Automatic configuration for system controllers
The use of natural language processing and semantic matching to create a digital twin of energy consumption systems automates data integration and analysis, addressing the complexity and cost issues in existing monitoring systems, enabling efficient energy management and optimization.
Patent Information
- Application Number
- PCT/EP2024/061874
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-08-01
- Filing Date
- 2024-04-30
- Publication Date
- 2024-11-07
AI Technical Summary
Current energy consumption monitoring systems face high complexity and costs due to heterogeneous communication protocols and semantics, requiring manual effort to integrate and analyze data points, which limits their widespread availability and effectiveness in optimizing energy usage in buildings.
A computer system utilizing natural language processing and semantic matching to classify and map control data from existing systems onto a uniform information model, creating a digital twin that automates data integration and analysis, reducing manual effort and enabling automated generation of monitoring dashboards.
This approach allows for efficient, automated processing and optimization of energy consumption data, achieving high accuracy in data classification and integration, thereby facilitating optimized energy management and reducing greenhouse gas emissions.
Smart Images

Figure EP2024061874_07112024_PF_FP_ABST
Abstract
Description
[0001] Automated configuration for plant controls
[0002] The invention relates to a computer system and an associated method for the automated creation of a configuration for a system control system, in particular for monitoring or consumption management of buildings. Data from an existing system control system or automation system is acquired, classified, and interpreted and assigned using language models and functions such as semantic matching (SM). In particular, the invention relates to an innovative method for mapping semantically heterogeneous control data or system information, in particular in the form of data points, of an existing technical system to a control model standard without manual configuration, thereby forming the basis for self-X monitoring applications, particularly for energy or consumption management.
[0003] In particular, 35% of Germany's final energy consumption is used for the use and operation of residential and non-residential buildings. This operation causes approximately 28% of greenhouse gas emissions in Germany. International agreements such as the Paris Climate Agreement call for a significant reduction in greenhouse gas emissions to prevent major climate change. To this end, the goal is to increase the renovation rate in Germany to 2%. In terms of control technology, technical systems generally face the same challenges and tasks as buildings.
[0004] It is often unclear which modernization measure will be the most effective if the current operating status of a system, particularly of a building, cannot be represented and thus assessed. This gives rise to the need to first comprehensively record the energy consumption status of a system. Significant savings can often be achieved by optimizing the existing technical components without the need for major modernizations. Therefore, the condition of the existing system must first be recorded and a system control system must be set up, particularly in the case of building automation or energy consumption monitoring. Various obstacles, such as the high complexity of setting up such systems, high costs and a lack of data, stand in the way of the widespread availability of energy consumption monitoring applications.The reasons for this high level of complexity are usually heterogeneous communication protocols at the so-called field level and the heterogeneous semantics of the transmitted information. In particular, a consistent nomenclature for naming identical, equivalent, or identically typed technical components in the existing system is often not used. To integrate the information into energy consumption monitoring, also referred to as energy monitoring, various protocols typically have to be processed, their information analyzed, and integrated into the energy consumption monitoring system. Analyzing the data points and integrating them at the correct point in the energy consumption monitoring system usually requires a significant amount of manual effort.
[0005] Due to a semantically heterogeneous description of a plant's technical components in a plant control system, often in the form of so-called data points, and / or due to inconsistent control standards, it is often not possible to search for, for example, the power consumption of specific consumers such as all pumps used and to automatically integrate this into a newly created plant control system or user application. The data points typically have to be analyzed manually and integrated into a newly created plant control system or user application.
[0006] Currently, no solutions are known that specifically process the semantic information of a system control system automatically. For example, a semantic matching service has been specified for building automation data, which specifically uses the BACnet protocol. Each BACnet data point has a name (object name) and a description (description). However, this depends on the manufacturer or the building automation company that implements the system. If the data points are to be integrated into a monitoring application or application at the management level, the data point in question must be manually analyzed. Only then can it be integrated into the correct location in a new user application or a new control program. Current solutions for the semi-automated processing of data are based on a behavioral analysis of the control data in combination with metadata. For example, there isSolutions that record the course of a data point and, together with the data point's unit, assign it to a system. For example, a data point from a building that has a temperature of approximately 65°C can be assigned to a temperature of a heating system. However, these approaches are not yet widespread. Currently, data, particularly from buildings, is integrated into monitoring applications through manual assignment by engineers: The data from a building is recorded, analyzed, and added to the correct locations in the monitoring application. This is time-consuming and costly. As a result, the majority of buildings do not have energy consumption monitoring, meaning optimization potential remains undiscovered and buildings are not operated optimally.
[0007] Based on this, the object of the invention is to improve the recording, analysis, and / or optimization of existing plant control systems. In particular, the recording of technical components for software-based optimization of existing systems, such as buildings, is to be improved.
[0008] This problem is solved by the subject matter of patent claim 1 and the independent method claim. Preferred developments can be found in the subclaims.
[0009] The invention provides a computer system for processing control data of a plant with a plant control, with a computer and an interface to a plant control, wherein the plant has at least one technical component which represents at least one data point with regard to the control data, wherein the interface is connected to the computer and the plant control, which exchanges control data with the technical components of the plant via a network, wherein the interface captures the control data and transmits it to the computer, and wherein the computer system is configured to classify a data point by means of natural language processing, to assign it to a technical component and / or to create a digital control image of the plant control.
[0010] The invention particularly encompasses the semantic processing and mapping of system and / or operating data from a technical system automation system, particularly a building, into a uniform information model. Part of the invention also encompasses a gateway that can be integrated into a system with automation. The gateway collects control data from the technical components of the system. The invention particularly focuses on the automated analysis and processing of non-uniformly named or labeled, and thus heterogeneous, control data from an existing system control system using an algorithm. This algorithm is based particularly on methods of natural language processing (NLP), a branch of artificial intelligence for language processing. Different data point designations are processed, interpreted, and presented in a structured, understandable information model.This makes it possible to provide a so-called semantic digital twin of an existing plant automation or plant control system, essentially independent of the nomenclature originally chosen when the plant was created.
[0011] In order to optimize the energy efficiency of a system, such as a building, the system or operating data, such as power consumption, energy consumption, or switching cycles of the technical components, must be systematically recorded, analyzed, and holistically optimized. An operator's various systems usually have a large amount of operating data, particularly relating to relevant consumers, which can be represented or characterized by a data point. Such a data point, also referred to as a data vector, can contain, in particular, the power consumption of a consumer and / or its designation and / or function and / or an energy consumption profile, and can also be in the form of a multidimensional matrix with data entries.A data vector contains, in particular, the data of a technical system or an individual system component that is made available externally by the system control system. Data points are also, in particular, acquired, processed, or transmitted by the system control system for system operation. This can be physical or virtual information, which is acquired or generated, in particular, by sensors, actuators, or other components of the technical system.
[0012] The computer can also be cloud-based, so that applications in particular can be executed remotely. The interface is in particular a gateway, which is typically located between the computer and the technical system and is in particular connected to an existing control system, in particular via its network such as a BUS system. A digital control image of the system control comprises in particular a control program-related and / or emulating and / or virtual computer program, which can replace and / or control and / or intervene in the system control of an existing system. The digital control image can comprise a control and / or information model and can in particular provide the control data of the existing system within the digital image.
[0013] By semantically processing and displaying the system information as a digital twin, the invention enables the automated generation of dashboards for monitoring applications, particularly for buildings. By implementing executable rules, different dashboards can be created based on the detected building data. For example, if data from a heat pump is found in one building and data from a gas boiler in another, different applications can be created. Dashboards are therefore generated automatically, tailored to the user data. This is based on a process that maps heterogeneous data to a unified model.
[0014] The result in the form of a digital control image, or a control and / or information model, can - as previously described - also be referred to as a digital twin of a plant with technical components. Preferably, all data points of a plant are processed and classified by the NLP models, with the data being assigned to categories or classes that were defined in a preliminary training run and learned by the models. This allows patterns and relationships in the control data to be learned, which enable a classification of the operating data. Based on the classifying data points, an entire digital control image of a plant can be created, which can look different for each plant.The basic structure of the digital control image is generally the same for different user applications, so the digital control image forms the basis for automated monitoring applications. Executable rules can be implemented for the applications, which search a digital control image for specific data points. This is made possible in particular by semantic processing and uniform representation. This allows, for example, the following rule to be implemented: Search for each flow temperature of a building's heating systems and create a line chart for it. These rules then only need to be implemented once and can then be executed each time the system is commissioned, allowing new user applications to be created automatically. Over time, trends in the system's data are recorded.These can form the basis for optimization suggestions or anomaly detection within the system.
[0015] A semantic matching service is based on methods from the research field of NLP. NLP is a branch of artificial intelligence that enables computers to understand natural language and then execute actions based on it. State-of-the-art NLP models are based primarily on the transformer architecture. It is characterized by an encoder-decoder structure. The encoder is used for language analysis and classification, and the decoder for language generation tasks. To cover more specific requirements in the areas of classification or generation, today's models are usually characterized by either an encoder or decoder structure and are referred to as pretrained language models (PLMs). These are first trained on large amounts of general text (pretraining) and then refined (fine-tuning) or further trained for specific NLP tasks (e.g., text classification).The combination of general pretraining followed by fine-tuning for a specific task is called transfer learning. Based on these PLMs and the specific NLP tasks of text classification (TC) and natural language inference (NLI), heterogeneous semantics can be processed.
[0016] Preferably, the computer system is configured to use natural language processing to assign and / or interpret different data point designations from the building automation system based on language models, preferably based on word embeddings generated from transformer models and / or long-short-term memory-based models. This allows, in particular, a uniform assignment with regard to the designation of the technical components and the control system as a whole to be achieved.
[0017] In a particularly preferred embodiment, the computer system is configured to train and / or refine the language models, in particular the transformer models and / or long-short-term memory-based models, with texts, so that they possess, in particular, a basic understanding of language and / or appropriate associations. These models are typically pre-trained on large amounts of text and, in a further step, are refined on labeled data points from the building automation system, so that they are capable of mapping heterogeneous data points from the building automation system to the developed information model.
[0018] In a likewise preferred embodiment, the computer system is configured to train, refine, and / or improve the transformer models and / or long-short-term memory-based models on preselected data points in a further step, thereby significantly improving the accuracy of the assignment results and the creation of a control image. During operation, the data point undergoes a multi-stage process, which classifies the data point with increasing finer detail (text classification). For example, in the case of building management, the basic function (e.g., supplying air) to which the data point is assigned is first classified. In a next - possibly second - step, the more precise function is classified (here, for example, providing air), and in a possible third step, the component (supply air fan). The last step classifies the data point, for example, as a switching command (e.g.,Supply air fan). Through a possible multi-stage process of increasingly fine classification, a very high mapping accuracy of sometimes over 95% can be achieved. The computer system is preferably configured to create a monitoring application from the control and / or information model. A monitoring application particularly comprises a customized software-based application that can control the technical components or with which the user can influence the controls, particularly via a user interface. The monitoring application preferably relates to an energy monitoring application.
[0019] In particular, the computer system is configured to create a visualization of the control and / or information model. This may include one or more dashboards that are created automatically and that the user can use to view data or make further configurations. This can achieve a high level of user-friendliness and / or operability.
[0020] The invention also relates to a method for processing control data of a system with technical components, wherein a technical component represents a data point with respect to the control data, comprising the following steps:
[0021] Recording the control data via an interface,
[0022] Sending the control data to a computer,
[0023] Classifying a data point using Natural Language Processing, assigning the data point to a technical component and
[0024] Creating a digital control and / or information model that represents a digital control image of the plant control system.
[0025] In particular, the method comprises the interpretation of different data point labels from the building automation system using natural language processing, in particular based on word embeddings generated from transformer models and / or long-short-term memory-based models; training the aforementioned models with texts, in particular so that they have a basic understanding of language, and refining the aforementioned models on preselected data points.
[0026] The method may also include creating a monitoring application from the control and / or information model. Visualizing the control and / or information model and / or creating one or more dashboards is also conceivable.
[0027] The invention is explained in more detail below using preferred embodiments with reference to the drawings.
[0028] The drawings show
[0029] Fig. 1 shows the computer system according to the invention in a preferred embodiment, Fig. 2 shows the method steps according to the invention in a preferred embodiment,
[0030] Fig. 3 shows an overview of data points and labels for a preferred embodiment of the invention,
[0031] Fig. 4 Labeling using the example of a data point,
[0032] Fig. 5 an illustration related to Natural Language Inference (NLI),
[0033] Fig. 6 a tabular overview of training results of the different models and
[0034] Fig. 7 a general structure for building management in companies.
[0035] Figure 1 schematically shows the computer system 1 according to the invention in a preferred embodiment, comprising a computer 5 and an interface 6. The interface 6 connects the computer system 1 to a system control system or system 4, which has several technical components 3 or is also networked in a network 7. Control data 2 are exchanged with the computer 5 via the interface 6.
[0036] Figure 2 shows schematically the most essential steps according to the invention, which steps include:
[0037] Sl : Acquisition of control data 2 via an interface 6,
[0038] S2: Sending the control data 2 to a computer 5,
[0039] S3: Classifying a data point using Natural Language Processing,
[0040] S4: Assignment of the data point to a technical component 3 and S5: Creation of a digital control and / or information model 8, which represents a digital control image of the plant control.
[0041] One way to avoid the need to manually enter and assign control data 2, especially building automation data, is to implement a Semantic Matching (SM) service, which automatically processes heterogeneous semantics and maps them to a semantic standard. The prerequisite for such applications is a uniform information structure, i.e., a uniform information model on which the components are based. In order to automatically integrate the components or their automation stations into, for example, an energy monitoring system, the information must first be integrated into a digital twin (e.g., the Asset Administration Shell (AAS)). The result is a AAS per automation station with the submodels "Asset Interface Description" and "Datapoint Information." The "Datapoint Information" submodel is subsequently processed by the SM service.The heterogeneous information is mapped to building services functions using artificial intelligence methods, specifically natural language processing (NLP). The results are then mapped to the "Operational Information" and "Monitoring" submodels, which are referred to as structural information. The structural information obtained through processing in the submodels is integrated into the energy monitoring system using rule-based automation, thus enabling so-called Self-X energy monitoring applications for existing systems.
[0042] In the field of Industry 4.0 (14.0), the concept of the VWS has been established as the basis for interoperability. The VWS is the digital representative of an asset in the digital world. The information model of the VWS defines how information about different assets must be structured. The central building block of the VWS are submodels that map the properties and functionalities of the assets. Submodels represent, for example, the topics of identification, design or configuration. Access to the information of a VWS is realized through interfaces provided by the VWS. To enable automated access to the interfaces of different VWS, these interfaces are standardized. The literature often describes four levels that are passed through during the standardization of interfaces. For example, levels 1 and 2 of various interfaces are described, e.g. “GetSubmodel”.This interface can be called to access a specific submodel of a VWS. The general interfaces can be provided in various services.
[0043] Figure 3 provides an exemplary overview of the data points and their assigned labels. The initial dataset consists of 8,809 data points from the cities of Cologne and Hamburg. These were labeled and assigned to the various functions and components of the classification method.
[0044] In the first step, all data points were assigned to one of the basic functions of the building services engineering (heat, air, cooling, media, power supply, security, transport). If a data point could not be assigned to any of the functions due to missing or incomprehensible names or descriptions, it was assigned to the “Other systems” category. The basic functions of transport, power and cooling supply were underrepresented in the data set, so the data points associated with them were only labeled up to the first level. The basic functions of air, heat and media supply were divided into further functions and the data points were each assigned to one of these functions (dashed line). In the third step, the data points were assigned to components or higher-level areas of the building services engineering (dotted line). In the final step, all data points are assigned a label that expresses the actual information of the data point.An example labeling process of a data point is shown in Figure 4. The individual data sets were used to train NLP models.
[0045] Current state-of-the-art models are typically trained using transfer learning. First, a PLM must be selected. Since the data points in the dataset consist of German names and descriptions, only German PLMs were considered. Several German models are described in [Chan, B., Schweter, S., and Möller, T.: German's Next Language Model. Proceedings of the 28th International Conference on Computational Linguistics. Barcelona,: International Committee on Computational Linguistics 2020]. Of these, the "gBERT-base" model was selected as the baseline because it delivered the best results in initial tests on the created dataset. The model is based on the architecture of the BERT-base model and is an encoder model. Encoder models are particularly suitable for classification tasks such as TC.The models of the first three levels (basic function, secondary function, component) were trained using TC, while the final level, the data point level, was trained using NLI. The final level was not trained using TC because, on the one hand, the number of resulting models was too large, and, on the other hand, the amount of training data for the individual components was too small.
[0046] Figure 3 shows 35 technical components 3. If fine-tuning on TC were also performed for the fourth stage, 35 models would have to be trained. Since an even larger number of components 3 is expected in the future due to the expansion of the training dataset, the number of models to be trained would continue to rise, which would lead to rising costs for training and the models in operation. The data points of the listed components 3 are also unevenly distributed; for example, 807 data points are assigned to the pump, while only 16 data points are assigned to the exhaust air damper. As a result, underrepresented components 3 do not have sufficient data to train high-performance models. By using the NLI method, it is possible to train just one model, which can be used for each component, as shown as an example in Figure 5.
[0047] In the NLI approach, the name and description of the data point are used as a premise. This is contrasted with a hypothesis consisting of a defined introduction and a label (e.g.: The data point describes: alarm message). Since a total of 241 labels are possible, each premise is contrasted with 241 hypotheses. The model learns to predict which of the hypotheses fits the data point (entailment) or does not describe the data point (contradiction). During training, all labels (241) are contrasted with all data points (5,485). The four-stage classification process makes it possible to reduce the number of candidate labels during operation to the number that are suitable for the predicted component. The results of the various models with training parameters are shown in tabular form in Figure 6.
[0048] The Fl score was used as the evaluation metric: Fl Score = 2 * (Precision * Recall) / (Precision + Recall). The results show that the different models can be successfully used to classify data points. Except for the Media model, all models achieved a score above 95%. The trained models were combined in a pipeline and are deployed as a 14.0 service.
[0049] For automated monitoring, a general structure was first designed to fit into the context of a company's building management (see Fig. 7). For this purpose, a hierarchical structure was created consisting of several levels with VWS. The top level is the company's VWS including the associated submodel, which contains the information about the company. A company can have several properties, on which in turn several buildings are constructed. Up to this level, the VWS must be created manually. The further levels and submodels "Building Operational Information" and "AMEV Building Monitoring" are created automatically using the information from the "NLP Classification Result" submodel. All automation stations in a building are read out and processed by the 14.0 NLP service.Depending on the classification result, different VWSs are created. If, for example, all data points of a building are assigned to the basic functions of heat and air supply, two VWSs are created (one per basic function). This is repeated for the next levels: system (e.g. two separate heating systems within a building), sub-system (three heating circuits and one generator per system), and component (e.g. pump). The appropriate data points are then assigned to the respective levels within the sub-models “Operational Information” and “Monitoring.” In this way, a hierarchical level structure (structural information) is created from VWSs automatically (Self-X). Monitoring applications can then be created using automatically executable rules. List of reference symbols.
[0050] 1 computer system
[0051] 2 Control data 3 Technical components
[0052] 4 Appendix
[0053] 5 computers
[0054] 6 Interface / Gateway
[0055] 7 Network 8 Control and / or information model
Claims
Patent claims 1. Computer system (1) for processing control data (2) of a plant (4) with a plant control system, with a computer (5) and an interface (6) to a plant control system, wherein the plant (4) has at least one technical component (3) which represents at least one data point with respect to the control data (2), wherein the interface (6) is connected to the computer (5) and the plant control system, which exchanges control data (2) with the technical components (3) of the plant (4) via a network (7), wherein the interface (6) records the control data (2) and transmits it to the computer (5), and wherein the computer system (1) is configured to classify a data point by means of natural language processing, to assign it to a technical component (3) and / or to create a digital control image of the plant control system.
2. Computer system according to claim 1, wherein the computer system (1) is configured to assign different designations for a technical component (3) by means of natural language processing based on language models.
3. Computer system according to the preceding claim, wherein the computer system (1) is configured to train the language models with example data.
4. Computer system according to the preceding claim, wherein the computer system (1) is configured to train and / or refine the language models with preselected data points.
5. Computer system according to one of the preceding claims, wherein the computer system (1) is configured to create a monitoring application from the control and / or information model (8).
6. Computer system according to one of the preceding claims, wherein the computer system (1) is configured to create a visualization of the control and / or information model (8).
7. Computer system according to the preceding claim, wherein the computer system (1) is configured to create one or more dashboards.
8. Method for processing control data (2) of a system (4) with technical components (3), wherein a technical component (3) represents a data point with respect to the control data (2), comprising the following steps: - Acquiring the control data (2) by means of an interface (6), sending the control data (2) to a computer (5), - Classifying a data point using Natural Language Processing, - Assignment of the data point to a technical component (3) and - Creating a digital control and / or information model (8) which represents a digital control image of the plant control.
9. Method according to the preceding claim, further comprising the steps - Interpretation of different names for technical components (3) using natural language processing based on language models, - Training the aforementioned models with texts and - Training and / or refining the aforementioned models on preselected data points.
10. Method according to one of the two preceding claims, further comprising creating a monitoring application from the control and / or information model (8).
11. Method according to one of the three preceding claims, further comprising a visualization of the control and / or information model (8) and / or the creation of one or more dashboards.
Citation Information
Patent Citations
Systems and methods for classifying audio into broad phoneme classes
US20040230432A1
Language Identification
US20160092427A1
Device-directed utterance detection
US20210295833A1
