Industrial plant machine learning system

Abstract

This invention relates to an industrial shop floor machine learning system, comprising: a machine learning model providing machine learning data; an industrial shop floor providing shop floor data; and an abstraction layer connecting the machine learning model and the industrial shop floor, wherein the abstraction layer is configured to provide standardized communication between the machine learning model and the industrial shop floor using a machine learning markup language.

Description

Technical Field

[0001] This disclosure relates to industrial shop floor machine learning systems, methods for industrial shop floor machine learning communication, applications of industrial shop floor machine learning systems in machine learning development, and computer programs. Background Technology

[0002] Connecting machine learning to shop floor data is a challenging task. Data is distributed across many different systems, and it's impossible to provide all the data specifically needed to train or rate machine learning models. Even when some relevant data is available, training a well-performing machine learning model is often insufficient. The application of transfer learning concepts, successfully used on image data, cannot be easily transferred to industrial data such as time-series / signal data, as well as alarm and event data. The high dimensionality of industrial data, often consisting of hundreds or thousands of data points, makes machine learning subject to the dimensionality curve, and algorithms are likely to overfit the provided training data.

[0003] Furthermore, machine learning (ML) models used in process control and automation require access to both historical and current process and shop floor data. Connecting a DCS to an ML model demands significant effort in selecting and configuring the necessary inputs. This configuration is highly dependent on the equipment topology, and even minor changes to the equipment necessitate substantial redesign and model relearning.

[0004] Especially in the manufacturing industry, each workshop has different automation systems, different types of sensors, and different components, even if the type of workshop and the products produced may be the same. Therefore, there is no guarantee that machine learning models can be generalized from one device to another.

[0005] Some ML models require training on labels that are traditionally very expensive. Larger companies have many people working on data labeling tasks.

[0006] When performing continuous operations, the ML model begins to provide predictions. These are typically only understandable to the person who trained the model, if not well documented. If the model changes, potential additional computations based on the results fail.

[0007] Currently, ML tasks require a significant amount of data engineering. In data science projects that include ML, up to 80% of the time is spent searching for data and developing data pipelines. For the next project on the same data source (e.g., a processing workshop), efforts are made to maintain the same approach.

[0008] Data science projects dedicate a significant amount of time to data exploration and understanding. In practice, data is rarely documented, which is why data scientists spend a considerable amount of time on this task.

[0009] Implementing an ML solution involves several design options. Predictive computations can be performed in multiple locations, in the cloud, or on-site. Finally, the application consuming the results needs to know where to look. This information is often hard-coded into the application, making changes and modifications difficult.

[0010] Therefore, an improved machine learning system for industrial workshops is needed. Summary of the Invention

[0011] According to one aspect, an industrial shop floor machine learning system includes: a machine learning model that provides machine learning data; an industrial shop floor that provides shop floor data; and an abstraction layer that connects the machine learning model and the industrial shop floor, wherein the abstraction layer uses a machine learning markup language to provide standardized communication between the machine learning model and the industrial shop floor.

[0012] The term "markup language" used here is configured to organize components in an industrial workshop, and in particular to identify the correct technical names of components in order to extract the corresponding data.

[0013] In process automation systems, signals concerning the state and performance of control loops, or measuring instruments, have technical names that depend on the automation system, the engineering and libraries used, and shop-side specific naming conventions. Markup languages, for example, organize these technical system names based on shop-side topology (process to cell, container, and final control loop) and provide additional information, such as which variable is being controlled. This enables simple, automated queries to identify the desired technical signals independently of the specific implementation of the automation system. Alternatives to markup languages ​​are simple mapping tables or key-value-based documents, such as JSON.

[0014] Machine learning markup languages ​​allow for changes to components or applications on the industrial floor (also known as the industrial floor floor) without having to change anything in the machine learning model, especially on machine learning computational pipelines.

[0015] In other words, the abstraction layer is configured to manage data transmission between machine learning units and the industrial workshop.

[0016] Preferably, the workshop data includes structured data, particularly time series, alarms and events, as well as unstructured data, particularly reports.

[0017] Preferably, the shop floor data is stored locally in the historical machine of the industrial shop floor. Since the historical machine's shop floor data is required to power the machine learning model via the DCS, the connection between the historical machine, the DCS, and the machine learning model is crucial for the machine learning model. The abstraction layer allows changes to the machine learning model without needing to establish new connections to the historical machine or the DCS, as it provides standardized communication.

[0018] Preferably, when a component of the system is changed, the new data source simply reconnects to the abstraction layer. Therefore, data consumers using the abstraction layer, such as users of the system or another component, will not notice any changes.

[0019] Preferably, users can connect to the abstraction layer using the system's input interface. Users can send requests to the abstraction layer. Requests trigger services in the abstraction layer, such as search or retrieval, and the abstraction layer provides a structured response to the user using a machine learning markup language.

[0020] In other words, the abstraction layer enables communication between machine learning models (especially machine learning applications) and industrial workshops (especially distributed control systems, DCS). The abstraction layer provides abstraction and translation between industrial operational technology (OT) and industrial information technology (IT) and machine learning.

[0021] Depending on the direction of data flow, machine learning models and industrial workshops include data consumers and / or data sources. An abstraction layer manages the data flow between data sources and data consumers (also known as data sinks).

[0022] Preferably, the abstraction layer provides an abstraction of the shop floor data. In this so-called bottom-up view, the abstraction layer is configured to provide abstract shop floor data to the machine learning model. More preferably, the abstraction layer provides an abstraction of the machine learning data, particularly machine learning predictions provided by the machine learning model. In this so-called top-down view, the abstraction layer is configured to provide abstract machine learning data to the industrial shop floor.

[0023] This standardized communication reduces configuration work and provides a simple way to reconfigure and relearn after changes in the industrial shop floor. Furthermore, it provides a mechanism for automatically generating finite state machines from DCS programs, which can be used to provide labels with state and phase information to supervised machine learning models.

[0024] This standardization reduces configuration work and provides a simple way to reconfigure and learn after changes occur on the industrial floor.

[0025] Thanks to the abstraction layer, all components of the industrial shop floor and machine learning model can be interchanged with similar components without requiring necessary modifications to other components of the industrial shop floor and machine learning model.

[0026] The abstraction layer also allows the execution of machine learning algorithms for machine learning models to be managed at various locations based on optimization criteria.

[0027] Therefore, the abstraction layer allows for the provision of industrial shop floor machine learning systems with improved development, implementation, and operational speed.

[0028] In a preferred embodiment, the abstraction layer is configured to enrich the received shop floor data with context data, wherein the context data includes shop floor status.

[0029] The term "workshop state," as used herein, includes the state of process variables and / or the state of components, such as a steady state or a startup state.

[0030] Therefore, the abstraction layer allows for the provision of industrial shop floor machine learning systems with improved development, implementation, and operational speed.

[0031] In a preferred embodiment, the industrial shop includes a distributed control system (DCS), wherein an abstraction layer is configured to determine contextual data by analyzing the DCS's code to automatically generate a finite state machine for automatically generating the shop's state.

[0032] The term "analyzing DCS code" as used here involves migrating the DCS code into a so-called expression tree, where the entire code is represented in the form of method -> branch -> expression -> operator -> binary operation. Context data, particularly the shop floor state, is the currently active node in the expression tree or a subtree within the expression tree. Subtrees in the expression tree correspond to subroutines such as steady-state control, automatic start-up or shutdown, and safety logic.

[0033] Therefore, the abstraction layer allows for the provision of industrial shop floor machine learning systems with improved development, implementation, and operational speed.

[0034] In a preferred embodiment, the abstraction layer is configured to analyze the DCS code using code expression tree analysis. The expression tree represents the automation code in a tree structure, where each node in the tree is an expression, subroutine, or binary operation, such as a>b. During execution, the program resides in the expression and a node in the tree, and the node or subtree can be mapped to the state of the DCS or workshop. This state is characterized by the ID of the currently active node in the expression tree.

[0035] In a preferred embodiment, the machine learning model is configured to use the workshop state as a label for training the machine learning model.

[0036] Therefore, the generation of labeled data for machine learning models becomes cheaper, and thus the quality of machine learning models is improved.

[0037] In a preferred embodiment, the abstraction layer is configured to abstract machine learning data and workshop data.

[0038] The term "abstraction" as used here refers to a complex set of data, which is then represented by an abstract version of that data. For example, all signals from a temperature sensor are abstracted from a single dataset. The abstraction layer abstracts away from the concrete implementation of automation, such as naming conventions and decisions about which control operates on which hardware and with which I / O. This allows machine learning systems to work with generic queries such as "all tank temperatures" or "all signals in all control loops in unit X".

[0039] In a preferred embodiment, the connection between the abstraction layer and the industrial workshop uses platform-independent communication technology.

[0040] In a preferred embodiment, platform-independent communication technologies include OPC Unified Architecture (OPC UA) or Message Queuing Telemetry Transport (MQTT).

[0041] In a preferred embodiment, the abstract shop floor data includes standardizing and abstracting supplier-specific parts and industrial shop floor-specific parts using machine learning markup languages.

[0042] In a preferred embodiment, the abstraction layer is located in an edge device near the industrial workshop.

[0043] Alternatively, the abstraction layer could reside in a cloud environment.

[0044] In a preferred embodiment, the abstraction layer includes an application programming interface (API) that provides standardized access to shop floor data.

[0045] Preferably, the API operates in a manner independent of the vendor and shop floor topology.

[0046] In a preferred embodiment, the application programming interface includes an access control unit that provides users with access control over industrial shop floor data and machine learning data.

[0047] Preferably, the access control unit ensures secure and controlled access to shop floor data and machine learning data.

[0048] Preferably, the access control unit performs restricted data exchange only on data necessary to meet confidentiality requirements.

[0049] According to one aspect of the present invention, a method for machine learning communication in an industrial workshop includes the following steps: In a first step, machine learning data is provided through a machine learning model. In a second step, workshop data is provided by the industrial workshop. In a third step, standardized communication between the machine learning model and the industrial workshop is provided using a machine learning markup language through an abstraction layer connecting the machine learning model and the industrial workshop.

[0050] According to one aspect of the invention, the use of the industrial shop floor machine learning system as described herein in machine learning development is provided.

[0051] According to one aspect of the invention, a computer program is provided that includes instructions which, when executed by a computer, cause the computer to perform the steps of the method used herein.

[0052] The subject matter of the invention will be explained in more detail below with reference to the preferred exemplary embodiments shown in the accompanying drawings. Attached Figure Description

[0053] Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings:

[0054] Figure 1 A schematic diagram of an industrial workshop machine learning system is shown; and

[0055] Figure 2 A schematic diagram of a method for industrial machine learning communication is shown.

[0056] The reference numerals used in the accompanying drawings and their meanings are listed in summary form in the reference numeral list. In principle, the same assembled parts have the same reference numerals in the drawings.

[0057] Preferably, the functional modules and / or configuration mechanisms are implemented as programmable software modules or processes; however, those skilled in the art will understand that the functional modules and / or configuration mechanisms can be fully implemented or partially assembled in hardware. Detailed Implementation

[0058] Figure 1 An industrial shop floor machine learning system 10 is illustrated, comprising a machine learning model 20, an industrial shop floor 30, and an abstraction layer 40 connecting the machine learning model 20 to the industrial shop floor 30. The machine learning model 20 includes a user unit 21, a training unit 22, a scoring unit 23, and a visualization unit 34. The industrial shop floor 30 includes a distributed control system (DCS) 31, a historical data unit 32, an enterprise resource planning (ERP) unit 33, a computerized maintenance management system (CMMS) 34, a content management system (CMS) 35, a laboratory information management system (LIMS) 36, and a process flow unit 37, including, for example, P&ID and IO lists. The abstraction layer 40 includes an access control unit 41, which manages data access between the machine learning model 20 and the industrial shop floor 30. Additionally, the abstraction layer 40 includes a directory service that manages the network resources of the abstraction layer 40.

[0059] Therefore, an abstraction layer 40 is defined between the industrial shop floor 30 (specifically, the DCS 31) and the machine learning model 20 (specifically, machine learning-related applications). A machine learning meta-language is used to standardize communication between the machine learning model 20 and the industrial shop floor 30. Specifically, the machine learning meta-language is used to standardize communication between the DCS 31, the historical machine 32, and other data sources. The abstraction layer 40 includes an application programming interface (API) to provide data on requests with strict access control, capable of distinguishing different receivers. Thus, a mechanism can be provided to automatically generate a finite state machine describing the industrial shop floor 30 and to provide labels for supervised machine learning models.

[0060] Therefore, machine learning can be provided to customers at a low cost. Furthermore, machine learning projects run by distributors or customers are accelerated. Applications in industrial shop floor 30 can easily consume machine learning results without needing to know how or where they were generated. Changes to the machine learning model do not require “rewiring” industrial shop floor applications, especially shop floor applications. Data access for machine learning can be securely and controlled through abstraction layer 30. The generation of labeled data is inexpensive, thus improving the quality of machine learning model 20. Shop floor data is provided in a structured manner using a machine learning markup language. The machine learning markup language allows for changes to the shop floor without requiring changes to machine learning model 20 or anything in the machine learning computation pipeline. Therefore, a mechanism can be provided to manage the execution of machine learning algorithms at various locations based on optimization criteria.

[0061] In existing technologies, when a machine learning model is connected to a DCS, the model's input is directly connected to some signals and process variables available in the control system and / or historical data set. This is a lengthy process, requiring domain expert skills in selecting appropriate signals, and is dependent on the shop floor topology, signal naming scheme, and the suppliers of the control system and historical data set.

[0062] However, the abstraction layer 40 between the workshop 30 and the machine learning model 20, particularly for machine learning applications, accelerates the development, implementation, and operation of machine learning. A machine learning markup language is used to standardize communication between the machine learning model 20 and the workshop 30. The abstraction layer 40 can reside in the cloud or on an edge device, managing the data flow between data sources and data destinations.

[0063] This standardization reduces configuration work and provides a simple way to reconfigure and relearn after shop floor changes. Furthermore, it provides a mechanism for automatically generating finite state machines from DCS programs, which can be used to provide labels with state and phase information to supervised machine learning models.

[0064] In its simplest version, abstraction layer 40 provides an abstraction of shop floor data.

[0065] In a bottom-up view, industrial shop floor 30 generates data, particularly structured data such as time series, alarms, and events, as well as unstructured data such as reports. This is stored locally on historical machine 32 or other systems. A subset can be sent from there to the cloud, for example, via edge devices. CMMS provides a local view of the data. Enterprise dashboard applications provide a global view of the data.

[0066] The design of Abstraction Layer 40 specifically addresses the needs of machine learning. It provides secure and structured access to industrial shop floor data. Users only see what they are authorized to see. Structure is imposed using a machine learning markup language. There, the data is enriched with the metadata and tags necessary for machine learning. To connect Abstraction Layer 40 to Shop Floor 30, technologies such as OPCUA and MQTT are used, which are capable of constructing shop floor data.

[0067] In addition to structured shop floor data, abstraction layer 40 also provides information about shop floor states and labels. Therefore, mechanisms for analyzing DCS code, such as code expression tree analysis, are used to automatically generate finite state machines from DCS programs. Abstraction layer 40 can provide machine learning engineers with automatically generated states as labels for training supervised machine learning models.

[0068] In the event of changes to the industrial shop floor machine learning system 10, such as changes to components, the new data source is simply connected to the abstraction layer 40. Therefore, those consuming data using the abstraction layer 40 will not notice any changes.

[0069] In this top-down view, users can connect to the abstraction layer 40, send requests to it, and trigger services such as searching, retrieving, and obtaining structured machine learning markup language responses. These can be used directly in machine learning design environments such as Python, R, and Matlab.

[0070] The machine learning engineer sends a search request to abstraction layer 40. The engineer doesn't need to know all the details, but can query abstraction layer 40 for all available data that meets specified criteria.

[0071] Machine learning engineers can send a fetch request to abstraction layer 40 to retrieve specified data in machine learning markup language.

[0072] During the data exploration phase, at the start of a machine learning project, engineers need to understand the available data. Abstraction layer 40 provides a search-like service that allows for the automated searching of available data. This data is provided in a structured manner using a machine learning markup language, making it directly usable by data exploration tools.

[0073] During the training phase, machine learning engineers run numerous experiments to build predictive models. If a supervised learning model is to be developed, labeled data can be automatically used by the machine learning development environment.

[0074] During the testing and validation phase, the developed model can be automatically compared with the testing and validation data.

[0075] During the deployment phase, the resulting machine learning model 20 can be run and "announced" to the abstraction layer 40 via MLML. Where the machine learning model 20 is deployed is not important. The results can be consumed through the abstraction layer 40.

[0076] During the operational phase, the workshop data required by machine learning model 20 is provided by abstraction layer 40. The results of the predictive model of machine learning model 20 can be consumed again via abstraction layer 40. As long as the same data is consumed and the same type of results are produced, any changes made to machine learning model 20 can be easily implemented.

[0077] If the same type of data can be provided, and if there is also an abstraction layer 40, then the migration of the machine learning model 20 to other industrial workshops is simplified.

[0078] In addition, the abstraction layer also handles data exchange between applications and analytics algorithms.

[0079] Instead of directly connecting to shop floor data, abstraction layer 40 is used to retrieve the data. Therefore, a subscription service can be used, which always provides new data as the shop floor data changes. Any shop floor data generated within these applications can again be provided via abstraction layer 40. This includes any machine learning models within the application itself.

[0080] The analysis includes machine learning algorithms and other computational functions. Instead of directly obtaining data from the source, the abstraction layer 40 can be used to provide the necessary data and deliver the computational results.

[0081] In addition, the abstraction layer can be used with existing software and BI solutions.

[0082] Existing software applications are often designed in a way that does not allow for automatic data extraction, and the data must be provided in some structure. These can be coupled to an abstraction layer 40 via connectors. The task of these connectors is to transform the data to suit the application. Connectors can be based on existing standards.

[0083] Data available to existing applications is typically not intended to be shared; it is usually only created as needed in machine-readable format as exported files. Connectors can read these and make them available to abstraction layer 40.

[0084] Decision-makers use BI solutions such as Power BI, Qlik, or Tableau to analyze the current state, identify the root causes of problems, and obtain predictions about the impact on shop floor performance. These can interact with the abstraction layer 40 to obtain lifetime data and filter it as needed.

[0085] Therefore, it greatly simplifies data engineering.

[0086] Abstraction layer 40 can provide user-specified specific queries to populate predefined machine learning templates, which semantically define the data requirements of machine learning algorithms, such as specifying certain features as input to the system, such as “reactor temperature, head pressure, tail pressure” or “drive-side vibration measurement on pumps”.

[0087] Abstraction layer 40 identifies data points either by using a mapping of statically defined data points in the IT / OT system, or by employing natural language processing techniques, particularly named entity recognition, and by using graph algorithms for shop floor topology analysis to identify the correct data points in the data source. This analysis may include data descriptions in IO lists, configuration data in the DCS, data point names, identifiers, names, etc. As a post-processing step, abstraction layer 40 can perform a "sanity check" on the extracted data. For example, if the recorded data actually behaves like a temperature or vibration signal, or shows a cross-correlation expected based on the shop floor or asset topology—for example, if vibration signals from vibration sensors on pumps, blowers, or gearless mill drives match electrical signals, in the simplest case, as on / off information.

[0088] Figure 2 A method for machine learning communication in an industrial shop floor is described, comprising the following steps. In a first step S10, machine learning data is provided by a machine learning model. In a second step S20, shop floor data is provided by an industrial shop floor 30. In a third step S30, standardized communication is provided between the machine learning model and the industrial shop floor 30 using a machine learning markup language through an abstraction layer 40 connecting the machine learning model and the industrial shop floor 30.

[0089] Reference Symbol List

[0090] 10 Industrial Shop Machine Learning Systems

[0091] 20 machine learning models

[0092] 21 User Units

[0093] 22 training units

[0094] 23 scoring units

[0095] 24 visualization units

[0096] 30 industrial workshops

[0097] 31 Distributed Control System

[0098] 32 historical machines

[0099] 33 Enterprise Resource Planning

[0100] 34 Computer Maintenance Management System

[0101] 35 Content Management System

[0102] 36 Laboratory Information Management System

[0103] 37 process units

[0104] 40 Abstraction layers

[0105] 41 Access Control Unit

[0106] 42 Directory Services

[0107] S10 First Step

[0108] S20 Second Step

[0109] S30 Third Step

Claims

1. An industrial workshop machine learning system (10), comprising: Machine learning model (20) provides machine learning data; An industrial workshop (30) provides workshop data, wherein the workshop data is stored in a historical machine (32) or other system, and the workshop data includes structured data and / or unstructured data, wherein the structured data includes time series, alarms and events, and the unstructured data includes reports; and An abstraction layer (40) connects the machine learning model (20) and the industrial workshop (30); wherein The abstraction layer (40) is configured to use a machine learning markup language to provide standardized communication between the machine learning model and the industrial workshop (30); The abstraction layer (40) is configured to enrich the received shop floor data with context data, wherein the context data includes shop floor states, and the shop floor states include the states of process variables and / or the states of components; and The machine learning model (20) is configured to use the workshop state as a label for training the machine learning model (20).

2. The system according to claim 1, The industrial workshop mentioned above includes a distributed control system (DCS) (31). The abstraction layer (40) is configured to determine the context data by analyzing the code of the DCS (31) to automatically generate a finite state machine for automatically generating the workshop state.

3. The system according to claim 2, The abstraction layer (40) is configured to use code expression tree analysis for analyzing the code of the DCS (31).

4. The system according to any one of claims 1-3, The abstraction layer (40) is configured to abstract the machine learning data and the workshop data.

5. The system according to any one of claims 1-3, The connection between the abstraction layer (40) and the industrial workshop (30) uses platform-independent communication technology.

6. The system according to claim 5, The platform-independent communication technologies mentioned include OPC Unified Architecture (OPCUA) or Message Queuing Telemetry Transport (MQTT).

7. The system according to claim 4, The abstract workshop data includes: The machine learning markup language is used to standardize and abstract supplier-specific parts and industrial workshop-specific parts.

8. The system according to any one of claims 1-3 and 6-7, wherein The abstraction layer (40) is located in an edge device near the industrial workshop (30).

9. The system according to any one of claims 1-3 and 6-7, wherein The abstraction layer (40) includes: An application programming interface (API) provides standardized access to the workshop data.

10. The system according to claim 9, wherein The API includes an access control unit (41) that provides users with access control to the industrial workshop data and the machine learning data.

11. A method for machine learning communication in an industrial shop floor, comprising the following steps: (S10) Machine learning data is provided through machine learning models; (S20) Workshop data is provided via an industrial workshop (30), wherein the workshop data is stored in a historical machine (32) or other system, and the workshop data includes structured data and / or unstructured data, wherein the structured data includes time series, alarms and events, and the unstructured data includes reports; and By using an abstraction layer (40) that connects the machine learning model and the industrial workshop (30), and by using a machine learning markup language, standardized communication is provided (S30) between the machine learning model and the industrial workshop (30). The received shop floor data is enriched with contextual data by the abstraction layer (40), wherein the contextual data includes shop floor status, and the shop floor status includes the status of process variables and / or the status of components; as well as The machine learning model (20) uses the workshop state as a label to train the machine learning model (20).

12. The method according to claim 11, The industrial workshop mentioned above includes a distributed control system (DCS) (31). The abstraction layer (40) is configured to determine the context data by analyzing the code of the DCS (31) to automatically generate a finite state machine for automatically generating the workshop state.

13. An application of the industrial shop floor machine learning system according to any one of claims 1-10 in machine learning development.

14. A computer program product comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method according to claim 11 or 12.

Images