System and method for digital reflex network (DRN) based control

Abstract

A digital reflex neural network (DRN)-based control system according to the present disclosure comprises: a sensor for measuring the state of a device to be analyzed; a digital reflex neural network slave connected to the sensor; a digital reflex neural network master connected to the digital reflex neural network slave; one or more actuators connected to the digital reflex neural network master; a semantic adapter connected to the digital reflex neural network master; and a management server connected to the semantic adapter. If one or more sensor signals correspond to a logic rule of a predefined format, the digital reflex neural network master may not receive a control command generated by the management server from the semantic adapter, but may independently generate and output a control command.

Description

Technology Field

[0001] The present disclosure relates to a digital reflex neural network (DRN)-based control system and method. It belongs to the technical field of controlling manufacturing machinery. Background Technology

[0003] Recently, industrial sites have been striving to implement Digital Twin technology to ensure the operational efficiency and safety of facilities. In particular, Digital Retrofitting is gaining attention for enhancing performance and efficiency by adding new technologies, such as sensors, networks, software, and artificial intelligence, to older machinery and systems. Through Digital Retrofitting, the remaining lifespan of existing equipment can be extended and production quality can be enhanced. Furthermore, it can lower the barriers to entry for digital transformation and contribute to achieving carbon neutrality through resource conservation and energy savings.

[0004] In control environments utilizing digital retrofits, there is a growing need for rapid and autonomous control systems capable of responding without delay to signs of equipment anomalies. However, conventional industrial control systems are typically configured to process data analysis and complex control commands through a central server (MES, ERP, QMS, etc.). Such centralized systems inevitably cause communication latency and pose a problem in that they are unsuitable for real-time control environments, as they make immediate response to equipment failures or hazardous situations difficult.

[0005] Furthermore, relying solely on physical sensor signals acquired from sensors attached to field equipment has limitations in accurately diagnosing the condition of the equipment. Since most physical sensors can only measure external conditions such as surface temperature or vibration, they cannot identify critical state changes within the equipment or at inaccessible points—namely, critical factors like core temperature, internal stress, or deep fatigue. This lack of information acts as a fundamental problem that undermines the accuracy of control systems and the reliability of predictive maintenance.

[0006] Therefore, there is a need for technology capable of independently generating real-time control commands at the edge of factories or work sites without the intervention of a central server. Furthermore, there is a demand for technology that overcomes the measurement limitations of sensors attached to equipment at factories or work sites and utilizes information regarding the interior of the equipment or inaccessible locations for control purposes. Prior art literature

[0008] Republic of Korea Registered Patent Publication No. 10-2695216 The problem to be solved

[0009] The present disclosure aims to provide a system capable of independently generating real-time control commands at the edge of a factory or work site without the intervention of a central server.

[0010] The present disclosure aims to provide a system that can generate and modify logic rules for control in real time at a factory or work site without requiring specialized knowledge of coding.

[0011] The present disclosure aims to provide a system that overcomes the measurement limitations of sensors attached to factory or work site equipment and utilizes information about the interior of the equipment or inaccessible points for control. means of solving the problem

[0013] A digital reflex neural network (DRN)-based control system according to the present disclosure comprises: a sensor for measuring the state of a device to be analyzed; a digital reflex neural network slave connected to the sensor; a digital reflex neural network master connected to the digital reflex neural network slave; one or more actuators connected to the digital reflex neural network master; a semantic adapter connected to the digital reflex neural network master; and a management server connected to the semantic adapter. When one or more sensor signals correspond to a logic rule of a predefined format, the digital reflex neural network master may not receive a control command generated by the management server from the semantic adapter, but may independently generate and output the control command.

[0014] According to one embodiment, the logic rule of the predefined format may include one or more sensor signal inputs; a conditional expression set based on the one or more sensor signal inputs; and one or more control command outputs that are output when the conditional expression is satisfied.

[0015] According to one embodiment, the digital reflex neural network master may include a logic editor that generates a logic rule of the predefined format based on user input.

[0016] According to one embodiment, the logic editor includes a Smaller Larger Language Model (sLLM), and the sLLM can receive natural language or a fishbone diagram as input and generate logic rules of the predefined format.

[0017] According to one embodiment, the digital reflex neural network master includes a lightweight FEM (Finite Element Method) engine that generates a cyber sensor signal based on a physical sensor signal obtained from the digital reflex neural network slave, and can generate the control command by inputting the physical sensor signal and the cyber sensor signal into a logic rule of a predefined format.

[0018] According to one embodiment, a pre-trained surrogate model that generates a cyber sensor signal based on a physical sensor signal acquired from the digital reflex neural network slave is included, and the control command can be generated by inputting the physical sensor signal and the cyber sensor signal into a logic rule of the pre-defined format.

[0019] According to one embodiment, if a first control command generated based on a first logic rule and a second control command generated based on a second logic rule require different actions from the same actuator, the digital reflex neural network master can transmit a control command according to a predefined order to the same actuator. Effects of the invention

[0021] According to one embodiment, a system can be provided that can independently generate real-time control commands at the edge of a factory or work site without the intervention of a central server.

[0022] According to one embodiment, a system can be provided that allows logic rules for control to be generated and changed in real time at a factory or work site without professional knowledge of coding.

[0023] According to one embodiment, a system can be provided that overcomes the measurement limitations of sensors attached to factory or work site equipment and utilizes information about the interior of the equipment or inaccessible points for control.

[0024] Meanwhile, the effects according to the embodiments are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the description below. Brief explanation of the drawing

[0026] FIG. 1 is an example diagram of a system configuration according to one embodiment. FIG. 2 is a diagram illustrating a digital reflex neural network (DRN)-based control operation according to one embodiment. FIG. 3 is a diagram illustrating a logic editor and logic rules according to one embodiment. FIG. 4 is a diagram illustrating a logic editor built-in sLLM (smaller Large Language Model) according to one embodiment. FIGS. 5a to 5e are drawings for explaining an FEM engine according to one embodiment. FIG. 6 is a flowchart illustrating a surrogate model learning operation according to one embodiment. FIG. 7 is an example diagram of the configuration of a device according to one embodiment. Specific details for implementing the invention

[0027] In the following, embodiments are described in detail with reference to the attached drawings. However, since various modifications may be made to the embodiments, the scope of the patent application is not limited or restricted by these embodiments. It should be understood that all modifications, equivalents, and substitutions to the embodiments are included within the scope of the rights.

[0028] Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be modified and implemented in various forms. Accordingly, the embodiments are not limited to the specific disclosed forms, and the scope of this specification may include modifications, equivalents, or substitutions that fall within the technical concept.

[0029] Terms such as "first" or "second" may be used to describe various components, but these terms should be interpreted solely for the purpose of distinguishing one component from another. For example, the first component may be named the second component, and similarly, the second component may be named the first component.

[0030] When it is stated that a component is "connected" to another component, it should be understood that it may be directly connected to or coupled with that other component, or that there may be other components in between.

[0031] The terms used in the embodiments are for illustrative purposes only and should not be interpreted as intended to be limiting. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "comprising" or "having" are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0032] Terms such as “…part,” “…module,” etc., as used in this specification refer to a unit that processes at least one function or operation, which may be implemented in hardware or software, or a combination of hardware and software. In this specification, embodiments may be represented by functional block configurations and various processing steps. These functional blocks may be implemented by various numbers of hardware and / or software configurations that execute specific functions. For example, integrated circuit configurations such as memory, processing, logic, and look-up tables may be employed, which can execute various functions under the control of one or more microprocessors or other control devices.

[0033] Spatially relative terms such as "below," "beneath," "lower," "above," and "upper" may be used to facilitate the description of the relationship between one component and other components as illustrated in the drawings. Spatially relative terms should be understood as encompassing different orientations of components during use or operation, in addition to the orientations depicted in the drawings. For example, if a component depicted in a drawing is inverted, a component described as "below" or "beneath" of another component may be placed "above" of that component. Therefore, the exemplary term "below" may encompass both the lower and upper directions. Components may also be oriented in other directions, and accordingly, spatially relative terms may be interpreted according to the orientation.

[0034] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the embodiments pertain. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0035] In addition, when describing with reference to the attached drawings, identical components are assigned the same reference numeral regardless of drawing symbols, and redundant descriptions thereof are omitted. In describing the embodiments, if it is determined that a detailed description of related prior art could unnecessarily obscure the essence of the embodiments, such detailed description is omitted.

[0037] FIG. 1 is an example diagram of a system configuration according to one embodiment.

[0038] A system according to one embodiment may include an analysis target device (101, 104), a sensor (111-119), a digital reflex neural network (DRN) slave (121-124), a digital reflex neural network (DRN) master (131, 132), an actuator (161, 162), a semantic adapter (141), and a management server (151).

[0039] In this specification, a Digital Reflex Network (DRN) refers to a system that includes a process in which a Digital Reflex Network (DRN) master (131, 132) does not receive a control command from a management server (151), but instead directly generates a control command and transmits it to an actuator (161, 1622). In this specification, the Digital Reflex Network (DRN) may be referred to simply as "Digital Reflex Network" without the abbreviation, or simply as the abbreviation DRN.

[0041] [System Components]

[0042] The device to be analyzed (101, 104) may be a manufacturing or work equipment or part thereof located at a factory or work site, but is not limited thereto. The device to be analyzed (101, 104) only needs to be configured to be able to attach a sensor (111-119) and is not required to be state-of-the-art equipment.

[0043] The sensor (111-119) may be configured to measure and collect mechanical, electrical, chemical, and operational information of the device to be analyzed (101, 104). The sensor (111-119) may measure various types of analog and digital data such as temperature, pressure, vibration, rotation, speed, acceleration, noise, sound waves, ultrasound, current, voltage, wavelength, frequency, humidity, viscosity, flow rate, flow velocity, and residual amount of the device to be analyzed (101, 104). Additionally, the sensor (111-119) may include a configuration for detecting specific states or operations, such as a failure state or operation status of the device to be analyzed (101, 104). The sensor (111-119) may be connected to transmit the measured information to a DRN slave (121-124) via an electrical signal or a communication protocol.

[0044] The DRN slaves (121-124) can collect and process data transmitted from the sensors (111-119). The DRN slaves (121-124) can perform pre-processing tasks such as refining analog and digital signals received from the sensors, converting communication protocols, or standardizing data formats. Additionally, the DRN slaves (121-124) can group or unify data collected from multiple sensors and transmit it to the DRN masters (131, 132).

[0045] The DRN master (131, 132) can perform control and communication based on data received from the DRN slaves (121-124). The DRN master (131, 132) can perform analysis on the received data to immediately detect abnormal signs or specific state changes of the device to be analyzed (101, 104). In particular, the DRN master (131, 132) can generate control commands independently without the intervention of the management server (151) and transmit them to the actuator (161, 162) to rapidly control the equipment in real time. Meanwhile, the DRN master (131, 132) can perform an upper communication gateway that transmits collected and processed sensor signals to the semantic adapter (141) or the management server (151).

[0046] The DRN master (131, 132) may include at least one processor. The DRN master (131, 132) may not be a single physical device and may be composed of multiple physical devices. Each device unit of the DRN master (131, 132) may be configured to perform all or part of the computational functions, storage / retrieval functions, input / output functions, and control functions that a conventional computer has.

[0047] The DRN master (131, 132) may include a logic editor. The DRN master (131) may generate control commands based on logic rules of a predefined format generated through the logic editor. The logic rules of a predefined format may include one or more sensor signal inputs, conditional expressions, and one or more control command outputs. Specific embodiments of the logic editor will be described later with reference to FIGS. 3 and 4.

[0048] The DRN master (131, 132) may include a lightweight FEM (Finite Element Method) engine and / or a surrogate model. The lightweight FEM engine and / or surrogate model may generate cyber sensor signals based on physical sensor signals obtained from DRN slaves (121-124). The cyber sensor signals may be used as inputs to logic rules of a predefined format. Specific embodiments of the lightweight FEM engine will be described later with reference to FIG. 5. Specific embodiments of the surrogate model will be described later with reference to FIG. 6.

[0049] The DRN master (131, 132) has lower computational power compared to the management server (151), but may have a hardware configuration optimized for lightweight computation. The DRN master (131, 132) may be deployed at the edge of the factory site. By performing lightweight computation and being located at the site, the DRN master (131, 132) can perform control quickly by omitting communication delays with the management server (151) and computations by the management server (151). Additionally, by deploying the DRN master (131, 132) at the factory site, it can provide convenience for field workers to easily change logic rules of a predefined format in real time.

[0050] The actuator (161, 162) may include a configuration that physically changes the state or operation of the device to be analyzed (101, 104) according to a control command received from the DRN master (131, 132). Additionally, the actuator (161, 162) may include a configuration that performs more extensive control, such as generating an alarm signal to notify an administrator of an abnormal condition or cutting off the power to some or all of the factory equipment.

[0051] The semantic adapter (141) can serve as an interface between the DRN masters (131, 132) and the management server (151). The semantic adapter (141) converts data collected from the DRN masters (131, 132) into a unified format or semantic and transmits it to the management server (151). Conversely, it can convert control commands received from the management server (151) to fit the individual format or unit required by each DRN master (131, 132) and transmit them. Through such conversion and unification operations, the semantic adapter (141) supports the consistent processing of various data units and communication protocols within the system from the perspective of the management server (151).

[0052] The management server (151) may include at least one processor. The management server (151) may not be a single physical device and may be composed of multiple physical devices. Each device unit of the management server (151) may be configured to perform all or part of the computational functions, storage / retrieval functions, input / output functions, and control functions that a conventional computer has.

[0053] The management server (151) may include a conventional Manufacturing Execution System (MES), Enterprise Resource Planning (ERP), Supply Chain Management (SCM), Quality Management System (QMS), Energy Management System (EMS), and Facility Energy Management System (FEMS) for managing manufacturing facilities.

[0054] The management server (151) can be configured to have superior computational capabilities compared to the DRN master (131, 132). Through this, the management server (151) can efficiently perform large-scale or complex computations that are difficult for the DRN master to handle.

[0055] The management server (151) may include a central control model. The central control model can generate subsequent control commands for sensor signals that individual DRN masters cannot process. Additionally, the central control model can interpret data flowing in from multiple DRN masters in a complex manner to generate integrated control commands from the perspective of overall facility operation that are beyond the processing range of individual DRN masters. The management server (151) can transmit the control commands generated by the central control model to the DRN masters (131, 132) through the semantic adapter (141).

[0056] The management server (151) may include a non-lightweight precision FEM engine. The non-lightweight precision FEM engine can set the mesh resolution to a high resolution exceeding the level of hundreds of nodes. In addition, precise physical analysis can be achieved by utilizing a non-linear model-based calculation method. Furthermore, boundary conditions can also be set in detail and in various ways to provide more in-depth analysis.

[0058] [Connection relationships between system components]

[0059] The connection relationships between system components according to one embodiment may be as follows. Specific implementation aspects may be adopted differently depending on the embodiment.

[0060] There may be one or more devices to be analyzed (101, 104). The first device to be analyzed (101) may be a motor brake of an overhead crane. The motor brake may have a motor part (102) and a brake part (103). The second device to be analyzed (104) may be a rolling facility of an overhead crane.

[0061] One or more sensors may measure individual devices to be analyzed. The first sensor (111) may be a vibration sensor that measures the vibration of the motor unit (102). The second sensor (112) may be a temperature sensor that measures the temperature of the motor unit (102). The third sensor (113) may be a vibration sensor that measures the vibration of the brake unit (103). The fourth sensor (114) may be a temperature sensor that measures the temperature of the brake unit (103). The fifth sensor (115) may be an acoustic sensor that measures the sound (noise) of the brake unit (103). The sixth sensor (116) may be a current sensor that measures the current of the motor brake, which is the first device to be analyzed (101). The seventh sensor (117) may be an oil sensor that measures the oil condition (viscosity, dielectric constant, conductivity, density, temperature, etc.) of the motor brake, which is the first device to be analyzed (101). The eighth sensor (118) may be a temperature sensor that measures the heat generated at the friction contact part of the rolling equipment, which is the second device to be analyzed (104). The ninth sensor (119) may be a vibration sensor that measures the vibration of the second device to be analyzed (104).

[0062] Multiple sensors may be connected to a single DRN slave. The first to fifth sensors (111-115) may be connected to the first DRN slave (121). The first to fifth sensors (111-115) may transmit a measurement signal for the first device to be analyzed (101) to the first DRN slave (121). The sixth sensor (116) may be connected to the second DRN slave (122). The sixth sensor (116) may transmit a measurement signal for the first device to be analyzed (101) to the second DRN slave (122). The seventh sensor (117) may be connected to the third DRN slave (123). The seventh sensor (117) may transmit a measurement signal for the first device to be analyzed (101) to the third DRN slave (123). The 8th and 9th sensors (118, 119) can be connected to the 4th DRN (124). The 8th and 9th sensors (118, 119) can transmit measurement signals for the 2nd analysis target device (104) to the 4th DRN slave (124).

[0063] Multiple DRN slaves can be connected to a single DRN master. The first to third DRN slaves (121-123) can be connected to the first DRN master (131). The first to third DRN slaves (121-123) can transmit preprocessed sensor signals for the first analysis target device (101) to the first DRN master (131). The fourth DRN slave (124) can be connected to the second DRN master (132). The fourth DRN slave (124) can transmit preprocessed sensor signals for the second analysis target device (104) to the second DRN master (132).

[0064] One or more actuators may be connected to a single DRN master. A first actuator (161) may be connected to a first DRN master (131). A second actuator (161) may be connected to a second DRN master (132). The actuators (161, 162) may be, but are not limited to, an LED (warning light), an alarm (warning sound), a shut-off circuit, an emergency power cut-off device, a motor speed / torque control device, a valve actuator, a pump on / off switch, a cooling fan drive device, or a motion control unit of a robot arm.

[0065] Sensors (111-119), DRN slaves (121-124), DRN masters (131, 132), and actuators (161, 162) can be placed at the edge of the factory site where the device to be analyzed (101, 104) is located. In particular, by having the DRN masters (131, 132) located at the factory site, communication delays with the management server (151) and calculations by the management server (151) can be omitted, and control can be performed quickly.

[0066] Multiple DRN masters can be connected to a single semantic adapter. The first and second DRN masters (131-132) can be connected to the semantic adapter (141). The semantic adapter (141) can be connected to a management server (151).

[0068] FIG. 2 is a diagram illustrating a digital reflex neural network (DRN)-based control operation according to one embodiment.

[0069] In the following, for convenience of explanation, the first analysis target device (101), first sensor (111), first DRN slave (121), first DRN master (131), and first actuator (161) of FIG. 1 are each abbreviated as analysis target device (101), sensor (111), DRN slave (121), DRN master (131), and actuator (161).

[0071] [Fig. 2 Dotted line]

[0072] The dotted line in FIG. 2 shows the process of transmitting a sensor signal from the sensor (111) to the DRN master (131). The sensor (111) can generate a signal by measuring the device to be analyzed (101) (S200). The sensor (111) can transmit the measurement signal to the DRN slave (121) (S210). The DRN slave (121) can preprocess the received sensor signal (S220). This process may include noise removal, data format standardization, or data compression. The DRN slave (121) can transmit the preprocessed sensor output to the DRN master (131) (S230).

[0073] The DRN master (131) can determine whether one or more sensor signals correspond to a logic rule of a predefined format (S240). Depending on the determination result, the control process can be branched into a solid line path or a dotted line path.

[0075] [Fig. 2 Solid Line]

[0076] If the DRN master (131) determines that one or more sensor signals correspond to a logic rule of a predefined format (S240, Yes), the DRN master (131) can independently generate a control command without the intervention of the semantic adapter (141) or the management server (151) (S250).

[0077] The DRN master (131) can generate control commands based on logic rules of a predefined format. A logic rule of a predefined format may include one or more sensor signal inputs, conditional expressions, and one or more control command outputs. A logic rule of a predefined format may be generated through a logic editor of the DRN master (131). Specific embodiments of the logic editor will be described later with reference to FIGS. 3 and 4.

[0078] Additionally, the DRN master (131) may include a lightweight FEM engine and / or a surrogate model. The lightweight FEM engine and / or surrogate model may generate cyber sensor signals based on physical sensor signals obtained from the DRN slave (121). The cyber sensor signals may be used as inputs to logic rules of a predefined format. Through this, more accurate and sophisticated control commands can be generated by comprehensively considering the internal state (e.g., internal temperature) of the device under analysis, which is difficult to determine solely by physical sensors (e.g., surface temperature). Specific embodiments of the lightweight FEM engine will be described later with reference to FIG. 5. Specific embodiments of the surrogate model will be described later with reference to FIG. 6.

[0079] The DRN master (131) can output a control command generated independently to the actuator (161) (S270). The actuator (S161) can promptly execute direct measures such as stopping the operation of the equipment, opening and closing a specific valve, or activating a warning light or alarm according to the control command generated by the DRN master (131), but is not limited thereto.

[0080] Through this, the DRN master (131) can eliminate communication delays with the management server (151) and respond quickly by performing immediate control on abnormal conditions of the device (101) to be analyzed. Preferably, even if the DRN master (131) processes control commands more slowly than the management server (151), the process according to the solid line of FIG. 2, which does not use the management server (151), can be configured to be faster than the process according to the dotted line of FIG. 2, which uses the management server (151). More preferably, the DRN master (131) can be configured to process control commands faster than the management server (151). Through this, the DRN master (131) can provide the advantage of being able to quickly control the site through independent control without using the management server (151).

[0082] [Fig. 2 Dotted line]

[0083] If the DRN master (131) determines that one or more sensor signals do not conform to a logic rule of a predefined format (S240, No), control can be performed via an integrated path through the management server (151).

[0084] The DRN master (131) can transmit data including feature data and reflection logic results to the semantic adapter (141) (S261). The feature data may be values ​​processed by the DRN master (131) for the purpose of analysis of physical sensor values, which are waveform signals. For example, if the physical sensor is a vibration sensor, the feature data may include the Crest Factor, RMS value, FFT peak frequency, etc., but is not limited thereto. The reflection logic results may include control information performed by the DRN master (131) itself. The semantic adapter (141) can preprocess the feature data and reflection logic results (S262). The semantic adapter (141) can perform a preprocessing operation to convert the feature data and reflection logic results received from the DRN master (131) into a unified meaning and format required by the management server (151). Afterward, the semantic adapter (141) can transmit the preprocessed feature data and reflection logic results to the management server (151) (S263). The management server (151) can make integrated or complex control decisions by analyzing feature data and reflection logic results using its own central control model (S264).

[0085] The management server (151) can transmit control commands to the semantic adapter (141) (S265). The semantic adapter (141) can perform preprocessing to convert the commands into a format and control unit that the DRN master (131) can interpret (S266). The semantic adapter (141) can transmit the converted control commands to the DRN master (131) (S267). The DRN master (131) can output control commands to the actuator (161) (S270). The actuator (S161) can execute complex measures such as cutting off the equipment power, changing the operating speed, or reporting status information to a remote control system, but is not limited thereto.

[0087] FIG. 3 is a diagram illustrating a logic editor and logic rules according to one embodiment.

[0088] The DRN master (131) may include a logic editor (300). The logic editor (300) may generate logic rules (310) of a predefined format based on user input. The DRN master (131) may generate control commands based on the logic rules (310) of a predefined format generated through the logic editor (300).

[0089] The logic editor (300) may include a GUI (Graphic User Interface) that allows a user to combine logic blocks to create a logic rule (310) of a predefined format. The logic blocks may include a sensor signal input (311), a conditional expression (312), and a control command output (313).

[0090] A logic rule (310) of a predefined format may include one or more sensor signal inputs (311). The sensor signal inputs (311) may include physical sensor signals corresponding to the sensors (111-119) of FIG. 1. Additionally, the sensor signal inputs (311) may include cyber sensor signals generated through a lightweight FEM engine or surrogate model described later.

[0091] A logic rule (310) of a predefined format may include a conditional expression (312). The conditional expression (312) may set one or more sensor signal inputs (311) as a combination of AND conditions, OR conditions, and NOT conditions. Additionally, the conditional expression (312) may be configured to determine that a specific condition is valid only when it persists for a specific period of time or is repeated more than a specific number of times. When the conditional expression is satisfied, a control command output (313) may be produced.

[0092] A logic rule (310) of a predefined format may include one or more control command outputs (313). The control command outputs (313) may include the type of actuator and the operation to be performed per actuator, etc.

[0093] Meanwhile, if multiple logic rules are executed simultaneously and individual control commands conflict with each other, the DRN master (131) may output only the control commands according to a predefined priority. For example, a first control command generated based on a first logic rule and a second control command generated based on a second logic rule may require different operations from the same actuator. In this case, the DRN master (131) may prevent operation conflicts by transmitting only the control commands according to the predefined priority to the corresponding actuator. The predefined priority may be set to a high-risk priority established based on field experience, but the method of setting the priority is not limited to this.

[0094] Through the above, the logic editor (300) supports the formalization of tacit knowledge (empirical knowledge) at the site and provides convenience and immediacy by allowing the site manager to directly set logic rules based on No-Code. In addition, when control commands conflict, by applying commands according to a predefined priority, it ensures that high-risk logic rules are executed first and can avoid behavioral conflicts of the actuators.

[0096] FIG. 4 is a diagram illustrating a logic editor built-in sLLM (smaller Large Language Model) according to one embodiment.

[0097] The logic editor (400) can include sLLM to enhance user convenience. Based on user input, sLLM can generate logic rules (402) of a predefined format that the DRN master (131) can utilize without having to set up logic blocks one by one.

[0098] sLLM can receive and interpret user input (401) in the form of natural language or fishbone diagrams. Natural language input can be in a user-friendly form, such as "stop the machine if oil pressure is less than 2.0 bar or vibration exceeds 15," and can be entered in the form of typing or speech. Fishbone diagrams can be configured to visually structure the cause and effect of a problem. sLLM can analyze user input and automatically generate logic rules (402) of a predefined format.

[0099] For example, in a situation where a sensor is attached to a press process, if a user inputs a command in natural language such as "If vibration is 12 mm / s or more and oil pressure is less than 2 bar, execute a shutdown command, server log, and field light within 0.03 seconds," sLLM can convert this into a logic rule of a predefined format such as 'INPUT(Vibration, Oil Pressure) IF (Vibration ≥ 12 AND Oil Pressure < 2) THEN (STOP Machine, LED On, MQTT Log)'.

[0100] Through this, sLLM can convert human-understood knowledge-based requests into logic rules that the DRN master (131) understands. This allows field workers to easily set the logic rules necessary to implement a digital reflex neural network without coding knowledge or the use of the GUI of the logic editor (300).

[0102] FIGS. 5a to 5e are drawings for explaining an FEM engine according to one embodiment.

[0103] FIG. 5a is a diagram illustrating the processing operation of a non-lightweight precision FEM engine (510) and a lightweight FEM engine (500) in comparison, FIG. 5b shows a structure (501) to be analyzed by a lightweight FEM engine (500) according to one embodiment, FIG. 5c shows an element model (502) according to one embodiment, FIG. 5d shows a lightweight model (503) according to one embodiment, and FIG. 5e shows the analysis result (506) of a lightweight FEM engine (500) according to one embodiment.

[0105] [Non-lightweight precision FEM engine (510)]

[0106] According to one embodiment, the management server (151) may include a non-lightweight precision FEM engine (510). The non-lightweight precision FEM engine (510) can perform high-precision finite element analysis on the device to be analyzed using a high-resolution mesh and complex boundary conditions.

[0107] Steps S511 through S16 represent a preprocessing process for data to be input into a non-lightweight precision FEM engine (510). To perform non-lightweight precision FEM, an actual object or structure is defined (S511), and the components constituting the actual object or structure can be idealized and modeled as analysis elements such as beams, shells, and solid elements (S512). Subsequently, the idealized structure is geometrically modeled in a 2D or 3D CAD environment (S513), and the geometric model can be converted into a finite element mesh model composed of multiple elements and nodes (S514).

[0108] Here, mesh resolution refers to the total number of nodes when the finite elements of the device to be analyzed are modeled as a graph data structure for finite element analysis. The lower the mesh resolution, that is, the fewer the number of nodes, the less computation is required for finite element analysis. The mesh resolution of the non-lightweight precision FEM engine (510) can be set to a high resolution exceeding the level of several hundred nodes. Preferably, the mesh resolution of the non-lightweight precision FEM engine (510) can be set to more than 1,000 nodes.

[0109] Next, material properties required for analysis, such as the elastic modulus, thermal conductivity, and specific heat of the material, are input (S515), and support conditions such as fixed, hinged, and slider components, as well as boundary conditions such as load, temperature, and pressure, can be applied (S516). Boundary and load conditions are manually set by the user considering the purpose of the analysis and experimental conditions, and user intervention may be required when changing the analysis conditions. Through the above preprocessing steps, a precise FEM model of the analysis target can be constructed.

[0110] In the numerical calculation step, the stiffness matrix and mass matrix for each element are generated and assembled to form the overall matrix equation, and then the displacement, strain, stress, temperature distribution, etc. can be calculated using a numerical analysis algorithm (direct solution, iterative solution, etc.) (S517). In the post-processing step, the calculated results can be processed to visualize them in the form of contour lines, curve graphs, animations, etc., and provided for review by an analyst (S518).

[0111] Meanwhile, if the user determines that the modeling assumptions or boundary conditions are inappropriate by referring to the visualized analysis results as needed, the user may manually return to one or more of steps S512 to S516 to modify the model and re-execute the analysis (S519). As such, the non-lightweight precision FEM engine (510) can handle high degrees of freedom and complex boundary conditions, but since it requires manual operation by the user and requires a lot of computational resources and time, it is desirable to operate it in a batch or near-real-time manner on a high-performance device such as a management server (151).

[0113] [Lightweight FEM Engine (500)]

[0114] According to one embodiment, the DRN master (131) may include a lightweight FEM engine (500). The lightweight FEM engine (500) may generate a cyber sensor signal (504) based on a physical sensor signal (510) obtained from a DRN slave (121-124). The cyber sensor signal (504) may include mechanical, electrical, chemical state and operation information of the device to be analyzed that cannot be measured by a physical sensor. For example, the cyber sensor signal (504) may be the core temperature of the device to be analyzed generated based on the surface temperature of the device to be analyzed measured by a physical sensor, but is not limited thereto.

[0115] The lightweight FEM engine (500) can improve the ability to analyze the state and operation of the device under analysis when performing real-time control on the DRN master (131) without the intervention of the management server (151). The main targets for analysis are mechanical, electrical, and chemical states and operations that cannot be measured by physical sensors, and may include strain, stress, temperature, vibration, etc., which indicate physical defects and abnormal signs, but are not limited thereto. By generating cyber sensor signals through the lightweight FEM engine (500) and constructing logic rules that use physical sensors and cyber sensors in combination, the accuracy of predicting the state of the device under analysis can be improved.

[0116] When the lightweight FEM engine (500) uses a mesh-based model, the mesh resolution can be set to a low resolution of tens to hundreds of nodes to minimize computational burden and reflect only characteristic physical quantities. Additionally, a high-speed analysis suitable for real-time can be achieved by utilizing an iterative calculation method based on a linear model. Simplified constraint and load conditions can be used for boundary conditions. Through this, even on a DRN master (131) that is lighter than a management server (151), the physical sensor values ​​can be mapped to boundary conditions in real-time and analysis can be performed within a few milliseconds to hundreds of milliseconds to output cyber sensor values.

[0117] The specific calculation procedure for the lightweight FEM engine (500) can be performed through steps S501 to S507. First, the actual structure of the device to be analyzed can be defined (S501).

[0118] The structure of the device to be analyzed according to one embodiment may be as shown in FIG. 5b, but is not limited thereto. Referring to FIG. 5b, the lightweight FEM engine (500) may set a planar frame structure consisting of three beam elements (①, ②, ③) and four nodes (1 to 4) as the target for analysis. At this time, the length, cross-sectional area (A), moment of inertia (I), and elastic modulus (E) of each element are defined in advance, and a situation may be defined in which a horizontal load of 40 kN is applied to node 2 as an external force and a moment of 500 N·m acts on node 3. Such an example may be an idealized model of the structure of the housing, support bracket, motor frame, etc. to which the actual DRN master (131) is attached.

[0119] According to one embodiment, the cyber sensor values ​​to be output by the lightweight FEM engine (500) may be the displacement and rotation (u2, v2, phi2, u3, v3, phi3) of nodes 2 and 3, and the reaction force and anti-moment (F1x, F1y, M1, F4x, F4y, M4) of nodes 1 and 4. These values ​​are reaction forces acting inside the structure or on the support, and can be assumed to be in a situation where it is difficult or economical to directly install and measure them with a physical sensor.

[0120] Next, the lightweight FEM engine (500) can select a predefined element model (S502). Element stiffness matrices, mode reduction models, surrogate models, etc., that are pre-calculated during the pre-design process may be stored in a library format. The lightweight FEM engine (500) can select a model corresponding to the structure to be analyzed from the library.

[0121] An element model according to one embodiment may be as shown in FIG. 5c, but is not limited thereto. Referring to FIG. 5c, for a beam element ①, the relationship between a displacement vector consisting of displacements and rotations (u1, v1, phi1, u2, v2, phi2) of nodes 1 and 2 and a load vector consisting of corresponding reactions and anti-moments (F1x, F1y, M1, F2x, F2y, M2) can be expressed by a stiffness matrix k. The stiffness matrix k can be expressed as a 6×6 matrix consisting of a scaling factor such as E / L and factors dependent on the cross-sectional area A and the second moment of area I, and such a generalized stiffness matrix template may be stored in advance and selected.

[0122] Next, the lightweight FEM engine (500) can generate a lightweight model based on a selected library model (S503). For example, after calculating the overall stiffness matrix based on the stiffness matrix of each element, the model dimension can be reduced to retain only the necessary degrees of freedom from the overall stiffness matrix, or a low-dimensional model can be constructed by selecting only the nodes corresponding to the location to be analyzed. The lightweight model generated in this way has significantly fewer degrees of freedom compared to the high-precision model, making it suitable for real-time analysis.

[0123] Additionally, when using a mesh-based model, the mesh resolution can be set to a low resolution of tens to hundreds of nodes to minimize computational burden and reflect only characteristic physical quantities. Preferably, the mesh resolution of the lightweight FEM engine (500) can be set to 1,000 nodes or less. In particular, when the mesh resolution of the lightweight FEM engine (500) is set to 1,000 nodes or less and the mesh resolution of the non-lightweight precision FEM engine (510) of the management server (151) is set to exceed 1,000 nodes, the independent control process of the DRN master (131) according to the solid line of FIG. 2 can obtain a significant advantage in computational speed compared to the integrated control process passing through the management server (151) according to the dotted line of FIG. 2.

[0124] Model lightweighting according to one embodiment may be as shown in FIG. 5d, but is not limited thereto. Referring to FIG. 5d, the overall stiffness matrix for all degrees of freedom of nodes 1, 2, 3, and 4 can be formed by assembling the element stiffness matrices corresponding to beams ①, ②, and ③. At this time, utilizing the fact that only the displacement and rotation values ​​of nodes 2 and 3 are to be known, the model can be lightweighted to retain only the submatrix corresponding to the corresponding degrees of freedom, as indicated by the dotted line in FIG. 5d. To this end, reduction techniques such as static condensation and Guyan condensation can be used, and if necessary, the order of rows and columns can be permuted to place the degrees of freedom of interest at the top of the matrix to perform the reduction operation. Through this, a reduced stiffness matrix defined only for the degrees of freedom of the nodes of interest can be generated. The reduced stiffness matrix can be used in subsequent numerical analysis steps.

[0125] In the following sequence, the lightweight FEM engine (500) can receive boundary conditions and loads (S504). In particular, unlike a conventional FEM engine, the boundary conditions and loads can change in real time according to the measurements of the physical sensor. Specifically, the physical sensor signal is transmitted in real time to the DRN master (131) via the DRN slave, and the lightweight FEM engine (500) can reset boundary conditions such as support point displacement, external force, moment, and temperature gradient in real time based on the measurements of the physical sensor. For example, the load acting on a specific node can be inversely calculated from the torque or thrust measured by the sensor, and the boundary conditions can be configured such that the fixed end displacement is set to 0 and the free end displacement is set to an unknown.

[0126] In the embodiment with reference to FIGS. 5b to 5d, a physical sensor for measuring horizontal force may be installed at node 2, and a physical sensor for measuring moment may be installed at node 3. In actual operating conditions, horizontal loads and moments different from the 40 kN and 500 N·m assumed in the design may act on node 2 and node 3, and these real-time measured values ​​may be continuously input as boundary conditions to the lightweight FEM engine (500).

[0127] Next, the lightweight FEM engine (500) can perform matrix generation and numerical analysis (S505). The lightweight FEM engine (500) can construct an overall stiffness matrix [K] for the degrees of freedom of interest as shown in FIG. 5d by transforming and assembling the reduced element stiffness matrix into a global coordinate system. The overall stiffness matrix [K] can be expressed as a symmetric matrix corresponding to the degrees of freedom of the selected node (e.g., u2, v2, phi2, u3, v3, phi3, etc.), and the lightweight FEM engine (500) can calculate the displacement of each degree of freedom by solving a set of linear equations of the form [K]{u}={F}. At this time, since the number of degrees of freedom of interest and the matrix structure are fixed in advance and only the values ​​input from the physical sensor change over time, the amount of computation can be minimized by using techniques such as LU decomposition, pre-calculation of the inverse matrix, and reuse of initial values ​​in the iterative solution.

[0128] Next, the lightweight FEM engine (500) can perform matrix generation and numerical analysis (S505). The lightweight FEM engine (500) can construct a reduced total stiffness matrix defined only for the degrees of freedom of interest (e.g., u2, v2, phi2, u3, v3, phi3) as illustrated by the dashed line in FIG. 5d by converting and assembling the reduced element stiffness matrix into the global coordinate system. The reduced total stiffness matrix is ​​expressed as a symmetric matrix corresponding to the degrees of freedom of the selected node, and the lightweight FEM engine (500) can calculate the degrees of freedom of interest by solving a set of linear equations. Since the number of degrees of freedom of interest and the matrix structure are fixed in advance, and only the boundary conditions and load vectors input from the physical sensor change over time, the amount of computation can be minimized by using techniques such as LU decomposition, pre-calculation of the inverse matrix, and reuse of initial values ​​in the iterative solution. If necessary, the removed degrees of freedom can be approximately restored from the degrees of freedom of interest according to the relationship used in the reduction process.

[0129] Next, the lightweight FEM engine (500) can output an analysis result (S506). The analysis result can be used as a cyber sensor signal, and can be a value that calculates the mechanical, electrical, chemical state and operation of the device being analyzed that cannot be measured by a physical sensor, or a value that compensates for the measurement error of the physical sensor.

[0130] The analysis results according to one embodiment may be as shown in 5e, but are not limited thereto. Referring to FIG. 5e, the displacement (u2, v2, phi2, u3, v3, phi3) and support point reaction force (F1x, F1y, M1, F4x, F4y, M4) of specific nodes (e.g., nodes 2, 3) that are difficult to measure as physical sensors can be calculated as an analysis result. The lightweight FEM engine (500) can calculate predefined physical quantities (e.g., maximum bending stress, core temperature at a specific location, residual deformation amount, etc.) from these analysis results and convert them into one or more cyber sensor signals (504) to provide to the control logic of the DRN master (131). For example, the displacements and stresses of multiple nodes can be combined and output as a single indicator, such as a structural integrity index or thermal margin.

[0131] In addition, the lightweight FEM engine (500) can perform automatic feedback (S507). Specifically, the lightweight FEM engine (500) compares the cyber sensor signal (504) calculated in step S506 with the physical sensor signal resulting from the control of the DRN master (131), and if the difference between the model and the actual system exceeds a predetermined threshold, it can automatically correct the boundary conditions or load model or readjust the parameters of the library model. In particular, unlike a conventional FEM engine, through the automatic feedback loop between step S504 and step S506, the lightweight FEM engine (500) can secure the analysis accuracy required for real-time control while maintaining the fit of the model without manual intervention by the user.

[0132] Preferably, the DRN master (131) equipped with a lightweight FEM engine (500) can be configured so that, even if the processing time for individual control commands is slower than that of the management server (151) equipped with a non-lightweight precision FEM engine (510), the process according to the solid line of FIG. 2 that does not use the management server (151) is faster in terms of overall response speed than the entire process according to the dotted line of FIG. 2 that uses the management server (151). More preferably, the DRN master (131) equipped with a lightweight FEM engine (500) can be configured to process individual control commands faster than the non-lightweight precision FEM engine (510) running on the management server (151) by using a lightweight model and a real-time optimized numerical analysis algorithm. Through this, the DRN master (131) can provide the advantage of being able to quickly control the site through independent control without using the management server (151).

[0134] FIG. 6 is a flowchart illustrating a surrogate model learning operation according to one embodiment.

[0135] According to one embodiment, the DRN master (131) may include a pre-trained surrogate model (not shown) that generates a cyber sensor signal based on a physical sensor signal obtained from a DRN slave (121-124).

[0136] The surrogate model can be employed as a method for the DRN master (131) to generate cyber sensor signals, either together with or separately from the lightweight FEM engine (500). While the lightweight FEM engine (500) directly performs mesh-specific calculations through lightweight finite element analysis (FEM), the surrogate model operates in the form of a pre-trained surrogate model. That is, instead of directly performing complex physical calculations, the surrogate model can immediately output an inference value for the cyber sensor signal according to the corresponding condition when a physical sensor signal is input. The surrogate model can significantly simplify the calculation process and improve speed compared to the lightweight FEM engine (500), thereby enabling more immediate control decisions.

[0137] According to the embodiment, a lightweight FEM engine (500) and a surrogate model may be employed together. The DRN master (131) may generate cyber sensor signals by using a surrogate model under specific conditions and a lightweight FEM engine (500) under other conditions. For example, if the control command output occurs for less than a predetermined period or number of times, it can be interpreted as a period in which the devices under analysis operate stably. In this case, the DRN master (131) may be configured to use a surrogate model that has a low number of operations and low power consumption. On the other hand, if the control command output occurs for more than a predetermined period or number of times, it can be interpreted as a situation in which at least some of the devices under analysis are unstable. In this case, the DRN master (131) may be configured to use a lightweight FEM engine (500) for fine control.

[0138] Since the surrogate model generates cyber sensor signals based on physical sensor signals, it may use a lightweight deep learning-based regression model (e.g., DNN) or a machine learning-based regression model (e.g., support vector regression, random forest regression), or a regression model that combines these. The surrogate model can be trained using supervised learning. The surrogate model can be trained through a management server (151) or a separate training device. The surrogate model can be trained using training data in which physical sensor signals for the device to be analyzed are used as features, and target measurements (e.g., core temperature) of the device to be analyzed analyzed by a non-lightweight precision FEM engine are used as labels corresponding to the features. The output of the surrogate model, which receives the features as input, can be trained to become increasingly closer to the labels.

[0139] The following describes the process of learning a surrogate model through a learning device.

[0140] First, the learning device can acquire features and labels as training data (S610).

[0141] The feature for training the surrogate model may be a physical sensor signal for the device under analysis. According to one embodiment, the feature may include the surface temperature of the device under analysis. The label for training the surrogate model may be a target measurement of the device under analysis analyzed by a non-lightweight precision FEM engine included in the management server (151) or a separate device. According to one embodiment, the label may include the core temperature of the device under analysis.

[0142] Now, the learning device can apply features to the surrogate model (S620).

[0143] The learning device may use features as input to the surrogate model as is, or undergo a conventional process such as dividing features into smaller units to match the input format that the surrogate model can accept. For learning efficiency, the learning device may use stochastic gradient descent in the surrogate model optimization step (S650), and to this end, it may generate inputs to the surrogate model by dividing features into batches. Features may be redistributed within the batch for each training epoch.

[0144] Next, the learning device can obtain an output from the surrogate model (S630).

[0145] The output of the surrogate model may be an inferred value for a target measurement of the device under analysis. According to one embodiment, the output of the surrogate model may be an inferred value for the core temperature of the device under analysis.

[0146] Afterwards, the learning device can compare the output and the label (S640).

[0147] The process of comparing the output of a surrogate model corresponding to inference with the label corresponding to the correct answer can be accomplished by calculating a loss function. Known loss functions such as the mean squared error (MSE) and cross entropy error (CEE) may be used. However, this is not limited to these, and any loss function used in various surrogate models can be utilized as long as the deviation, error, or difference between the output of the surrogate model and the label can be measured.

[0148] Next, the learning device can optimize the surrogate model based on the comparison value (S650).

[0149] The learning device can increasingly align the output of the surrogate model corresponding to the inference with the label corresponding to the correct answer by updating the weights of the surrogate model nodes so that the value of the loss function or the sum of the loss function value and the regularization term becomes smaller. Through this, the surrogate model can be optimized to output an inference that is close to the correct answer.

[0150] L1 regularization (Lasso), L2 regularization (Ridge), Elastic-net, etc., may be employed as regularization terms. Known algorithms such as stochastic gradient descent (SGD), coordinate descent, and subgradient methods may be used to update the weights of the surrogate model nodes. Algorithms such as backpropagation may be employed to improve computational speed. However, this is not limited to these, and regularization terms and weight optimization algorithms used in various surrogate models may be employed.

[0151] The learning device can train a surrogate model on a single set of hyperparameters by repeating this process (S620-S560). Subsequently, the optimal set of hyperparameters can be determined through grid search, cross-validation, etc., on the set of hyperparameters.

[0152] Through the above, the surrogate model can be trained to receive a physical sensor signal for the device to be analyzed and to infer a target measurement value of the device to be analyzed that is difficult to measure with a physical sensor. The trained surrogate model can be included in the DRN master (131) and used in the aforementioned cyber sensor signal generation operation.

[0154] FIG. 7 is an example diagram of the configuration of a device according to one embodiment.

[0155] A device (701) according to one embodiment may include a processor (702) and a memory (703). The device (701) may be the device described above with reference to FIGS. 1 to 6. The device (701) may perform at least one of the operations described above with reference to FIGS. 1 to 6.

[0156] Memory (703) may store information related to the methods described above or store a program in which the methods described above are implemented. Memory (703) may be volatile memory or non-volatile memory.

[0157] The processor (702) can execute a program and control the device (701). The code of the program executed by the processor (702) can be stored in memory (703). The device (701) is connected to an external device (e.g., a personal computer or a network) through an input / output device (not shown) and can exchange data through wired or wireless communication.

[0158] The device (701) may be used to train a deep learning or machine learning model or to use a trained deep learning or machine learning model. The memory (703) may contain a deep learning or machine learning model that is being trained or has been trained. The processor (702) may train or execute a deep learning or machine learning model algorithm stored in the memory (703). The device (701) for training the deep learning or machine learning model and the device (701) for using the trained deep learning or machine learning model may be the same or separate.

[0160] The embodiments described above may be implemented as hardware components, software components, or a combination of hardware and software components. For example, the devices, methods, and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on the operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0161] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to a person skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. Hardware devices may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0162] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or command the processing unit independently or collectively. Software or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0163] Although the embodiments have been described above with reference to the limited drawings, those skilled in the art can apply various technical modifications and variations based thereon. For example, appropriate results can be achieved even if the described techniques are performed in a different order than described, or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or if they are replaced or substituted by other components or equivalents.

[0164] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A digital reflex neural network (DRN)-based control system comprising: a sensor for measuring the state of a device to be analyzed; a digital reflex neural network slave connected to the sensor; a digital reflex neural network master connected to the digital reflex neural network slave; one or more actuators connected to the digital reflex neural network master; and a semantic adapter connected to the digital reflex neural network master. The system includes a management server connected to the semantic adapter, and if one or more sensor signals conform to a logic rule of a predefined format, according to a first process, the digital reflex neural network master independently generates a first control command and transmits it to the one or more actuators; if one or more sensor signals do not conform to a logic rule of a predefined format, according to a second process, the semantic adapter converts data collected from at least one digital reflex neural network master, including the digital reflex neural network master, into a predefined unified format and transmits it to the management server; the management server generates a second control command based on the unified format data received from the semantic adapter and transmits it to the semantic adapter; the semantic adapter receives the second control command from the management server, converts the second control command into a predefined individual format required by each of the at least one digital reflex neural network masters, and transmits it to each of the at least one digital reflex neural network masters; and the digital reflex neural network master receives the second control command converted into its own individual format from the semantic adapter and transmits it to the one or more actuators A digital reflex neural network (DRN)-based control system configured to transmit, wherein the first process is faster than the second process. Claim 2 A digital reflex neural network (DRN)-based control system according to claim 1, wherein the logic rule of the predefined form comprises one or more sensor signal inputs; a condition expression set based on the one or more sensor signal inputs; and one or more control command outputs when the condition expression is satisfied. Claim 3 In paragraph 2, the digital reflex neural network master comprises a logic editor that generates logic rules of the predefined format based on user input, in a digital reflex neural network (DRN)-based control system. Claim 4 In paragraph 3, the logic editor includes a Smaller Larger Language Model (sLLM), and the sLLM receives natural language or a fishbone diagram as input and generates logic rules of the predefined format, a digital reflex neural network (DRN)-based control system. Claim 5 In paragraph 2, the digital reflex neural network master includes a lightweight FEM (Finite Element Method) engine that generates a cyber sensor signal based on a physical sensor signal acquired from the digital reflex neural network slave, and a digital reflex neural network (DRN)-based control system that generates the control command by inputting the physical sensor signal and the cyber sensor signal into a logic rule of the predefined format. Claim 6 A digital reflex neural network (DRN)-based control system according to paragraph 2, comprising a pre-trained surrogate model that generates a cyber sensor signal based on a physical sensor signal acquired from the digital reflex neural network slave, and generating the control command by inputting the physical sensor signal and the cyber sensor signal into a logic rule of the pre-defined format. Claim 7 A digital reflex neural network (DRN)-based control system according to paragraph 2, wherein the digital reflex neural network master transmits a control command according to a predefined order to the same actuator when a first control command generated based on a first logic rule and a second control command generated based on a second logic rule require different actions from the same actuator. Claim 8 A sensor for measuring the state of a device to be analyzed; a digital reflex neural network slave connected to the sensor; a digital reflex neural network master connected to the digital reflex neural network slave; one or more actuators connected to the digital reflex neural network master; and a semantic adapter connected to the digital reflex neural network master. A digital reflex neural network (DRN)-based control method performed by a management server connected to the semantic adapter, comprising: a step in which, when one or more sensor signals conform to a logic rule of a predefined format, the digital reflex neural network master independently generates a first control command and transmits it to one or more actuators according to a first process; and when one or more sensor signals do not conform to a logic rule of a predefined format, the semantic adapter converts data collected from at least one digital reflex neural network master, including the digital reflex neural network master, into a predefined unified format and transmits it to the management server according to a second process; a step in which the management server generates a second control command based on the unified format data received from the semantic adapter and transmits it to the semantic adapter; and a step in which the semantic adapter receives the second control command from the management server, converts the second control command into a predefined individual format required by each of the at least one digital reflex neural network masters, and transmits it to each of the at least one digital reflex neural network masters. A digital reflex neural network (DRN)-based control method comprising the step of the digital reflex neural network master receiving a second control command converted into its own individual format from the semantic adapter and transmitting it to the one or more actuators, wherein the first process is configured to be faster than the second process.

Images