Industrial automation system and method

Abstract

An industrial automation system (100, 100A) for implementing at least one industrial process comprises a plurality of process components (200; the present invention relates to a system (100, 200, 300, 301, 302, 303, 304), each component being capable of being classified into a queue corresponding to a queue allocation criterion. At least some of the process components (301, 302, 303, 304) are configured to perform an ML process involving machine learning ML model parameters. At least one of the process components (200) is configured to host at least a portion of the at least one ML model of each queue, and is further configured to pass ML model parameters between the plurality of process components. The system is configured to allocate one or more of the process components to one of the queues according to a queue allocation criterion; attributing ML model parameters of process components in a selected queue in the queues to ML models belonging to the selected queue; determining a proximity value of each pair of queues; assigning a pair of queues to a respective group of neighboring queues if the proximity value satisfies a predetermined proximity criterion; and sharing ML model related data among the process components belonging to the same adjacent queue group. A computer-implemented method performed in an industrial automation system is also provided.

Description

Technical Field

[0001] The present invention relates to an industrial automation system for implementing at least one industrial process. The present invention also relates to a computer-implemented method performed in the industrial automation system. Background Art

[0002] In industrial automation, it is often necessary to detect abnormal phenomena, predict maintenance needs of the factory to which the industrial automation is applied, etc. Some industrial automation solutions use machine learning (ML) technology, such as ML-based soft sensors. ML relies on robust training of the corresponding ML model. In order to train, sufficient data capturing different scenarios in the factory and sufficient data for each scenario are required.

[0003] When training a machine learning model, it is necessary to be able to obtain data such as local data. An industrial site may include multiple factories, each of which performs processes. Factories may also be spread across multiple industrial sites. In a traditional example, the ML model is trained at the factory or site (i.e., within the area of ​​the factory or site where data is readily available and accessible). In this case, only factory local data or site local data is used to train the ML model, which may not be sufficient to develop a robust machine learning model.

[0004] As a non-limiting example, in process automation at a cement plant, soft sensors based on machine learning (ML) are used, for example, to detect abnormal phenomena, predict attributes related to maintenance work, etc. It is expected that any ML model used in these exemplary ML-based techniques is a robust ML model. For example, robustness can be achieved by using captured data of different scenarios and / or using a sufficient amount of those data to train the ML model.

[0005] A typical way to train an ML model is to train the ML model at the site (site, location) of the industrial process. For example, the ML model is trained on-site on the edge / server IT infrastructure using local data. In this case, learning the ML model is limited to local data. Local data may not be sufficient to make the ML model robust. Data from multiple sites of the same or different industrial processes can also be shared, for example, in order to achieve and / or improve the robustness of the ML model. For example, data from multiple factories can be shared with a server (such as a central server that can be cloud-hosted). In this case, sensitive data may be involved, and it is not expected and / or allowed to share or leak this data across site boundaries. Moreover, sharing data for training ML models typically involves a large amount of data to be transferred, which increases communication overhead, storage costs, and power consumption. The quality of the ML model needs to be enhanced. Summary of the invention

[0006] According to one aspect, an industrial automation system for implementing at least one industrial process is provided. The industrial automation system includes a plurality of process components. Each process component can be classified as a queue corresponding to a queue allocation criterion. At least some of the process components are configured to perform ML processes involving ML model parameters. At least one of the process components is configured as at least a portion of at least one ML model hosting each queue. At least one process component is also configured to transfer ML model parameters between a plurality of process components. The system is configured to assign one or more of the process components to a queue in the queue according to the queue allocation criterion; attribute the ML model parameters of the process components in the selected queue in the queue to the ML model belonging to the selected queue; determine the proximity value of each pair of queues; if the proximity value meets the predetermined proximity criterion, assign a pair of queues to corresponding adjacent queue groups; and share ML model-related data between process components belonging to the same adjacent queue group.

[0007] According to another aspect, a computer-implemented method for implementing at least one industrial process performed in an industrial automation system is provided. The industrial automation system includes a plurality of process components. Each process component can be classified into a queue corresponding to a queue allocation criterion. In the method, an ML process involving model parameters is performed on one of the process components. At least a portion of at least one ML model for each queue is hosted on the same or another process component of the system's process components. ML model parameters are passed between the plurality of process components. Based on the queue allocation criterion, one or more of the process components are assigned (typically, automatically assigned) to one of the queues. ML model parameters of the process components in a selected one of the queues are attributed (typically, automatically attributed) to the ML model belonging to the selected queue.

[0008] According to another aspect, a non-volatile storage medium is provided. The storage medium stores a computer program. The computer program includes instructions, which, when executed on a processor of an industrial automation system, cause the processor to perform the method described herein.

[0009] The following examples, features, and details may be considered and / or applied to any one or more of the following: an industrial automation system as described herein, a computer-implemented method as described herein, a non-volatile storage medium as described herein.

[0010] As used herein, a process component refers to a device or equipment involved in executing a process belonging to an industrial automation application for which an industrial automation system is used. For example, a process is an automated process (such as automated production of chemical substances), a technical device, etc. Typically, each process component performs a specific physical, chemical, technical or virtual function in an automated process.

[0011] Typically, two or more of the process components host at least a portion of at least one ML model for each queue. For example, one process component hosts a complete ML model for the queue to which the process component belongs (e.g., is classified as / assigned to). Another process component belonging to a different queue, for example, hosts another complete ML model for the queue.

[0012] As used herein, a queue is a group of process components that share a predetermined characteristic. Typically, process components of the same queue perform the same or comparable physical, chemical, technical, or virtual functions in an automated process. As an example and not limitation, when considering an industrial automation system across sites or across plants, a first process component at a first site or a first plant performs the same or comparable function as a second process component at a second site or a second plant; the first process component and the second process component may share the same queue.

[0013] That is, different process components may be part of the same queue. For example, process components that contribute to comparable processes at different sites may be part of the same queue, ie, be classified as being in the same queue.

[0014] As used herein, queue allocation criteria refers to the principle of assigning or attributing a process component to a queue. For example, but not limited to, the queue allocation criteria may be to determine whether two process components perform the same or equivalent functions, and based on the determination, when the two process components perform the same or equivalent functions, assign the same queue to the two process components.

[0015] Therefore, assigning one or more of the process components to one of the queues according to the queue assignment criteria may include: determining whether two process components perform the same or equivalent functions, and based on the determination, assigning the same queue to the two process components when the two process components perform the same or equivalent functions.

[0016] Attributing ML model parameters of process components in a selected one of the queues to the ML model belonging to the selected one of the queues may include providing, to process components in the selected one of the queues that lack an ML model and / or have little data for performing ML operations on their own, ML model parameters of process components in the same selected one of the queues that have sufficient ML models.

[0017] A proximity value as used herein is a measure or degree of similarity that indicates the functionality of a process component in one queue to the functionality of a process component in another queue. A proximity criterion as used herein is, for example, a mathematical operation that allows the proximity value to be determined. For example, from an ML perspective, when process components have no reason to share the same queue, but their functionality is still sufficiently similar, they may have proximity values ​​that indicate such sufficient similarity.

[0018] Sharing ML model related data between process components belonging to the same adjacent queue group generally includes transferring ML model related data from one process component to another process component. Transferring may include any suitable form of data transmission and / or data reception, such as wired data communication or wireless data communication.

[0019] In one example, the technology described herein also includes: determining a performance value for each queue in a selected queue group in the adjacent queue group; selecting a queue in the selected adjacent queue group indicating expected performance as a performance queue based on the performance value; and using ML model-related data and / or ML model of the performance queue in at least one different queue in the selected adjacent queue group.

[0020] As used herein, a selected queue group for one of the neighboring queue groups may be determined by selecting one of the neighboring queue groups for selection.

[0021] As used herein, a performance value may be a measure or degree of performance of a process component in a corresponding queue in its corresponding function (one or more) (such as a physical, chemical, technical, or virtual function). As a non-limiting example, a performance value may include the actual output of an intermediate product in a function (subprocess), the actual control deviation of a function (subprocess), etc. As used herein, an expected performance may be a measure or degree of performance of a process component in a corresponding queue in its corresponding function (one or more) (such as a physical, chemical, technical, or virtual function). In addition to the above non-limiting examples, the expected performance may include the nominal output of an intermediate product in a function (subprocess), the nominal control deviation of a process (subprocess), etc. The relationship between the performance value and the expected performance may indicate how the actual performance of the process component in the corresponding queue is.

[0022] As used herein, a performance queue may indicate that the queue meets expected performance. For example, some or all of the process components included in the performance queue work well enough to perform their functions (sub-processes) as expected or intended.

[0023] In one example, the queue assignment criteria is defined as a hierarchical filter cascade. Each filter is configured to assign a filter output group to each process component as an output of a corresponding filter stage according to one or more component attributes.

[0024] As used herein, a hierarchical filter cascade refers to a series of filters in which the output of a filter belonging to a first filter stage is input to a filter belonging to a second filter stage, where the second filter stage is a successor to the first filter stage. However, the cascade is not limited to two filter stages and may include more than two (such as three, four or more) filter stages.

[0025] As used herein, a filter output group refers to the distinctions actually made by filters in a hierarchical filter cascade. That is, a filter output group typically includes the classifications performed by the corresponding filters.

[0026] As used herein, a component attribute refers to a characteristic or quality of the component under consideration. As long as the filter assigns a filter output group to each process component according to one or more component attributes, it effectively characterizes the component (eg, according to its quality).

[0027] In an example also related to hierarchical filter cascades, the techniques described herein also include assigning all process components exiting the last filter stage in the same filter output group to the same respective queue.

[0028] In one example, a proximity criterion for assigning a pair of queues to respective adjacent queue groups is met when a process component exits the second to last filter stage in the same filter output group but exits the last filter stage of a different filter output group.

[0029] In one example, the techniques described herein are applied to industrial automation involving a cement plant or multiple cement plants. For example, the industrial automation system includes one or more cement plants, that is, the industrial automation system involves controlling one or more cement plants. For example, in the case of multiple cement plants, each plant may have multiple industrial automation systems (such as but not limited to multiple site industrial automation systems per plant) and a superordinate system that takes into account or implements any interrelationships between the industrial automation systems of each plant required to perform the techniques described herein.

[0030] In an example relating to one or more cement plants, the component attributes include one or more of the following: a cyclone blockage detector, the type of cement produced in one or more plants, the type of fuel used in one or more plants, and data distribution of plant parameters (including one or more of fuel consumption, pressure, and temperature).

[0031] In an example relating to the component attributes just described, a cyclone blockage detector is included in a first filter stage of a graded filter cascade, while the type of cement produced in one or more plants is included in a second filter stage of the graded filter cascade.

[0032] In the examples related to the component properties just described, the fourth filter stage is the last filter stage and / or the third filter stage is the second to last filter stage.

[0033] In one example, at least one process component hosting the ML model for each respective queue is configured to pass ML model parameters to a server of at least some of the other process components acting as clients.

[0034] In one example, a cement plant cluster is provided, comprising the industrial automation system described herein, wherein process components are distributed across different cement plants in the cluster.

[0035] This summary is provided only to introduce certain concepts and is not intended to identify any key or essential features of the claimed subject matter. Other features of the inventive arrangement will become apparent from the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The arrangements of the present invention are illustrated by way of example in the accompanying drawings. However, the accompanying drawings should not be interpreted as limiting the arrangements of the present invention to the specific implementations shown. After reading the following detailed description and referring to the accompanying drawings, various aspects and advantages will become apparent.

[0037] Figure 1 illustrates a schematic diagram of an industrial automation system according to an embodiment;

[0038] Figure 2 illustrates a schematic diagram of an industrial automation system according to an embodiment;

[0039] Figure 3 A schematic example of a collaborative learning environment with a cohort is illustrated;

[0040] Figure 4 A schematic example of a cascade filter is illustrated;

[0041] Figure 5 A schematic example of cohort neighborhood based learning is illustrated;

[0042] Figure 6 A flowchart illustrating a method according to an embodiment; and

[0043] Figure 7 A program flow chart of a part of a method according to an embodiment is illustrated. DETAILED DESCRIPTION

[0044] It is believed that the various features described in the present disclosure will be better understood by considering the specification in conjunction with the accompanying drawings. The processes, systems, methods, etc. described herein and any variations thereof are provided for illustrative purposes and should be interpreted as a representative basis for teaching those skilled in the art to employ the features described herein (including variations thereof).

[0045] The present disclosure relates to technologies related to implementing industrial process automation using industrial automation systems. Collaborative learning (CL) in industrial automation refers to technologies in which multiple process components collaborate in training machine learning (ML) models. Each process component is related to industrial automation, for example, contributing to the implementation of an automated process (such as but not limited to the production of cement or other industrial products). Typically, in CL, the process components include multiple clients and at least one server, where the server hosts one or more ML models; however, it should be envisioned that some or all process components host at least a portion of one or more ML models in a peer-to-peer class configuration. For ease of explanation, and not intended to be considered limiting, the key concepts of CL are described in the context of a client-server model rather than a peer-to-peer model.

[0046] In a typical CL framework, each customer's raw data (such as measurement data, process data, etc.) is stored locally and is not exchanged or transmitted across site boundaries (e.g., factory boundaries). Instead, focused ML model updates intended for immediate aggregation are shared with a server (such as a central server) to achieve the goal of collaborative model learning.

[0047] However, CL can be challenging in some cases. In order to collaborate in CL, the raw data has to be transferred to a (central) server. When the raw data is transferred to a server, security risks (such as compromised data integrity, accidental proliferation of data, etc.) and / or expensive IT challenges (such as high overhead of encryption and data transfer, etc.) may arise. This is especially true when dealing with multiple plants spread across different industrial sites. Other kinds of challenges may stem from the fact that multiple plants on different sites may have fundamental differences in their data and its distribution, processes, environmental conditions and operating conditions.

[0048] Federated learning (FL) is a machine learning technique that trains algorithms across multiple distributed edge devices or servers that hold local data samples without exchanging them. FL enables multiple process components to build a common ML model without sharing raw data.

[0049] Figure 1FIG. 1 shows a schematic diagram of an industrial automation system 100 according to an embodiment. The industrial automation system 100 includes process components 200, 301, 302, 303, 304. Figure 1 In the example, the process component 200 is configured as a server, such as a centralized server, and the process components 301, 302, 303, 304 are each configured as a client. Each client 301, 302, 303, 304 is connected to the server 200 via a corresponding data connection 311, 312, 313, 314 for transmitting, receiving and / or exchanging data. The data connections 311, 312, 313, 314 include but are not limited to wired data connections or wireless data connections, such as bus systems such as CANBUS, Ethernet connections, etc. The server 200 hosts the machine learning model M1 and the machine learning model M2. Figure 1 In the embodiment, some or all clients 301 , 302 , 303 , 304 perform ML processes involving model parameters, and communicate some or all ML model parameters between the server 200 .

[0050] Figure 2 FIG. 1 shows a schematic diagram of an industrial automation system 100A according to another embodiment. The industrial automation system 100 includes process components 301, 302, 303, and 304. Figure 1 In the example of FIG. 3 , process components 301, 302, 303, 304 are connected to each other via data connections 321, 322, 323 for transmitting, receiving and / or exchanging data. Data connections 321, 322, 323 are not limited to Figure 2 The topology shown may include direct and / or indirect connections between any of the process components 301, 302, 303, and 304. The data connections 321, 322, and 323 include, but are not limited to, wired data connections or wireless data connections, such as bus systems such as CANBUS, Ethernet connections, and the like. Figure 2 In the example of , no central server 200 is provided; instead, some or each process component 301, 302, 303, 304 hosts at least a portion of the machine learning model M1 and / or the machine learning model M2. Figure 2 In the non-limiting example of , process component 301 hosts ML model M1 and process component 303 hosts ML model M2. Figure 2 In the embodiment, some or all of the process components 301 , 302 , 303 , 304 perform ML processes involving model parameters, and some or all of the ML model parameters are communicated between some or all of the other process components 301 , 302 , 303 , 304 .

[0051] It should be understood that the configuration of the industrial automation system 100, 100A is merely an example, and other configurations and / or topologies may be employed, such as combining some features of the industrial automation system 100, 100A or omitting some features, as long as at least one of the process components 301, 302, 303, 304 is configured to execute an ML process involving ML model parameters, and at least one of the process components 301, 303, 304 is configured to host at least a portion of at least one ML model and pass ML model parameters between multiple process components 301, 302, 303, 304.

[0052] Unless otherwise stated, the following descriptions apply to Figure 1 Industrial Automation Systems 100 and Figure 2 Industrial Automation System 100A.

[0053] In the industrial process automation system 100, 100A, some or all of the process components 301, 302, 303, 304 are configured to perform an ML process, such as an FL process. In the ML process, machine learning model parameters are derived by training the ML model to transform raw data (such as, for example, measurement data, process data, etc.) from the process components into the ML model. Further, in the industrial process automation system 100, 100A, some or all of the process components 301, 302, 303, 304 are configured to host at least a portion of at least one ML model, and to transfer ML model parameters belonging to the hosted model between the process components 301, 302, 303, 304.

[0054] Each of the process components 301, 302, 303, 304 can be classified into a queue corresponding to a queue allocation criterion (i.e., the principle by which the process components 301, 302, 303, 304 are assigned to or attributed to a queue). The ML models M1, M2 can be considered as common models for each queue. For example, in Figure 1 In the client-server approach of , the server 200 hosts the public model M1 and the public model M2, and the clients 301, 302, 303, 304 run local models delivered (eg, broadcasted) by the server 200 according to their respective queues. Figure 2 In the peer-to-peer approach, process component 301 hosts the public model M1, process component 303 hosts the public model M2, and process components 301, 302, 303, 304 run local models passed between process components 301, 303, 304 and according to their respective queues.

[0055] In either case, a typical workflow iteration may be as follows: common models M1, M2 are passed between process components 301, 302, 303, 304; anomaly prediction is performed locally on process component(s) 301, 302, 303, 304 running models M1, M2; retraining is performed locally on process component(s) 301, 302, 303, 304 running models M1, M2; retrained models are sent back to server 200 or peer process components 301, 302, 303, 304; server 200 or peer process components 301, 302, 303, 304 perform accuracy and model aggregation; aggregated models are communicated between process components 301, 302, 303, 304; next iteration begins. In other words, in Figure 1 In the exemplary client-server model of , first, the server 200 sends a global consensus model, for example, an ML model for anomaly prediction, to the client. The clients 301, 302, 303, 304 receive the model from the server 200 and retrain the model using local data. The client then shares the updated model weights with the server 200. These tasks are collectively referred to as communication rounds or iterations.

[0056] Figure 3 An illustrative example of a collaborative learning environment with a queue is illustrated. A cement plant 500 includes laboratory data 501, which is sent to multiple plants 601, 602, 603, 604. Each plant 601, 602, 603, 604 communicates with a server 200 via a corresponding CL enabler 611, 612, 613, 614. Server 200 hosts ML model M1 and ML model M2. Plant 601 and plant 602 are classified as the same queue Co1. Plant 603 and plant 604 are classified as the same queue Co2. Plant 601 hosts 600 items of laboratory data. Plant 602 hosts 500 items of laboratory data. Plant 603 hosts 200 items of laboratory data. Plant 604 hosts 300 items of laboratory data. Therefore, queue C1 is considered a strong queue. Queue Co2 is considered a weak queue. Inter-queue Co1, Co2 knowledge (i.e., domain knowledge such as fuel type, cement and / or production process) contributes to the queue allocation.

[0057] The CL enablers 611, 612, 613, 614 perform one or more of the following: queue attribute collection, model validation, communication, model adaptation, optimized retraining. In queue attribute collection, attributes for queue allocation are collected from the factory nodes 601, 602, 603, 604. In model validation, the new global model is validated before replacing the existing local model. In communication, the communication protocol (such as but not limited to gRPC or OPC UA) is adapted. In model adaptation, adaptation is performed on the base ML model. In optimized retraining, the retraining frequency is optimized based on parameters such as data sampling frequency.

[0058] The CL enablers 611, 612, 613, 614 facilitate the division of the plants 601, 602, 603, 604 into different clusters (or queues) based on different parameters such as environmental or operating conditions, task types, etc. Non-limiting examples of attributes that may be considered for queue assignment include tasks (e.g., the purpose of the ML-based soft sensor running in the plant), environmental / operating conditions, attributes of the data recorded by the plant, and data distribution in different plants such as mean, variance, skew, kurtosis, etc. The CL enablers 611, 612, 613, 614 collect these attributes from the plants 601, 602, 603, 604 and pass them to the process components 200, 301, 302, 303, 304 for queue assignment.

[0059] Figure 4 An illustrative example of a cascade filter CF for performing a queue allocation operation is illustrated. Queue allocation may include dividing nodes (factories 601, 602, 603, 604; process components 301, 302, 303, 304) into queues and finding the neighboring queues of each queue. Figure 4 In FIG. 4 , the cascade filter CF includes filters on filter stages F1, F2, F3, and F4. The output of the filter in filter stage F1 is input to the filter in filter stage F2. The output of the filter in filter stage F2 is input to the filter in filter stage F3. The output of the filter in filter stage F3 is input to the filter in filter stage F4. F4 is the last filter stage. Figure 4The process components in are represented by clients C1 to C25. The filters in the cascade filter CF have different properties, examples are given below. Based on the filter properties, clients C1...C25 are divided into queues. Filter stage F4 shows queues as outputs. Adjacent queues are defined as queues that are different from the last filter. For example, clients C1, C2, C3, C4 belong to queue 1, C5, C6, C7 belong to queue 2, they are adjacent queues and they are different with respect to filter F4. Process components 301, 302, 303, 304 belonging to one queue contribute to one ML model M1, M2.

[0060] As an example related to the cement industry, consider the following filter to group various cement plants:

[0061] -F1: Soft sensor for cyclone separator blockage detection;

[0062] -F2: Type of cement being produced;

[0063] -F3: Type of fuel used in the plant;

[0064] -F4: Data distribution of parameters such as fuel, fuel consumption, pressure and temperature.

[0065] Figure 5A schematic example of queue neighborhood-based learning for explaining a queue neighborhood learning method for improving model accuracy of poorly performing queues is illustrated. For example, consider that queue Co1 is training model M1 and queue Co2 is training model M2, which are neighbors, i.e., adjacent queues. Data and features Xa=[x1 x2x3] and corresponding target set L1 can be obtained at queue Co1. Queue Co2 has data with features Xb=[x1 x2 x3 x4 x5] and no targets or noisy or incomplete targets. We can train model M2 with the support of model M1 of queue Co1. Since the feature space of queue Co1 is a subset of queue Co2, the partial soft target L2s is obtained by sending Xbs=[x1 x2 x3] as input using ML model M1. A heuristic function (here as an example, Hf) is determined based on domain knowledge to derive partial soft targets using the feature set [x4, x5], i.e., L2f=Hf(x4, x5). Hf is the alkali-sulfur ratio, x4 represents alkali, and x5 represents sulfur. A combined function F(L2s, L2f) such as the weighted sum W1*L2s+W2*L2f is used to obtain the complete soft target set for queue Co2, that is, L2=W1*L2s+W2*L2. Now, the model M2 is trained using the feature set Xb and the soft target L2. In the example of a cement plant, M1 can be a machine learning model for cyclone blockage prediction, which is built using parameters x1 as fuel type, x2 as fuel consumption, and x3 as kiln oxygen.

[0066] Figure 6 A program flow chart of a method according to an embodiment starting from 1001 is illustrated. In 1001, an ML process involving ML model parameters is executed on one of the process components of the system. In 1002, at least a portion of at least one ML model for each queue is hosted on the same or another process component of the system, and the ML model parameters are transferred between multiple process components. In 1003, one or more of the process components are assigned (usually, automatically assigned) to one of the queues according to the queue assignment criteria. In 1004, the ML model parameters of the process components in the selected queues of the queues are attributed (usually, automatically attributed) to the ML model belonging to the selected queue. In 1005, a proximity value for each pair of queues is determined (usually, automatically determined). In 1006, if the proximity value meets the predetermined proximity criteria, a pair of queues is assigned (usually, automatically assigned) to a corresponding adjacent queue group. In 1007, ML model related data is shared (usually, automatically shared) between process components belonging to the same adjacent queue group.

[0067] Figure 7 Illustration of the supplementary procedure flow chart Figure 6In 1008, a performance value is determined (typically, automatically determined) for each queue in a selected one of the adjacent queue groups. In 1009, a queue in the selected adjacent queue group indicating desired performance is selected (typically, automatically selected) as a performance queue based on the performance value. In 1010, the ML model-related data and / or the ML model of the performance queue is used in at least one different queue in the selected adjacent queue group.

[0068] The benefit of adopting the technology as described herein is that common CL / FL problems, such as limited data availability and localized learning, are solved. Therefore, robust predictions can be achieved and productivity can be improved. Further, the benefit of adopting the technology as described herein is that it is impossible to take sensitive data off-site, thereby making data-related services more widely accepted by the operating entities of the factory. Further, the benefit of adopting the technology as described herein is that the technology described herein is sustainable, which can make model learning more easily transferable to new factories. Further, the benefit of adopting the technology as adopted herein is that the engineering workload is reduced. Further, the benefit of adopting the technology as described herein is that the solution is scalable because the solution can be used across multiple factories with minimal changes.

[0069] Ongoing engagement with the experts and operators results in ongoing service opportunities for additional MPC performance monitoring and improvements as well as additional MPC implementations.

[0070] The description of the arrangement of the present invention provided herein is for illustration only and is not intended to be exhaustive or limited to the disclosed forms and examples. The terms used herein are selected to explain the principles, practical applications or technical improvements of the technology found in the market for the arrangement of the present invention, and / or to enable those of ordinary skill in the art to understand the arrangement of the present invention disclosed herein. For those of ordinary skill in the art, modifications and variations may be obvious without departing from the scope and spirit of the described arrangement of the present invention. Therefore, reference should be made to the following claims rather than to the above disclosure to indicate the scope of these features and implementations.

Claims

1. An industrial automation system (100, 100A) for implementing at least one industrial process, the industrial automation system (100, 100A) comprising a plurality of process components (200; 301, 302, 303, 304), each process component being capable of being classified into a queue corresponding to a queue allocation criterion, in: At least some of the process components (301, 302, 303, 304) are configured to perform ML processes involving machine learning ML model parameters; and At least one of the process components (200) is configured to host at least a portion of at least one ML model for each queue and is further configured to communicate the ML model parameters between the plurality of process components; The system is configured as follows: assigning one or more of the process components to one of the queues according to the queue assignment criteria; attributing the ML model parameters of process components in a selected one of the queues to the ML model belonging to the selected queue; Determine the proximity value of each pair of queues; assigning a pair of queues to corresponding adjacent queue groups if the proximity values ​​satisfy a predetermined proximity criterion; as well as The ML model related data is shared between process components belonging to the same adjacent queue group.

2. The industrial automation system (100, 100A) according to claim 1, wherein the system is further configured to: determining a performance value for each queue in a selected one of the adjacent queue groups; selecting the queue indicating expected performance in the selected adjacent queue group as a performance queue based on the performance value; as well as The ML model-related data and / or the ML model of the performance queue are used in at least one different queue in the selected group of adjacent queues.

3. An industrial automation system (100, 100A) according to any of the preceding claims, wherein the queue allocation criteria is defined as a cascade of hierarchical filters, wherein each filter is configured to assign a filter output group to each process component as an output of the corresponding filter level based on one or more component attributes.

4. The industrial automation system of claim 3, wherein the system is further configured to assign all process components exiting the last filter stage in the same filter output group to the same respective queue.

5. An industrial automation system according to claim 3 or 4, wherein the proximity criterion for assigning a pair of queues to corresponding adjacent queue groups is met when a process component leaves the second to last filter stage in the same filter output group but leaves the last filter stage of a different filter output group.

6. An industrial automation system according to any one of the preceding claims, wherein the industrial automation system comprises one or more cement plants.

7. An industrial automation system according to claim 6 when dependent on any one of claims 3 to 5, wherein the component attributes include one or more of the following: a cyclone separator blockage detector; the type of cement produced in one or more factories; the type of fuel used in the one or more factories; data distribution of factory parameters including one or more of fuel consumption, pressure and temperature.

8. The industrial automation system of claim 7, wherein the cyclone separator blockage detector is included in a first filter stage of the graded filter cascade, the type of cement produced in the one or more plants is included in a second filter stage of the graded filter cascade, the type of fuel used in the one or more plants is included in a third filter stage of the graded filter cascade, and the data distribution of the plant parameters is included in a fourth filter stage of the graded filter cascade.

9. The industrial automation system of claim 8, wherein the fourth filter stage is the last filter stage and / or wherein the third filter stage is the penultimate filter stage.

10. An industrial automation system according to any one of the preceding claims, wherein at least one of the process components hosting the ML model of each corresponding queue is configured as a server that passes the ML model parameters to at least some of the other process components acting as clients.

11. A group of cement plants, comprising an industrial automation system according to any one of the preceding claims, wherein the process components are distributed over different cement plants in the group.

12. A computer-implemented method executed in an industrial automation system for implementing at least one industrial process, the industrial automation system comprising a plurality of process components, each process component being capable of being classified into a queue corresponding to a queue allocation criterion, the method include: On one of the process components of the system, executing (1001) an ML process involving machine learning (ML) model parameters; as well as hosting (1002) at least a portion of at least one ML model for each queue on the same or another of the process components of the system, and communicating the ML model parameters between the plurality of process components; automatically assigning (1003) one or more of the process components to one of the queues according to the queue assignment criteria; automatically attributing ( 1004 ) the ML model parameters of process components in a selected one of the queues to the ML model belonging to the selected queue; automatically determining (1005) a proximity value for each pair of queues; If the proximity value satisfies a predetermined proximity criterion, automatically assigning (1006) a pair of queues to corresponding adjacent queue groups; as well as The ML model related data is automatically shared (1007) between process components belonging to the same adjacent queue group.

13. The method according to claim 12, further comprising: include: automatically determining (1008) a performance value for each queue in a selected one of the adjacent queue groups; automatically selecting (1009) the queue indicating desired performance in a selected group of adjacent queues as a performance queue based on the performance value; as well as The ML model-related data and / or the ML model of the performance queue are used (1010) in at least one different queue in the selected group of adjacent queues.

14. The method according to any one of claims 12 to 13, further comprising: include: Filters in a hierarchical filter cascade implementing the queue assignment criteria are applied to the process components, wherein each filter as an output of a corresponding filter stage assigns a filter output group to each process component according to one or more component attributes.

15. The method according to claim 14, further comprising: include: All process components leaving the last filter stage in the same filter output group are assigned to the same corresponding queue.

16. The method according to any one of claims 14 to 15, further comprising: include: When a process component leaves the second to last filter stage in the same filter output group but leaves the last filter stage of a different filter output group, a pair of queues is assigned to corresponding adjacent queues.

17. A method according to any one of claims 12 to 15, wherein a proximity criterion for assigning a pair of queues to respective adjacent queue groups is met when a process component leaves the second to last filter stage in the same filter output group but leaves the last filter stage of a different filter output group.

18. A non-volatile storage medium having a computer program stored thereon, the computer program comprising instructions which, when executed on a processor of an industrial automation system, cause the processor to perform the method according to any one of claims 12 to 17.

Images