," and other states can be harmonized (by network 300 if appropriately trained). From components to systems

[0209] Figure 9 shows the first, second, and third industrial machines 100, 100', and 100', each having multiple components 110. The components can be implemented as R, V, or RV (see Figure 8). Once the components provide source data, the transformation by the network 300 yields target data. If the data remains isolated, the same network can be used for all transformation tasks. Of these multiplexing techniques, isolation techniques are known (see Figure 4). Transformation may not be necessary in all cases; that is, the source data may already be in the target format (identifier, value, etc.). Optionally, the network 300 determines whether the source data is already in the target format.

[0210] Therefore, the rightward arrows represent harmonized data (in terms of identifier / syntax and value / meaning) directed towards the twin components 210, 210' and 210'' of the twin 200, 200' and 200''. The digital equivalents 210 / 210' / 210'' can provide control signals. The diagram is simplified by showing only one control line.

[0211] Since the parameter data (from the components) and the control data (generally for the components, especially their actuators) have the same structure (i.e., by harmonizing with the conversion function 300, as described, by deharmonizing or adapting), the parameter and control data can be placed into the pool of historical data. The pool can be distinguished from the original (data associated with the first, second, or third system).

[0212] Those skilled in the art can provide historical data by applying calculations. For example, if the machine is a first blast furnace having a first volume, the activity of charging a first amount of material into this first furnace has the same cross-multiplication as charging a second amount into a second furnace having a second volume.

[0213] When a new component (or new system) is commissioned, historical data is used for various purposes, among others, • Even if the new component or system is not yet operational, it can be used in the prediction model to forecast the behavior of the new component (or system). It is possible.

[0214] Therefore, in Figure 9, system 100" is shown with a dashed line. Its twin system 200" can be installed in front of system 100". Relationship between parameters

[0215] Here, we will explain the relationships between parameters, and therefore between parameter data. Harmonization is also available.

[0216] Two components with essentially the same structure can have varying numbers of sensors. Whether the sensors are implemented as R or V sensors is irrelevant.

[0217] Continuing with the above examples of containers or tanks, the components not only need to hold a liquid stored in a specific volume, but also need to store the liquid at a specific temperature. The type of liquid should be the same in both cases.

[0218] The first component should have a level sensor and a temperature sensor, while the second component should have only a level sensor. Alternatively, both components should each have two sensors, but in component 110-B, the temperature sensor may fail.

[0219] The data can be collected as {{X}} for the first component and as {{Y}} for the second component, and the level and sensor data are variables.

[0220] A level sensor can be implemented using a digital meter to measure the distance from the top of a container to the surface of the liquid inside the container. As the volume of the liquid increases, the distance decreases.

[0221] Since the volume of a liquid is a function of temperature, its volume increases with increasing temperature and decreases with decreasing temperature. The relationship between volume and temperature can also be used as an absolute relationship; a liquid has a specific volume at a specific temperature. (Volume changes related to the temperature of the container are ignored here.)

[0222] The network can learn the relationship between both phenomena (i.e., volume and temperature in the examples), thereby "transforming" missing data from the source data into target data. Here, the transformation function responds to specific situations where source data is unavailable (at least for certain variables), thereby establishing the target data by processing the available source data.

[0223] Optionally, the network can use context, such as state data indicating the container storing the liquid. Considering context can improve accuracy, for example, when establishing target parameters as volume parameters. Identifying missing identifiers

[0224] It is possible to have a situation where multiple sensors for the same parameter (for example, both volumes) provide source data with various source parameter identifiers (see 231-xx in Figure 3). However, in some cases, the source parameter identifier may simply be missing (for one of the sensors), or the data field of the identifier may be incomplete.

[0225] Nevertheless, it is possible for the network to identify the syntax (i.e., derive parameter identifiers). In that case, the network learns the correlations between the data from a first component having three time series (source data), similar to an autoencoder.

[0226] for example, Regarding the data from the first component, parameter A correlates with parameter B at a level of 0.7. Parameter B correlates with parameter C, but there is a 10-minute delay at a level of 0.2. Regarding the data from the second component, the parameter identified as "PP" (an arbitrary name here) correlates with the parameter identified as "GG" at level 0.6, and "GG" correlates with the parameter identified as "WW" but with a 12-minute delay at level 0.3.

[0227] This yields a first correlation pattern for a first component having sensors for parameters A, B, and C, and a second correlation pattern for a second component having sensors for parameters identified as "PP," "GG," and "WW."

[0228] Processing (by pattern matching, optimization, trial and error, or "brute force") Parameter A is associated with "PP," so identifiers A and "PP" refer to the same phenomenon. Parameter B is associated with "GG," so identifiers B and "GG" refer to the same phenomenon. Parameter C is associated with "WW," so that identifiers C and "WW" refer to the same phenomenon. It is possible to decide that.

[0229] This embodiment can be modified, for example, by ignoring the time dimension. If the correlation is not yet discernible, the correction can be assumed to be a permutation (e.g., using two alternatives), and the computer can ultimately determine which alternative is applied.

[0230] Autoencoders are networks, and they are well known in this field, as seen in the following papers: (i) Dor Bank, Noam Koenigstein, Raja Giryes: "Autoencoders" arXiv:2003.05991, and (ii) Michael Tschannen, Olivier Bachem, Mario Lucic: "Recent Advances in Autoencoder-Based Representation Learning" arXiv:1812.05069. Referring to this may be helpful. Other forms of identifiers

[0231] As shown in the example, there is one syntactic element for each parameter. Both "In" and "Vol" represent "vol," and the (network-based) conversion function is trained accordingly.

[0232] However, the number of syntactic elements may vary. Pressure parameters are a convenient example to illustrate this.

[0233] Pressure can be measured either in absolute terms (by force over an area) or in relative terms (also in relation to two media, but by applying force over an area).

[0234] Identifiers can be written, for example, by writing "P=1000hPa" (to indicate ambient air pressure, atmospheric pressure) or by writing "P=1...over" (for example, a hypothetical example of a container pressure that is twice the ambient air pressure).

[0235] Decades ago, the unit "at" was sometimes written as "atu" (in the case of Uberdruck = Overpressure).

[0236] In other words, the trained network derives missing data from the context of the source data.

[0237] In a further embodiment, the context should be "ambient air." The context does not need to be named with such a label, but can be derived from the {{X}} or {{Y}} data. In such a context, for example, a measurement from a pressure sensor with values ​​between "730" and "790" can be associated by learning that the measuring unit is "Torr" or "mm Hg." simulation

[0238] When simulations are involved, harmonizing the data can be advantageous. The virtual sensor (120-V in Figure 5) may include an (auxiliary) network that can provide training data by using the harmonization technique of Method 400. • Parameter values ​​(and states) may be the results of a simulation. • Simulations have different accuracy than actual measurements using real digital sensors. The network (trained or in use, Figure 2A) can distinguish between data from real and virtual sensors. • Distinction may be related to error correction. For example, a virtual sensor can optionally provide a confidence index. • Auxiliary networks or rules can detect suspicious values, and these networks (or rules) may differ between real and virtual sensors. Other networks

[0239] In the context of industrial machinery, (auxiliary) networks can be used for a variety of purposes, such as predicting or forecasting the behavior of the machine.

[0240] Since such networks need to be trained, harmonization provides an additional method for obtaining training data. Training data from two different machines (or components) with different source data conventions are merged into a harmonized pool containing the training data. Machine control

[0241] The method 400 described above for harmonizing parameter data can be used by a computer system that specifies control data (see computer 200 in Figure 1, left-pointing arrow to the machine) and transmits control signals to multiple components (110-xx) of one or more industrial machines (100). As described, method 400 uses neural networks 330,340 (Figures 7A, 7B) to determine transformation rules by processing source parameter datasets (221-xx). The resulting target parameter datasets (222-xx) can then be used by a computer to obtain control signals.

[0242] In other words, method 400 has a further step for obtaining a control signal.

[0243] The background art section of this application has already described a control loop extended to a control room. Method 400 can have its use in such a control loop. Various data conventions can still lead to improper operation, but the risk of such differences causing improper operation can be reduced. Reversal of the conversion direction

[0244] In the explanations so far, data has been distinguished into source data and target data by viewing the constituent elements (or systems) as sources and the constituent element twins (or system twins) as targets.

[0245] However, twins processing harmonized data (component twins and system twins at both granularities) ultimately do not distinguish between the control signals to the actuators.

[0246] Digital actuators can be considered a complement to digital sensors. They receive control data and act to change physical phenomena. Pressure sensors can open and close valves, and temperature actuators can be heaters (or even cooling devices).

[0247] The above constraints also apply. Various actuators may require different control signals, even if they share common parameters.

[0248] The conversion function 300 can be applied in the reverse direction, having control data (or control signals from a computer or twin) at its input and control data (to its components) at its output. The training and use of the network are similar. Application Use Cases

[0249] Since this disclosure focuses on data processing, the descriptions and drawings refer to industrial machinery in a simplified, illustrative manner. Machines and components are shown as containers, etc. The publication identified above, International Publication No. 2022 / 194871, also refers to use cases in blast furnaces. This disclosure may also be applied to such furnaces.

[0250] Figure 10 shows an embodiment of a general-purpose computer device that may be used in conjunction with the techniques described herein. Figure 10 shows an embodiment of a general-purpose computer device 900 and a general-purpose mobile computer device 950, which may be used in conjunction with the techniques described herein. The computing device 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. The computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, mobile phones, smartphones, driver assistance systems or vehicle board computers, and other similar computing devices. For example, the computing device 950 may be used as a front-end by a user (e.g., a blast furnace operator) to interact with the computing device 900. The components shown herein, their connections and relationships, and their functions are illustrative only and do not limit the forms of implementation of the invention described herein and / or claimed herein.

[0251] Computing device 900 includes a processor 902, a memory 904, a storage device 906, a high-speed interface 908 that connects to the memory 904 and a high-speed expansion port 910, and a low-speed interface 912 that connects to a low-speed bus 914 and the storage device 906. Each of the components 902, 904, 906, 908, 910, and 912 may be interconnected using various buses and may be mounted on a common motherboard or in other manners as required. The processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or the storage device 906 for displaying graphic information for a GUI on an external input / output device such as a display 916 coupled to the high-speed interface 908. In other implementations, multiple processors and / or multiple buses may be used, along with multiple memories and memory types as required. Also, multiple computing devices 900 may be connected, and each device may provide a portion of the required operations (e.g., as a server bank, a group of blade servers, or a multi-processor system).

[0252] The memory 904 stores information within the computing device 900. In one implementation, the memory 904 is one or more volatile memory units. In another implementation, the memory 904 is one or more non-volatile memory units. The memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.

[0253] Storage device 906 can provide large-capacity storage for computing device 900. In one implementation, storage device 906 can be or include a computer-readable medium such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid-state memory device, or an array of devices including a storage area network or other configured devices. A computer program product can be tangibly embodied in an information carrier. The computer program product can also include instructions that, when executed, perform one or more of the methods as described above. The information carrier is a computer or machine-readable medium such as memory 904, storage device 906, or memory on processor 902.

[0254] High-speed controller 908 manages the bandwidth-intensive operations of computing device 900, while low-speed controller 912 manages the low-bandwidth-intensive operations. Such function assignments are merely illustrative. In one implementation, high-speed controller 908 is coupled to a high-speed expansion port 910 that can receive memory 904, display 916 (e.g., through a graphics processor or accelerator), and various expansion cards (not shown). In this implementation, low-speed controller 912 is coupled to storage device 906 and low-speed expansion port 914. The low-speed expansion port, which can include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet), can be coupled to one or more input / output devices such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.

[0255] The computing device 900 may be implemented in several different forms, as shown in the figure. For example, it may be implemented as a standard server 920, or multiple times in a group of such servers. It may be implemented as part of a rack server system 924. It may also be implemented in a personal computer, such as a laptop computer 922. Alternatively, components from the computing device 900 may be combined with other components in a mobile device (not shown), such as device 950. Each such device may contain one or more computing devices 900, 950, and the entire system may consist of multiple computing devices 900, 950 communicating with each other.

[0256] Computing device 950 includes, among other components, a processor 952, memory 964, input / output devices such as a display 954, a communication interface 966, and a transceiver 968. Device 950 may also include storage devices, such as a microdrive or other devices, to provide additional storage. Each of components 950, 952, 964, 954, 966, and 968 are interconnected using various buses, and some of the components may be mounted on a common motherboard or in other ways as needed.

[0257] The processor 952 can execute instructions within the computing device 950, including instructions stored in memory 964. The processor may be implemented as a chipset of chips including multiple separate analog and digital processors. The processor may provide coordination of other components of the device 950, such as control of the user interface, applications run by the device 950, and wireless communication by the device 950.

[0258] The processor 952 can communicate with the user through a control interface 958 and a display interface 956 coupled to the display 954. The display 954 may be, for example, a TFT LCD (Thin-Film-Transistor Liquid Crystal Display), an OLED (Organic Light Emitting Diode) display, or other suitable display technology. The display interface 956 may include appropriate circuitry for driving the display 954 to present graphics and other information to the user. The control interface 958 may receive commands from the user and translate them for submission to the processor 952. In addition, an external interface 962 may be provided to communicate with the processor 952 to enable short-range communication between device 950 and other devices. The external interface 962 may provide, for example, wired communication in some implementations or wireless communication in other implementations, and multiple interfaces may be used.

[0259] Memory 964 stores information within the computing device 950. Memory 964 can be implemented as one or more of a computer-readable medium, a volatile memory unit, or a non-volatile memory unit. Extended memory 984 may also be provided to and connected to device 950 through an extended interface 982, which may include, for example, a SIMM (Single In Line Memory Module) card interface. Such extended memory 984 may provide additional storage space to device 950, or it may store applications or other information for device 950. Specifically, extended memory 984 may include instructions for executing or supplementing the processes described above, or it may also include secure information. Therefore, for example, extended memory 984 may function as a security module for device 950, or it may be programmed with instructions that enable the secure use of device 950. Furthermore, secure applications may be provided via a SIMM card, along with additional information, such as placing identification information on the SIMM card in a way that makes it impossible to hash.

[0260] The memory may include, for example, flash memory and / or NVRAM memory, as described later. In one implementation, the computer program product is tangibly embodied in an information carrier. The computer program product includes instructions that, when executed, perform one or more of the methods described above. The information carrier is a computer or machine-readable medium, such as memory 964, extended memory 984, or memory on processor 952, which can be received, for example, via transceiver 968 or external interface 962.

[0261] Device 950 may communicate wirelessly through a communication interface 966, which may optionally include digital signal processing circuitry. The communication interface 966 may provide communication under various modes or protocols, including, among others, GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS. Such communication may be conducted, for example, through a radio frequency transceiver 968. Furthermore, short-range communication may be conducted, such as using Bluetooth, WiFi, or other such transceivers (not shown). Additionally, a GPS (Global Positioning System) receiver module 980 may provide device 950 with additional navigation and location-related radio data that can be appropriately used by applications running on device 950.

[0262] Furthermore, device 950 may communicate via voice using an audio codec 960 capable of receiving voice information from the user and converting it into usable digital information. The audio codec 960 may also generate audible sound for the user, for example, through a speaker in the handset of device 950. Such sound may include sounds from voice calls, recorded sounds (e.g., voice messages, music files, etc.), or sounds generated by applications running on device 950.

[0263] The computing device 950 may be implemented in several different forms, as shown in the figure. For example, it may be implemented as a mobile phone 980. Alternatively, it may be implemented as part of a smartphone 982, a personal digital assistant, or other similar mobile device.

[0264] The various implementations of the systems and technologies described herein can be realized in digital electronic circuits, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs executable and / or interpretable on a programmable system which includes at least one programmable processor, which may be specialized or general-purpose, coupled to receive data and instructions from a memory system, at least one input device, and at least one output device, and to transmit data and instructions to them.

[0265] These computer programs (known as programs, software, software applications, or code) include machine instructions for programmable processors and can be implemented in high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms “machine-readable medium” and “computer-readable medium” mean any computer program product, apparatus and / or device (e.g., magnetic disks, optical disks, memory, programmable logic devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, and include machine-readable medium that receives machine instructions as machine-readable signals. The term “machine-readable signal” means any signal used to provide machine instructions and / or data to a programmable processor.

[0266] To provide user interaction, the systems and technologies described herein can be implemented on a computer having a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) and a keyboard and pointing device (e.g., a mouse or trackball) on which the user can provide input to the computer. Other types of devices can also be used to bring about user interaction; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback), and the input from the user can be received in any form, including acoustic, voice, or tactile input.

[0267] The systems and technologies described herein can be implemented in a computing device that includes backend components (e.g., data servers), middleware components (e.g., application servers), or frontend components (e.g., client computers having a graphical user interface or web browser on which a user can interact with the implementation of the systems and technologies described herein), or any combination of such backend, middleware, and frontend components. The components of the system can be interconnected by digital data communication (e.g., communication networks) of any form or medium. Embodiments of communication networks include local area networks ("LANs"), wide area networks ("WANs"), and the Internet.

[0268] Computing devices can include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The client-server relationship arises from computer programs running on each computer that have a client-server relationship with each other.

[0269] Some embodiments have been described. Nevertheless, it will be understood that various modifications can be made without departing from the spirit and scope of the present invention.

[0270] Furthermore, the logic flows shown in the figures do not require the specific order, or sequential order, shown to achieve the desired result. Further, other steps may be provided from the described flow, steps may be eliminated, other components may be added to the described system, or other components may be removed from the described system. Accordingly, other embodiments are within the scope of the following claims.

Explanation of Reference Numerals

[0271] 100 Industrial machine 101 Component group 102 Component group 110 Component 120 Digital sensor 200 User interface 210 Digital equivalent 221 Source data 222 Target data 261 Context data 290 Operator 300 Conversion function (neural network) 315 Context network 330 Subnetwork 331 Input 332 Output 340 Subnetwork 341 Input 342 Output Method having 4xx steps 501 Context 502 Context 9xx General-purpose computer

Prior Art Documents

Patent Documents

[0272] [Patent Document 1] U.S. Patent Application Publication No. 2018 / 0356800

Claims

1. A computer implementation method (400) for harmonizing parameter data, Step (410) of receiving a source parameter dataset (221-xx) representing technical parameters from a component (110-xx) of an industrial machine (100), wherein the technical parameters belong to the component (110-xx), and the source parameter dataset (221-xx) has a source parameter identifier (231-xx) and a source parameter value (241-xx), Step (420) of matching the source parameter identifier (231-xx) with the target parameter identifier (232-xx) using a first subnetwork (330) of a pre-trained neural network (300), wherein the first pre-trained subnetwork (330) is trained with a plurality of historical parameter identifier pairs (231-hh, 232-hh) having a historical source parameter identifier (231-hh) at the input (331) of the first pre-trained subnetwork (330) and a historical target parameter identifier (232-hh) at the output (332) of the first pre-trained subnetwork (330), Step (430) of matching the source parameter value (241-xx) with the target parameter value (242-xx) using the second subnetwork (340) of the pre-trained neural network (300), wherein the second subnetwork (340) of the pre-trained neural network (300) - Selected according to the target parameter identifier (232-xx), Step (430) is trained with a plurality of historical parameter value pairs (241-hh, 242-hh) having a historical source value (241-hh) at the input (341) of the second pre-trained subnetwork (340) and a historical target value (242-hh) at the output (342) of the second pre-trained subnetwork (340), Step (440) of transferring both the target parameter identifier (232-xx) and the target parameter value (242-xx) to a user interface (200) that shows a user interface element (210-xx) corresponding to a component (110-xx) of the industrial machine (100) and for visualizing the target parameter value (242-xx), Computer implementation method (400), including.

2. The method according to claim 1 (400), wherein the first subnetwork (330) is trained with a set of parameter identifier pairs (231-h, 232-h), and is also trained with historical context data (261-hh) in the input (331) of the first subnetwork (330).

3. The method according to claim 1 or 2 (400), wherein the second subnetwork (340) is trained with a set of parameter value pairs (241-h, 242-h), and is also trained with historical context data (271-hh) in the input (341) of the second subnetwork (340).

4. The method (400) according to any one of claims 1 to 3, wherein the receiving step (410) further includes the step of identifying a real-time context (501, 502) in which the component (110-xx) is currently operating, the real-time context being a set of data points related to a data point (Xmn) in the source parameter dataset (221-xx) of the technical parameters.

5. The method according to claim 4 (400), wherein the receiving step (410) includes the step of identifying the real-time context, the step of accessing state data (S0, S1).

6. The method according to any one of claims 1 to 5 (400), wherein in the transfer step (440), the user interface (200) visualizes the target parameter value (242-xx) by a symbol (252-xx).

7. The method (400) according to any one of claims 1 to 6, wherein, after the step (410) of receiving the source parameter dataset (221-xx), the first and second subnetworks (430, 440) determine whether the source parameter dataset (221-xx) conforms to the specifications of the target parameter dataset (222-xx), and if the determination is positive, the method (400) continues with the transfer step (450), thereby skipping the matching step (430, 440).

8. After the step (410) of receiving the source parameter dataset (221-xx), the computer, The context network (315) is used to perform the step (415) of identifying the context type. The method according to any one of claims 1 to 7 (400), wherein the step (420) of matching the source parameter identifier with the target parameter identifier is performed by selecting a specific first subnetwork (330) according to the context type.

9. The method (400) according to any one of claims 1 to 8, wherein the transfer step (400) is followed by a step of obtaining control signals to control the operation of the components (110-xx) of the industrial machine (100).

10. A step of using the method (400) for harmonizing parameter data according to any one of claims 1 to 9, by a computer system (200) that specifies control data and transmits control signals to one or more components (110-xx) of one or more industrial machines (100).

11. A computer system (200) adapted to perform a method (400) for harmonizing parameter data from components (110-xx) of an industrial machine (100), wherein the method (400) is the computer implementation method (400) according to any one of claims 1 to 9.

12. A computer implementation method for training a neural network (300) having a first subnetwork (330) and a second subnetwork (340), The neural network (300) is trained for subsequent use in a computer implementation method (400) for harmonizing parameter data from multiple source conventions to a single target convention, using a data flow from input (331, 341) to output (332, 342) through the first and second subnetworks (330, 340). The parameter data relates to a component (110-xx) of an industrial machine (100), and the parameter dataset (221-xx) represents the technical parameters belonging to the component (110-xx), and the parameter dataset (221-xx) has a parameter identifier (231-xx) and a parameter value (241-xx). The aforementioned source specification includes a source parameter identifier (231-xx) and a source parameter value (241-xx), and the aforementioned single target specification includes a target parameter identifier (232-xx) and a target parameter value (242-xx), The method described above is The steps include training the first subnetwork (330) with a plurality of historical parameter identifier pairs (231-hh, 232-hh), each having a historical source parameter identifier (231-hh) at the input (331) of the first subnetwork (330) and a historical target parameter identifier (232-hh) at the output (332) of the first subnetwork (330), The steps include training the second subnetwork (340) with a plurality of historical parameter value pairs (241-hh, 242-hh) having a historical source value (241-hh) at the input (341) of the second subnetwork (340) and a historical target value (242-hh) at the output (342) of the second subnetwork (340), Computer implementation methods, including those mentioned above.

13. The computer implementation method according to claim 12, further comprising the steps of the computer implementation method (400) according to any one of claims 1 to 9.

14. A computer program product that, when loaded into the memory of a computer and executed by at least one processor of the computer, causes the computer to perform the steps of the method according to any one of claims 1 to 9, or the steps of the method according to claim 12.

Images