A robot virtual-real process parameter closed-loop regulation method based on GNN

Abstract

The present application relates to the technical field of industrial robot intelligent manufacturing and digital twin control, in particular to a robot virtual-real process parameter closed-loop regulation and control method based on GNN; the method comprises the following steps: collecting multiple types of state data of a trial production line, current process parameters and control equipment information, uniformly accessing an edge processing unit through a protocol conversion gateway, completing timestamp alignment, data quality verification, filtering correction and edge preprocessing; synchronously updating a production line simulation model, constructing a graph structure taking process stations as nodes and material flow relationship and process dependency relationship as edges, analyzing process failure probability by using GNN and checking abnormal stations; optimizing process parameters based on the updated simulation model, generating closed-loop control instructions and issuing them to programmable logic controllers and robot controllers, realizing virtual-real process parameter closed-loop regulation and control; the present application improves assembly consistency, beat stability and production line operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and digital twin control technology for industrial robots, specifically a closed-loop control method for virtual and real process parameters of robots based on GNN. Background Technology

[0002] Pilot production lines for embodied intelligent robots typically integrate multiple units such as joint modules, end effectors, testing equipment, automated guided vehicles (AGVs) logistics equipment, programmable logic controllers (PLCs), robot controllers, and simulation systems. During production line operation, a large amount of status data and process parameters are continuously generated. Due to the inconsistent communication protocols, sampling periods, and data representation methods of different devices, field data is often scattered across multiple controllers and interfaces, making it difficult to directly aggregate into a unified data processing link. Unified access, time alignment, and quality verification of heterogeneous data are performed to perform status analysis and process adjustments.

[0003] Status data from different devices is prone to issues such as clock drift, sampling jitter, missing values, and outliers. Directly using this data for analysis can lead to inconsistencies between virtual and real states, thus affecting the updating effectiveness of the simulation model. Furthermore, there are direct material flow relationships and process dependencies between workstations on the production line. Traditional methods often rely on manual experience or single-point threshold judgments to identify faults, making it difficult to analyze the propagation impact of faults between workstations in a timely manner and to effectively optimize current process parameters. Especially in closed-loop control processes, the lack of a unified mechanism for status modeling, fault impact analysis, and simulation verification can lead to delayed control command issuance, resulting in inaccurate parameter adjustments and consequently affecting assembly consistency, cycle time stability, and production line operating efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a closed-loop control method for robot virtual and real process parameters based on GNN, which avoids the inconsistency between virtual and real states caused by the direct application of heterogeneous state data, thus affecting the update effect of the simulation model. It is also easier to identify the propagation of process faults at the workstation level in advance and realize the closed-loop optimization of process parameters, thereby effectively improving the assembly consistency and workstation cycle stability of the pilot production line.

[0005] The objective of this invention can be achieved through the following technical solution: acquiring multiple types of state data, current process parameters, and information on production line control equipment from a robot pilot production line; connecting the multiple types of state data to an edge processing unit via a protocol conversion gateway for timestamp alignment and data quality verification; performing filtering correction and edge preprocessing on the timestamp-aligned multiple types of state data to obtain real-time state data, and synchronizing it to the production line simulation model to update its virtual state; constructing a graph structure based on the real-time state data, with process stations as nodes and material flow and process dependencies as edges, and using the real-time state data as node state parameters as input to a process relationship analysis model employing a graph neural network. A process failure impact analysis is performed to obtain the failure probability value of each process station. When the failure probability value of any process station is greater than or equal to the preset failure probability threshold, a failure backtracking is performed based on process dependencies, and early warning information and abnormal process station identification results are output. When the failure probability values ​​of all process stations are less than the preset failure probability threshold, real-time monitoring continues. The current process parameters are simulated and optimized based on the production line simulation model updated by the virtual state to obtain optimized process parameters. The optimized process parameters are used to generate closed-loop control commands and sent to the production line control equipment. The production line control equipment automatically adjusts the corresponding execution parameters based on the closed-loop control commands.

[0006] Preferably, the multiple status data include at least joint module status data, end effector status data, and automated guided vehicle (AGV) logistics equipment status data, and the joint module status data includes at least motor status, reducer status, and encoder status.

[0007] Preferably, the protocol conversion gateway supports one or more protocols among Controller Area Network (CAN) protocols, Ethernet control automation technology protocols, serial communication protocols, and Open Platform Communication Unified Architecture (OPA) protocols, and converts data from different protocols into message formats of Message Queuing Telemetry Transmission Protocol (MQTP) or Open Platform Communication Unified Architecture (OPA) protocols.

[0008] Preferably, timestamp alignment includes aligning data streams from different devices based on a hardware clock or a unified clock for edge processing units, and filtering correction includes interpolation compensation for missing data and smoothing of continuous sampled values.

[0009] Preferably, edge preprocessing includes moving average filtering and outlier removal.

[0010] As a preferred embodiment, in the graph structure, the nodes include at least assembly stations, inspection stations, and logistics stations, and the edges include at least material transfer edges and process constraint edges; the process relationship analysis model is used to analyze the impact of one or more faults, such as assembly stress exceeding limits, joint return abnormalities, and cable entanglement.

[0011] As a preferred approach, using a process relationship analysis model to analyze the impact of process failures includes: summarizing information and transferring data between adjacent nodes within a graph structure.

[0012] As a preferred option, simulation optimization includes optimizing one or more of the following: joint torque compensation values, assembly sequence, and cycle time, and generating process parameter adjustment schemes based on the optimization results.

[0013] Preferably, the closed-loop control commands include parameter update commands and motion adjustment commands. The parameter update commands are used to update the servo drive parameters, and the motion adjustment commands are used to adjust the automated guided vehicle scheduling path and / or the tightening tool torque.

[0014] As a preferred option, the closed-loop control method also includes: evaluating the effect of the simulation-optimized process parameters. The evaluation includes comparing the simulation results with the preset process index range, generating a compliance mark when all process indices meet the requirements, and solidifying the optimized process parameters as the current production line process benchmark parameters. If any process index does not meet the requirements, the optimized process parameters are used only as the result of this adjustment.

[0015] The beneficial effects of the present invention are as follows: 1. By uniformly encapsulating, protocol converting and accessing, aligning timestamps and verifying data quality for multiple types of status data, the present invention can effectively solve the problems of scattered data sources, inconsistent sampling periods and inconsistencies between virtual and real states caused by clock drift from heterogeneous devices. Furthermore, by reducing the interference of missing values, outliers and sampling jitter on subsequent analysis through filtering correction and edge preprocessing, the present invention improves the accuracy and availability of real-time status data.

[0016] 2. This invention constructs the process station, material flow relationship and process dependency relationship into a graph structure, and uses graph neural network to summarize and analyze the information of adjacent station status. It can break through the limitations of the traditional single-point threshold judgment method, realize the early identification of the impact of fault propagation between stations and the accurate location of abnormal process stations.

[0017] 3. This invention synchronizes the updated real-time status to the production line simulation model, optimizes the current process parameters based on the simulation results, and then generates closed-loop control instructions to be sent to the programmable logic controller and robot controller. This provides clear simulation verification basis and execution path for process parameter adjustments, reduces adjustment deviations caused by control instruction lag, and improves assembly consistency, cycle stability, and production line operating efficiency. Attached Figure Description

[0018] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0019] Figure 1This is a flowchart illustrating a closed-loop control method for robot virtual and real process parameters based on GNN, provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 A closed-loop control method for robot virtual-real process parameters based on GNN is proposed, comprising: acquiring multi-class state data, current process parameters, and production line control equipment information of the robot pilot production line; connecting the multi-class state data to the edge processing unit via a protocol conversion gateway for timestamp alignment and data quality verification; performing filtering correction and edge preprocessing on the timestamp-aligned multi-class state data to obtain real-time state data, and synchronizing it to the production line simulation model to update its virtual state; constructing a graph structure with process stations as nodes and material flow and process dependencies as edges based on the real-time state data, and using the real-time state data as node state parameter input to perform process fault impact analysis using a graph neural network process relationship analysis model to obtain the fault probability value of each process station; when the fault probability value of any process station is greater than or equal to a preset fault probability threshold, performing fault backtracking based on process dependencies, and outputting early warning information and abnormal process station identification results; when the fault probability values ​​of all process stations are less than the preset fault probability threshold, continuing real-time monitoring; and updating based on virtual state. The subsequent production line simulation model optimizes the current process parameters to obtain optimized process parameters; the optimized process parameters generate closed-loop control commands and send them to the production line control equipment; the production line control equipment automatically adjusts the corresponding execution parameters based on the closed-loop control commands; multiple types of status data include at least joint module status data, end effector status data, and automated guided vehicle (AGV) logistics equipment status data, and the joint module status data includes at least motor status, reducer status, and encoder status; the protocol conversion gateway supports one or more protocols among Controller Area Network (CAN) protocols, Ethernet control automation technology protocols, serial communication protocols, and open platform communication unified architecture protocols, and uniformly converts data from different protocols into message queue telemetry transmission protocol or open platform communication unified architecture protocol message format; timestamp alignment includes aligning data streams from different devices based on hardware clocks or edge processing unit unified clocks; filtering correction includes interpolation compensation for missing data and smoothing of continuous sampled values; edge preprocessing includes moving average filtering and outlier removal.

[0022] This implementation method is applicable to a pilot production line for embodied intelligent robots; the production line includes joint assembly-related equipment, testing-related equipment, automated guided vehicle (AGV) logistics equipment, programmable logic controllers (PLCs), robot controllers, protocol conversion gateways, edge processing units, and computing devices housing the production line simulation model; multiple types of status data refer to measurable operating parameters from different physical devices, which correspond to objective physical quantities in the actual operation of the production line, such as motor operating status, encoder feedback values, end effector operating status, and automated guided vehicle (AGV) operating status.

[0023] Current process parameters refer to process settings that are currently in effect or awaiting execution in the production line control equipment, such as joint torque compensation values, assembly sequence associated parameters, and cycle time parameters; production line control equipment information refers to information used to determine the parameter distribution targets and parsing methods, including at least the equipment type and equipment communication interface type; at the beginning of a system cycle, the edge processing unit collects various types of status data, current process parameters, and production line control equipment information from the pilot production line.

[0024] Preferably, the joint module status data includes at least the motor status, reducer status, and encoder status; the end effector status data is used to reflect the current working status of the actuator; and the automated guided vehicle logistics equipment status data is used to reflect the material transfer status between logistics workstations.

[0025] Since these data come from different controllers or sensor interfaces, the edge processing unit encapsulates them according to a unified data frame format. This data frame format contains at least four fields: device identifier, timestamp, parameter type, and parameter value. The device identifier is used to distinguish the data source, the timestamp is used for subsequent alignment, the parameter type is used to mark which category the value belongs to: motor state, reducer state, encoder state, end effector state, or automated guided vehicle state, and the parameter value is used to carry the sampled value. Through the above encapsulation, the originally semantically inconsistent data is organized into a unified structure, which facilitates subsequent unified parsing.

[0026] After encapsulation, the protocol conversion gateway performs unified access to data from different buses or different industrial interfaces; preferably, the protocol conversion gateway supports one or more protocols among the following: Controller Area Network (CAN) protocol, Ethernet control automation technology protocol, serial communication protocol, and open platform communication unified architecture protocol; when a device uploads status data through the CAN protocol, the gateway reads the device identifier and parameter content according to the data frame format of the protocol.

[0027] When another device uploads data via Ethernet control automation technology protocol or serial communication protocol, the gateway disassembles it according to the corresponding protocol fields. When the detection device has directly output the open platform communication unified architecture protocol format, the gateway retains its object semantics and completes the unified fields. Then, the gateway converts the data from different protocols into the message queue telemetry transmission protocol or the open platform communication unified architecture protocol message format before sending it to the edge processing unit. Using the message queue telemetry transmission protocol or the open platform communication unified architecture protocol as the unified message format ensures that differences between upstream devices do not affect the downstream data processing flow, thereby reducing the number of subsequent software interfaces.

[0028] After receiving the unified message, the edge processing unit performs timestamp alignment and data quality verification on multiple data streams. Timestamp alignment refers to mapping data generated by different devices within similar sampling periods to the same time base. This time base can come from a hardware clock or a unified clock of the edge processing unit. If a device has its own hardware clock, the edge processing unit directly reads the timestamp and sorts it according to the unified time base. If some devices do not have independent hardware clocks, the edge processing unit adds a unified clock mark when receiving data and then aligns it according to the receiving order and sampling period.

[0029] To address the potential clock drift issue between heterogeneous devices, the edge processing unit maintains a bidirectional sliding time alignment window. Upon receiving a device data frame, it calculates the time difference between the device timestamp and the heartbeat synchronization signal uniformly sent from the edge. If this time difference exhibits a cumulative unidirectional offset over multiple consecutive system cycles and exceeds a preset time difference tolerance, a clock drift compensation control branch is triggered. If the time difference is less than or equal to the preset time difference tolerance, the current timestamp alignment is maintained, and a first-order smoothing exponential algorithm with a set smoothing coefficient is used to dynamically lag or lead the timestamp of the received data for interpolation correction, ensuring strict alignment of the time axis of the spliced ​​state parameters.

[0030] Data quality verification checks at least two categories of content: one is field integrity, which means whether the device identifier, timestamp, parameter type, and parameter value are complete; the other is value availability, which means whether the parameter value is empty, invalid, or a preset abnormal value that exceeds the device's working range.

[0031] For data frames with incomplete fields, the edge processing unit can directly mark them as data to be compensated; for data frames with abnormal parameter values, the edge processing unit first retains their source information and then enters the subsequent outlier processing flow; the aligned data enters the filtering and correction step; the filtering and correction includes interpolation compensation for missing data and suppression of sampling jitter; the aligned data entering the filtering and correction step is the multi-class state data that has been timestamped in the aforementioned steps, and the two represent the same data processing stage and data set, thereby maintaining strict consistency between the specific implementation method in the specification and the feature terms recorded in the claims.

[0032] In one implementation, the edge processing unit uses the same parameter values ​​within adjacent system cycles as a basis to interpolate and compensate for missing parameters within a single cycle, ensuring that the subsequent analysis model will not be interrupted by single-point sampling failures. For sampling jitter, the edge processing unit uses a Kalman filter fusion method to smooth continuous sampled values, making the state curve closer to the actual operating trend of the equipment. If data containing jitter and missing samples is directly input into the subsequent graph neural network, the state transmission results between adjacent workstations will be affected by instantaneous noise, resulting in excessive fluctuations in the probability of failure.

[0033] After filtering and correction, the edge processing unit continues to perform edge preprocessing; edge preprocessing refers to the preliminary sorting of data at the edge to reduce transmission burden and retain status information that is meaningful for process judgment; preferably, edge preprocessing includes moving average filtering and outlier removal.

[0034] Moving average filtering is used to average the state values ​​over a continuous system period to reduce random fluctuations in short periods; outlier removal is used to remove outliers that deviate from a preset trend threshold and have been marked by data quality verification; after the above processing, real-time state data is obtained.

[0035] Reducing cloud transmission bandwidth means that when real-time status data is not only used for updating local production line simulation models, but also simultaneously sent to cloud monitoring, historical archiving, or cross-production line analysis platforms, the edge side first performs compressed preprocessing, thereby reducing the amount of data uploaded.

[0036] When the production line simulation model is deployed on a local computing device, the aforementioned edge preprocessing can still be directly used to reduce invalid data transmission between the edge processing unit and the simulation model.

[0037] Real-time status data is a set of status parameters mapped to each workstation and each piece of equipment and used to directly update the simulation model.

[0038] After obtaining the real-time status data, the edge processing unit synchronizes the data to the production line simulation model to update its virtual status. The production line simulation model is a virtual production line model used to describe the operating status of pilot production line equipment, workstation relationships, and the effects of process parameters.

[0039] During synchronization, the edge processing unit establishes a correspondence between the device identifier and the device object in the simulation model, and writes the joint module status, end effector status and automated guided vehicle status within the same system cycle into the status area of ​​the corresponding object in the model. As a result, the device status in the simulation model is consistent with the physical production line status. When the process parameters are subsequently optimized through simulation, the input basis is no longer the static set value, but the updated operating status.

[0040] While the simulation model is being updated, the edge processing unit constructs a process relationship diagram structure based on real-time status data; a process station refers to a specific work unit on the production line that undertakes assembly, inspection, or logistics tasks.

[0041] Material flow relationship refers to the order in which materials are transferred between workstations; process dependency relationship refers to the constraint relationship in which the result of the subsequent workstation is affected by the state of the preceding workstation.

[0042] In this embodiment, each process station corresponds to a node in the graph. If there is material transfer between adjacent stations, a material transfer edge is established; if there are process constraints before and after, a process constraint edge is established. The edge processing unit maps the real-time status data corresponding to the station to node status parameters, and then inputs them into the process relationship analysis model using a graph neural network.

[0043] The process relationship analysis model of graph neural network refers to a model that transmits and summarizes data according to the relationship between workstation nodes and their adjacent edges. In this embodiment, a graph convolutional network architecture is preferred. Its input includes an adjacency matrix describing the graph topology and a feature matrix of each node. The input state parameters of each process workstation node are constructed as a feature vector of fixed dimension. This vector is formed by splicing normalized physical quantities such as motor running current, encoder position deviation, and end effector force feedback in a preset order. Masking and zero-padding operations are performed for the types of physical quantities missing in different workstations to ensure the uniformity of the feature dimensions of each node.

[0044] The processing involves each node first reading its own real-time state feature vector; the process relationship analysis model preprocesses the graph adjacency matrix containing the node's own self-loops using the corresponding angle matrix to perform symmetric normalization, thus limiting the tensor scale divergence during feature aggregation.

[0045] By performing tensor multiplication on the normalized adjacency matrix, the feature matrix of each node, and the learnable weight matrix of the network layer, the linear aggregation of the state features of adjacent workstations along the edge to the current node is achieved in mathematical mapping. The aggregated hidden layer feature tensor is input into a nonlinear activation function to filter out negative numerical disturbances that have no process significance. Through a fully connected output layer and a normalized exponential function, the summarized high-dimensional features are mapped to the probability space, and the failure probability value of each process workstation is output.

[0046] The fault probability value ranges from 0 to 1. The closer the value is to 1, the higher the probability that the workstation is affected by a process fault. In this embodiment, the preset fault probability threshold is set between 0.7 and 0.9, preferably 0.8. The threshold is set based on the fault identification results of historical samples from the pilot production line: when the fault probability threshold is less than the preset lower limit, the normal state fluctuation is judged as abnormal; when the fault probability threshold is greater than the preset upper limit, the fault identification time delay is greater than the preset time tolerance. By limiting the fault probability threshold within the above range, the conditions for early warning response time and false alarm prevention control are met.

[0047] When the failure probability value of any process station is greater than or equal to a preset failure probability threshold, the edge processing unit performs a fault backtracking based on process dependencies. Fault backtracking refers to determining the abnormal workstation by executing the following decision tree logic: starting from the current failure probability value... Starting from a high-risk workstation with a probability of ≥0.8, traverse the set of all directly preceding workstations along the process dependency edges; if there exists a failure probability value for any preceding workstation i in this set... If the deviation of the effective physical quantity in the feature vector of the current high-risk workstation is ≥0.6 and exceeds 15% of the nominal maximum value, then the source of the anomaly is determined to be the preceding workstation i, and it is marked as an abnormal process workstation; if all directly preceding workstations If the value is less than 0.6, the current high-risk workstation is determined to be an abnormal process workstation; once confirmed, the edge processing unit outputs a warning message.

[0048] The warning information includes at least the abnormal process station identifier, fault category indication, and corresponding fault probability value, used to indicate the process position that needs to be adjusted in the control equipment or monitoring interface. If the fault probability values ​​of all nodes are less than the preset fault probability threshold, the edge processing unit keeps the currently issued and executed process parameters unchanged and continues to perform real-time monitoring in subsequent system cycles. At this time, the production line simulation model can still perform background consistency verification on the current process parameters based on the updated virtual state to confirm whether the current parameter combination continues to meet the operating conditions, but does not generate new closed-loop control instructions. Through the above processing, the conditional relationship between fault warning triggering, simulation optimization execution, and control instruction issuance becomes clearer.

[0049] For workstations where abnormal risks have been identified, the production line simulation model further optimizes the current process parameters to obtain optimized process parameters.

[0050] In this implementation, the premise of simulation optimization is that the simulation model has been updated by real-time state data. Therefore, the optimization object is no longer an isolated parameter, but a combination of parameters that act on the current real operating state.

[0051] After reading the current process parameters, the simulation model changes the parameter values ​​and observes the impact on the station status and the constraint relationship between adjacent stations. A target cost function is then constructed. Where J is the objective cost function, The output is the failure probability value. The absolute value of the time deviation between the running beat and the standard beat. For standard beat time, and Dividing is used to unify dimensions, making it equal to the dimensionless probability value. Maintain dimensional consistency. These are preset weighting coefficients that measure the proportion of the impact of failure probability to the impact of cycle time deviation, and their sum is 1.

[0052] The numerical calculation logic of the target cost function is the product of the first weighting coefficient and the failure probability value plus the product of the second weighting coefficient and the time deviation ratio. The aforementioned time deviation ratio is obtained by dividing the absolute value of the time deviation between the operating cycle time and the standard cycle time by the quotient of the standard cycle time. This clearly defines the division relationship in the formula and achieves dimensional uniformity, thereby providing effective support for the claims.

[0053] A grid search algorithm is used to perform discrete sampling traversal with a step size of 1% within a continuous interval of 10% above and below the current process parameters. The set of parameters that minimizes the objective cost function J is taken as the optimized process parameters. Considering the time granularity difference between the millisecond-level real-time control of the physical production line and the minute-level calculation of the simulation model, this implementation adopts an asynchronous closed-loop control link: when the edge processing unit outputs early warning information, it can issue a short-cycle safety hold command to the corresponding abnormal process station to avoid the abnormality from worsening; at the same time, it triggers the production line simulation model to asynchronously execute the above simulation optimization process on the computing device.

[0054] The edge processing unit generates control commands from the optimized process parameters and sends them to the production line control equipment. Due to the minute-level delay in simulation calculations, while the physical production line cycle time is on the second level, the edge processing unit temporarily stores the commands in the asynchronous sending queue after receiving the optimized process parameters. It continuously monitors the operating cycle time of the physical production line and preferably activates the closed-loop control command when the production line reaches the nearest idle cycle time start point or safe reversal node. This is to prevent irreversible mechanical shocks caused by sudden changes in parameters during the continuous movement of the physical actuator due to communication and calculation delays.

[0055] The production line control equipment includes programmable logic controllers (PLCs) and robot controllers. The edge processing unit generates corresponding instructions based on the different controlled objects: when the target object is a PLC, the instructions include the process parameters to be updated and the execution object identifier; when the target object is a robot controller, the instructions include parameter updates related to robot movement or execution. After receiving the instructions, the PLCs and robot controllers automatically adjust the corresponding execution parameters so that the equipment in the physical production line runs according to the new process parameters.

[0056] Thus, the status data from the physical production line, the optimization calculation results from the simulation model, and the parameter instructions returned to the physical equipment form a complete data round-trip process, enabling the production line to adjust process parameters during continuous operation. Through the above processing, the status information originally generated by heterogeneous equipment is organized into a unified and computable data stream. Then, with the help of edge processing, graph neural network analysis, and simulation optimization, parameter adjustment is completed. The direct result is a reduction in data deviation between the physical production line and the simulation model. Process fault propagation can be identified in advance at the workstation level. At the same time, the programmable logic controller and robot controller can update the execution parameters based on the optimization results, thereby improving the assembly consistency and workstation cycle stability of the pilot production line.

[0057] In some possible implementations, the nodes in the graph structure include at least assembly stations, inspection stations, and logistics stations, and the edges include at least material transfer edges and process constraint edges. The process relationship analysis model is used to analyze the impact of one or more faults, such as excessive assembly stress, abnormal joint return, and cable entanglement. The process fault impact analysis using the process relationship analysis model includes: summarizing information and transferring data between adjacent nodes in the graph structure. The simulation optimization includes optimizing one or more of the joint torque compensation value, assembly sequence, and cycle time, and generating process parameter adjustment schemes based on the optimization results.

[0058] The closed-loop control commands include parameter update commands and motion adjustment commands. The parameter update commands are used to update the servo drive parameters, and the motion adjustment commands are used to adjust the automated guided vehicle scheduling path and / or the tightening tool torque.

[0059] This implementation method focuses on illustrating the coordination process between process relationship diagrams, process failure impact analysis, simulation optimization, and control command issuance.

[0060] The application scenario can be a robot pilot production line containing assembly stations, inspection stations, and logistics stations. The assembly station is responsible for assembling joint modules or complete machine components, the inspection station is responsible for inspecting the assembly results, and the logistics station is responsible for material transfer by automated guided vehicles (AGVs). Within a system cycle, the edge processing unit obtains real-time status data from each station and constructs a graph structure according to the station type. Nodes include at least the assembly station, inspection station, and logistics station mapped to specific stations in the production line. Assembly station nodes can receive joint module status, end effector status, or tightening-related status. Inspection station nodes can receive status parameters corresponding to inspection results. Logistics station nodes can receive AGV running status and logistics transfer status. Material transfer edges in the graph represent the sequential relationship of products or components flowing from one station to the next, and process constraint edges represent how the result of one station affects the execution conditions of another station.

[0061] For example, the status of the assembly station will affect the inspection results of the subsequent inspection station, while the arrival status of the logistics station will affect the material loading conditions of the assembly station. Therefore, the edge processing unit establishes corresponding edge relationships respectively. After the graph structure is established, the edge processing unit maps the real-time status data of each station to node status parameters. The mapping is configured as a field association processing step. The edge processing unit first determines which station the data belongs to based on the equipment identifier, and then writes the real-time status data into the corresponding parameter area of ​​the station node according to the parameter type.

[0062] If a workstation corresponds to multiple devices, such as a joint module and an end effector existing in the same assembly workstation, the edge processing unit will combine the status parameters from these devices as the input data for that node. Through the above processing, each node in the diagram carries status parameters consistent with the actual operation of the workstation.

[0063] A process relationship analysis model using graph neural networks is used to analyze the impact of process failures. The process relationship analysis model is used to analyze one or more failures, such as excessive assembly stress, abnormal joint return, and cable entanglement. The analysis process does not directly give the final conclusion, but first performs data transfer and information aggregation between adjacent nodes in the graph.

[0064] Specifically, each node first reads its own state parameters to construct an initial feature vector. Since the types and dimensions of measurable physical quantities generated by assembly, inspection and logistics workstations are completely different, for example, the assembly workstation has torque parameters, while the logistics workstation only has position coordinates. If the network layer features are summed directly, it will lead to tensor dimension misalignment and confusion of physical meaning.

[0065] In this implementation, when constructing the graph structure, corresponding feature zero-padding masks and independent linear projection operators are assigned to different types of nodes. The projection operator is used to map initial feature vectors with different dimensions and lengths to a shared latent space with consistent dimensions, forming a node latent layer feature matrix with a unified dimension. Then, the latent layer feature vectors of the preceding and following workstations are received along the material transfer edge, and the latent layer feature vectors of the workstations that have a dependency relationship with them are received along the process constraint edge. After that, based on the message passing mechanism, this information is summarized and processed to output the fault probability value of the node.

[0066] Information aggregation is performed hierarchically according to the message aggregation mechanism of workstation relationship: the hidden features of the mapped node are linearly transformed by the shared weight matrix; the hidden features from directly adjacent nodes are extracted and aggregated by combining the transfer weights corresponding to the types of process constraint edges and material transfer edges, that is, the transformed features of the node itself are aligned and superimposed with the weighted features of adjacent nodes in the tensor dimension and summed.

[0067] The hidden layer representation of the current node is updated by an activation function. After feature aggregation by the network layers, the final node representation is input into the output layer for probability judgment, so that the output result can be accurately matched to the specific workstation by combining the comprehensive status of the preceding and following processes.

[0068] For example, if the joint module status of an assembly station deviates beyond the preset tolerance, and the subsequent inspection station reports abnormal inspection results, the failure probability of that assembly station in the diagram will increase. If the Automated Guided Vehicle (AGV) at a logistics station malfunctions, causing unstable material delivery, and the cable handling process at a subsequent assembly station experiences abnormal fluctuations, the information transmission between adjacent nodes in the diagram will increase the risk of cable entanglement at the relevant stations. To ensure clear judgment conditions, the failure probability value is still limited to between 0 and 1, and the preset failure probability threshold is set between 0.7 and 0.9, preferably 0.8. When the failure probability value of any node is greater than or equal to this threshold, the edge processing unit first identifies the node as an abnormal candidate process station and triggers an early warning message.

[0069] Then, by combining the process dependency relationship, a fault retrospective is performed to distinguish whether the high-risk node is the source of the anomaly or an abnormal behavior caused by the influence of the preceding workstation, so as to obtain the final result of the abnormal process workstation determination. The warning information includes at least the abnormal process workstation identifier, fault category prompt and fault probability value. If the fault probability value of all nodes is less than the preset fault probability threshold, the edge processing unit will not start new parameter adjustment, but will continue to perform real-time monitoring in subsequent system cycles.

[0070] The reason for the above settings is that the status of the workstations in the pilot production line fluctuates normally. The process parameters are only adjusted when the failure probability reaches a preset level, which can avoid frequent changes to the control parameters under normal operating conditions. After the abnormal process workstation is identified, the production line simulation model performs simulation optimization on the current process parameters. The optimization objects include at least one or more of the following: joint torque compensation value, assembly sequence, and cycle time.

[0071] Taking the assembly scenario of joint modules as an example, when the process relationship analysis model determines that the risk of assembly stress exceeding the limit is greater than or equal to the preset risk threshold, the simulation model first changes the joint torque compensation value and observes whether the failure probability of the assembly station node decreases after the adjustment; when the correlation coefficient between abnormal data and the coordination between the preceding and following processes is greater than the preset correlation threshold, the simulation model then changes the assembly sequence and examines whether the influence of the corresponding nodes of the material transfer edge and the process constraint edge is weakened; when it is determined that the problem is caused by the lack of coordination of the station cycle time, the simulation model adjusts the cycle time to observe the state changes between the logistics station and the assembly station; the result of the above process is not a single value, but forms a process parameter adjustment plan, which includes the parameter items that need to be changed and the corresponding adjustment direction; if all nodes in the current system cycle do not reach the preset failure probability threshold, the simulation model only retains the verification result of the current process parameters and does not output a new process parameter adjustment plan.

[0072] The edge processing unit generates control commands based on the process parameter adjustment scheme. Parameter update commands are used to update servo drive parameters, and motion adjustment commands are used to adjust the automated guided vehicle (AGV) scheduling path and / or tightening tool torque. If the anomaly is mainly concentrated in joint assembly-related workstations, the edge processing unit sends parameter update commands to the programmable logic controller (PLC), which then writes the updated content into the servo drive parameter area. If the anomaly is related to the logistics workstation, the edge processing unit sends motion adjustment commands to the PLC or logistics control unit to ensure that the AGV scheduling path is executed according to the optimized scheme. If the anomaly is related to tightening quality, the motion adjustment commands include tightening tool torque adjustment content, which is then sent to the tightening tool execution end by the corresponding controller.

[0073] Through the issuance of the above instructions, the risk results obtained from the process relationship analysis are directly converted into parameter or action adjustments that can be executed by the control equipment.

[0074] In this embodiment, there is a clear data connection between the graph structure, the process relationship analysis model, the simulation optimization, and the control instructions: the graph structure is responsible for representing the real dependencies between workstations, the process relationship analysis model uses the relationship between nodes and edges to calculate the failure probability, the simulation model tests the effect of different process parameters based on the updated workstation state, and the edge processing unit converts the optimization results into control instructions that can be recognized by the programmable logic controller and the execution device.

[0075] Through the above processing steps, process faults such as excessive assembly stress, abnormal joint return, and cable entanglement can be identified in advance during production line operation. The servo drive parameters, automated guided vehicle (AGV) path, and tightening tool torque are adjusted according to the working conditions, thereby reducing the continued impact of abnormal workstations on subsequent workstations. Among them, the process relationship analysis model preferably adopts a two-layer graph convolutional network structure and a one-layer fully connected output layer. In order to avoid the dimensional confusion and misalignment caused by directly using the superposition of original physical quantities, the original input node feature matrix is ​​preferably N×12-dimensional, where the 12-dimensional features correspond to preset physical quantities, including at least motor current, reducer temperature, encoder position deviation, encoder speed deviation, end effector force, end effector torque, AGV position deviation, AGV speed, joint vibration, servo temperature, task cycle deviation, and state mask.

[0076] Before inputting into the network, the edge processing unit divides each effective physical quantity by the nominal maximum value of its corresponding device to perform dimensionless normalization processing. When the system faces boundary anomalies such as damage or interruption of data acquisition by a certain type of sensor on the edge side, resulting in missing physical quantities, the engineering control program will trigger a fault-tolerant degradation control strategy: attach a low-confidence status label to the physical quantity field of the missing data, switch to the backup interpolation branch, and take the moving average of the previous three system cycles as the filler. If the number of consecutive missing cycles exceeds the preset missing number threshold, the physical characteristic of the node will be forcibly set to zero to avoid noise pollution.

[0077] Using a linear projection operator specific to the node type, the original N×12-dimensional feature matrix, after being masked and padded with zeros, is projected into an N×D-dimensional shared latent space, for example, D=64, forming a node latent layer feature matrix of uniform dimension. When entering a two-layer graph convolutional network to perform message aggregation calculation between adjacent workstations, the system adopts an adaptive attention allocation mechanism to perform weighted aggregation based on the physical topological properties of the edges.

[0078] The following adaptive attention allocation mechanism is used to calculate edge weights: For an edge between node i and its neighboring node j, the attention coefficient is calculated using the following formula: ;in, , Let be the hidden features of nodes i and j respectively, W be the shared weight matrix, a be the learnable parameter vector, the superscript T denotes the transpose of a matrix or vector, || denotes the concatenation operation, and LeakyReLU is the linear rectified function with leakage. Set a weight scaling penalty factor for the process constraint edges. =1.5, setting the material transfer edge =1.0; The final weights are obtained by normalization using the Softmax function: ;in, Let k be the set of neighboring nodes of node i, and k be the index of a node in the set of neighboring nodes. Let be the attention coefficient between node i and node k, and exp be the exponential function operation; this ensures that process anomalies with strict timing constraints can achieve higher cross-station transmission sensitivity.

[0079] To clarify the specific mathematical operation logic of the normalization process, the numerator of the final weight is the exponential operation result of the attention coefficient between the node itself and the target neighboring nodes, and the denominator is the sum of the exponential operation results of the attention coefficient between the node itself and all nodes in the neighboring node set. This undoubtedly reflects the summation and normalization process implicitly contained in the formula, ensuring that the sum of the weights passed by all neighboring nodes is one.

[0080] The training samples consist of state snapshots from the historical operation cycle of the pilot production line and subsequent confirmed fault results. Fault labels are generated from the workstation anomaly records after manual review. The loss function is preferably cross-entropy loss so that the output node fault probability value matches the historical fault label.

[0081] In some possible implementations, the closed-loop control method further includes: performing an effect evaluation on the simulation-optimized process parameters, which includes comparing the simulation results with the preset process index range, generating a compliance mark when all process indices meet the requirements, and solidifying the optimized process parameters as the current production line process baseline parameters; if any process index does not meet the requirements, the optimized process parameters are used only as the result of this adjustment.

[0082] This implementation describes the evaluation process of the execution effect after the optimization of process parameters is completed. This process is completed by the computing device where the production line simulation model is located and the edge processing unit. Its input is the process parameters that have been optimized through simulation, and its output is the judgment result of whether to solidify the set of process parameters as the current production line process baseline parameters.

[0083] Within a system cycle, after the production line simulation model completes the optimization of the current process parameters, it does not immediately retain all the optimization results as default parameters for a long period of time. Instead, it first enters the effect evaluation stage. The effect evaluation refers to re-verifying the process parameters after simulation optimization to confirm whether the execution results in the simulation model meet the pre-set process requirements of the production line.

[0084] The simulation model is loaded with optimized process parameters and run in conjunction with the updated virtual state to obtain simulation results. These simulation results correspond to the process results shown in the model after the optimized parameters are applied to the production line, rather than simply mathematical output.

[0085] The edge processing unit or the computing device where the simulation model is located compares the simulation results with the preset process index range. The preset process index range refers to the range of values ​​allowed for the production line in terms of assembly consistency, station cycle time and stability of related execution parameters. Assembly consistency is measured by the assembly torque error range, and station cycle time is measured by the time deviation between the preset running cycle time and the standard cycle time.

[0086] The preset process compliance conditions are clearly defined as all process indicators in the simulation results being within the range of the preset process indicators.

[0087] This implementation method requires that all process indicators participating in the comparison meet the range conditions before the optimized process parameters of the group are considered to have reached a curable state.

[0088] During the specific comparison, the computing device where the simulation model is located reads the corresponding simulation output according to the process parameter item and checks whether each item falls within the preset range. If any process indicator exceeds the preset range, it is determined that the current optimization result has not yet met the process compliance conditions. This set of parameters is only used as the result of this adjustment and is not written into the current production line process benchmark parameter area. If all process indicators are within the preset range, a compliance mark is generated.

[0089] The compliance mark is a confirmation result output by the computing device, used to notify the edge processing unit and production line control equipment that the set of process parameters has been verified by simulation and can be saved as the benchmark parameters for subsequent operation.

[0090] After generating the compliance mark, the edge processing unit will solidify the optimized process parameters as the current production line process baseline parameters. Solidification means writing the parameters that passed the evaluation into the production line control equipment or its associated parameter storage area, so that in subsequent system cycles, unless a new anomaly is triggered, this set of parameters will be used as the current default process setting.

[0091] The purpose of the above-mentioned solidification operation is to retain the verified and usable optimization results as a baseline, so as to avoid the repeated adjustment process when similar problems recur.

[0092] In this implementation, the effect evaluation is located between parameter optimization and long-term parameter storage, and the order cannot be reversed. If this step is omitted, even if a set of parameters is obtained through simulation optimization, it will be impossible to distinguish whether the parameter is only effective for the current local state or has the conditions to be used as a general benchmark for the production line.

[0093] By setting up a process to compare simulation results with preset process parameter ranges, and using the fact that all process parameters meet the range conditions as the standard for compliance, the parameters written into the production line process baseline parameter area can have a clear verification basis, thereby improving the consistency and stability of process parameters during subsequent production line operation.

Claims

1. A closed-loop control method for virtual and real process parameters of a robot based on GNN, characterized in that, include: Acquire various status data, current process parameters, and information on production line control equipment from the robot pilot production line; The protocol conversion gateway connects the various types of status data to the edge processing unit for timestamp alignment and data quality verification. The timestamp-aligned multi-class state data are filtered, corrected, and preprocessed to obtain real-time state data, which is then synchronized to the production line simulation model to update its virtual state. Based on the real-time status data, a graph structure is constructed with process stations as nodes and material flow and process dependencies as edges. The real-time status data is used as node status parameters as input to a graph neural network process relationship analysis model to analyze the impact of process failures and obtain the failure probability value of each process station. When the failure probability value of any process station is greater than or equal to a preset failure probability threshold, a failure backtracking is performed based on the process dependencies, and early warning information and abnormal process station identification results are output. When the failure probability values ​​of all process stations are less than the preset failure probability threshold, real-time monitoring continues. Based on the production line simulation model updated by the virtual state, the current process parameters are simulated and optimized to obtain the optimized process parameters; The optimized process parameters are used to generate closed-loop control commands, which are then sent to the production line control equipment. The production line control equipment automatically adjusts the corresponding execution parameters based on the closed-loop control command.

2. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The multiple types of status data include at least joint module status data, end effector status data, and automated guided vehicle (AGV) logistics equipment status data, and the joint module status data includes at least motor status, reducer status, and encoder status.

3. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The protocol conversion gateway supports one or more protocols among Controller Area Network (CAN) protocols, Ethernet control automation technology protocols, serial communication protocols, and open platform communication unified architecture protocols, and converts data from different protocols into message formats of message queue telemetry transmission protocol or open platform communication unified architecture protocol.

4. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The timestamp alignment includes aligning data streams from different devices based on a hardware clock or a unified clock for the edge processing unit, and the filtering correction includes interpolation compensation for missing data and smoothing of continuous sampled values.

5. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The edge preprocessing includes moving average filtering and outlier removal.

6. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, In the graph structure, nodes include at least assembly stations, inspection stations, and logistics stations, and edges include at least material transfer edges and process constraint edges; the process relationship analysis model is used to analyze the impact of one or more faults, such as excessive assembly stress, abnormal joint return, and cable entanglement.

7. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The process failure impact analysis using the aforementioned process relationship analysis model includes: summarizing information and transferring data between adjacent nodes in the graph structure.

8. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The simulation optimization includes optimizing one or more of the following: joint torque compensation value, assembly sequence, and cycle time, and generating a process parameter adjustment scheme based on the optimization results.

9. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The closed-loop control commands include parameter update commands and motion adjustment commands. The parameter update commands are used to update the servo drive parameters, and the motion adjustment commands are used to adjust the automated guided vehicle scheduling path and / or the tightening tool torque.

10. The closed-loop control method for robot virtual and real process parameters based on GNN according to claim 1, characterized in that, The closed-loop control method further includes: evaluating the effect of the simulation-optimized process parameters. The evaluation includes comparing the simulation results with the preset process index range, generating a compliance mark when all process indices meet the requirements, and solidifying the optimized process parameters as the current production line process benchmark parameters. If any of the process indices does not meet the requirements, the optimized process parameters are used only as the result of this adjustment.

Images