Edge computing collaborative control method and system for heterogeneous interconnection of multiple devices in intelligent manufacturing units

Abstract

This invention relates to the field of control, and in particular to an edge computing collaborative control method and system for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit. Specifically, by collecting functional metadata and real-time status data of heterogeneous devices, a weighted directed acyclic service graph is constructed based on a unified resource representation framework, and the weights of the edges in the graph are initialized as collaborative costs. Then, the production task is parsed as the target state of the service graph, and multiple candidate collaborative paths are generated through reverse search. A resource occupancy spatiotemporal tensor is constructed for each path and the execution risk coefficient is calculated. The path with the smallest coefficient is selected as the optimal task sequence. Subsequently, control commands are distributed and theoretical state vectors are predicted. The feedback from the collected devices is fused into actual state vectors. The Mahalanobis distance between the two is calculated. If the distance exceeds a threshold, the current sequence is frozen, a compensation path is replanned, and the collaborative costs of the corresponding edges of the service graph are updated nonlinearly based on the distance and abnormal task segments.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to an edge computing collaborative control method and system for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit. Background Technology

[0002] Modern intelligent manufacturing units deploy a large number of heterogeneous devices with varying functions and from different manufacturers, such as CNC machine tools, industrial robots, automated guided vehicles, and machine vision systems. Differences in communication protocols, data interfaces, and control command systems make reliable information exchange and behavioral collaboration difficult. Traditional manufacturing execution systems and other upper-level applications typically employ a centralized control architecture, using a cloud-edge model for unified device scheduling. However, when handling tasks requiring close collaboration among multiple devices, this approach faces challenges such as high network latency, poor real-time performance, and significant data transmission bandwidth pressure, making it difficult to meet the low-latency, high-reliability collaborative control requirements of flexible production scenarios. Furthermore, the control logic in this model is often pre-programmed, lacking sufficient flexibility and scalability to adapt quickly to changes in production line layout or production tasks.

[0003] By shifting computing and decision-making capabilities to the network edge closer to the devices, the real-time responsiveness of control systems can be improved. As mentioned above, most methods rely on static rules or predefined workflow templates to represent the collaborative relationships between devices, failing to construct service association models based on the real-time status and capabilities of the devices and autonomously generate optimal task sequences. Moreover, during task execution, the failure to reliably assess and mitigate the execution risks of collaborative paths can easily lead to operational interference, efficiency bottlenecks, and even safety incidents. The lack of a closed-loop adjustment capability that can monitor execution deviations online, assess the degree of anomalies, and accordingly perform path compensation and self-optimization of control strategies results in low system response efficiency to anomalies. Summary of the Invention

[0004] This invention proposes an edge computing collaborative control method for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit, which addresses the difficulty of existing technologies in meeting the requirements for low latency and high reliability collaborative control in flexible production scenarios. The method includes the following steps:

[0005] Obtain the functional metadata and real-time status data of each heterogeneous device in the intelligent manufacturing unit, and construct a weighted directed acyclic service graph representing the behavioral association between devices based on the unified resource representation framework. The weight of the edges in the graph is initialized as the collaboration cost.

[0006] The system receives production tasks and parses them into target states in the service graph. It then uses a reverse search algorithm to backtrack from the target state to the current device state, generating multiple candidate collaborative paths. For each candidate collaborative path, it constructs a spatiotemporal tensor representing the resource occupancy of each device and calculates the collaborative execution risk coefficient. Finally, it selects the path with the smallest collaborative execution risk coefficient as the optimal collaborative task sequence.

[0007] Control commands are distributed to each device according to the optimal collaborative task sequence, and the theoretical state vector of the sequence execution is predicted synchronously; the actual state data fed back by each device is collected and fused into the actual state vector of the unit.

[0008] Calculate the Mahalanobis distance between the actual state vector and the theoretical state vector. When the distance exceeds a preset threshold, freeze the execution of the current task sequence. Using the current actual state vector as a new starting point, replan the compensation path to the target state. Based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, increase the collaborative cost of the edge corresponding to the task sequence segment in the service graph nonlinearly through a preset weight update function.

[0009] Optionally, the weighted directed acyclic service graph representing the behavioral associations between devices is constructed based on the unified resource representation framework, wherein the weights of the edges in the graph are initialized as collaboration costs, including:

[0010] The robotic arms, CNC machine tools, and automated guided vehicles within the intelligent manufacturing unit are abstracted as nodes in the diagram;

[0011] The processing, handling, and testing services provided by the equipment are abstracted as directed edges connecting the nodes;

[0012] After normalizing the historical average execution time, rated energy consumption, and historical failure rate of each service, the collaborative cost C of the edge is initialized by weighted summation.

[0013] Optionally, the step of using a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate cooperative paths includes:

[0014] Using the target state as the starting point, the preceding service nodes that can reach the state are traversed in reverse in the service graph.

[0015] The reverse traversal process is continuously iterated until the starting node of the search path matches the current device state, thus obtaining a complete candidate cooperative path;

[0016] Repeat this process until all paths that meet the preset path length or cost limits are found.

[0017] Optionally, the construction of the spatiotemporal tensor representing the resource occupancy of each device and the calculation of the collaborative execution risk coefficient include:

[0018] Construct a spatiotemporal tensor with device ID as the first dimension, time step as the second dimension, and resource type as the third dimension. ;

[0019] For each task step in the candidate collaborative path, the estimated resource consumption of each step is filled into the tensor. The corresponding position;

[0020] The collaborative execution risk coefficient R is calculated by weighted summation of all resource occupancy points exceeding a preset safety threshold, using the following formula:

[0021]

[0022] in, This represents the resource usage value of type k for device i at time step j. The rated capacity of resource type k for device i, For the risk weight of type k resources, The security threshold for type k resources. For indicator functions, when the condition is true, that is... The value is 1 if it is true, and 0 otherwise.

[0023] Optionally, the step of distributing control commands to each device according to the optimal cooperative task sequence and synchronously predicting the theoretical state vector for the execution of the sequence includes:

[0024] A digital twin model of the intelligent manufacturing unit is established, and simulation is performed in the model according to the control commands distributed to each device;

[0025] At each preset time sampling point, key state parameters of each virtual device in the model are extracted, and the parameters are arranged and combined in a predetermined order to obtain the theoretical state vector.

[0026] Optionally, the process of collecting the actual state data fed back by each device and fusing it into the actual state vector of the unit includes:

[0027] The system periodically reads all joint angles from the robotic arm controller, spindle speed and feed rate from the CNC machine tool system, and real-time coordinates from the automated guided vehicle positioning system.

[0028] All the acquired data are concatenated into a one-dimensional column vector according to the same order and dimensions as the theoretical state vector, which serves as the actual state vector of the unit.

[0029] Optionally, calculating the Mahalanobis distance between the actual state vector and the theoretical state vector includes:

[0030] The covariance matrix is ​​calculated based on the set of deviation vectors between multiple actual state vectors and their corresponding theoretical state vectors under historical normal operating conditions. ;

[0031] Using the inverse of the covariance matrix Calculate the Mahalanobis distance between the current actual state vector and the theoretical state vector. , calculate Compare with a confidence threshold set based on a statistical distribution.

[0032] Optionally, the step of non-linearly increasing the collaborative cost of the edges corresponding to the task sequence segment in the service graph based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the occurrence of the anomaly, through a preset weight update function, includes:

[0033] For each edge in the task sequence segment corresponding to the occurrence of the anomaly, the new collaborative cost is... From the original collaborative cost The updated formula is: ;

[0034] in, For the calculated Mahalanobis distance, The preset Mahalanobis distance threshold, and These are preset normal values, which control the basic magnitude and growth rate of the cost increase, respectively.

[0035] Furthermore, this invention also relates to an edge computing collaborative control system for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit, comprising the following modules:

[0036] The acquisition module is used to acquire the functional metadata and real-time status data of each heterogeneous device in the intelligent manufacturing unit, and to construct a weighted directed acyclic service graph representing the behavioral association between devices based on the unified resource representation framework. The weight of the edges in the graph is initialized as the collaboration cost.

[0037] The construction module is used to receive production tasks and parse them into target states in the service graph. It uses a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate collaborative paths. For each candidate collaborative path, it constructs a spatiotemporal tensor representing the resource occupancy of each device and calculates the collaborative execution risk coefficient. The path with the smallest collaborative execution risk coefficient is selected as the optimal collaborative task sequence.

[0038] The fusion module is used to distribute control commands to each device according to the optimal collaborative task sequence, and synchronously predict the theoretical state vector of the sequence execution; collect the actual state data fed back by each device, and fuse them into the actual state vector of the unit.

[0039] The calculation module is used to calculate the Mahalanobis distance between the actual state vector and the theoretical state vector. When the distance exceeds a preset threshold, the execution of the current task sequence is frozen, and a compensation path to the target state is replanned with the current actual state vector as the new starting point. Based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, the collaborative cost of the edge corresponding to the task sequence segment in the service graph is nonlinearly increased through a preset weight update function.

[0040] Preferably, the weighted directed acyclic service graph representing the behavioral relationships between devices is constructed based on the unified resource representation framework, wherein the weights of the edges in the graph are initialized as collaboration costs, including:

[0041] The robotic arms, CNC machine tools, and automated guided vehicles within the intelligent manufacturing unit are abstracted as nodes in the diagram;

[0042] The processing, handling, and testing services provided by the equipment are abstracted as directed edges connecting the nodes;

[0043] After normalizing the historical average execution time, rated energy consumption, and historical failure rate of each service, the collaborative cost C of the edge is initialized by weighted summation.

[0044] Preferably, the step of using a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate cooperative paths includes:

[0045] Using the target state as the starting point, the preceding service nodes that can reach the state are traversed in reverse in the service graph.

[0046] The reverse traversal process is continuously iterated until the starting node of the search path matches the current device state, thus obtaining a complete candidate cooperative path;

[0047] Repeat this process until all paths that meet the preset path length or cost limits are found.

[0048] Preferably, the construction of the spatiotemporal tensor representing the resource occupancy of each device and the calculation of the collaborative execution risk coefficient include:

[0049] Construct a spatiotemporal tensor with device ID as the first dimension, time step as the second dimension, and resource type as the third dimension. ;

[0050] For each task step in the candidate collaborative path, the estimated resource consumption of each step is filled into the tensor. The corresponding position;

[0051] The collaborative execution risk coefficient R is calculated by weighted summation of all resource occupancy points exceeding a preset safety threshold, using the following formula:

[0052]

[0053] in, This represents the resource usage value of type k for device i at time step j. The rated capacity of resource type k for device i, For the risk weight of type k resources, The security threshold for type k resources. For indicator functions, when the condition is true, that is... The value is 1 if it is true, and 0 otherwise.

[0054] Preferably, the step of distributing control commands to each device according to the optimal collaborative task sequence and synchronously predicting the theoretical state vector for the execution of the sequence includes:

[0055] A digital twin model of the intelligent manufacturing unit is established, and simulation is performed in the model according to the control commands distributed to each device;

[0056] At each preset time sampling point, key state parameters of each virtual device in the model are extracted, and the parameters are arranged and combined in a predetermined order to obtain the theoretical state vector.

[0057] Preferably, the process of collecting the actual state data fed back by each device and fusing it into the actual state vector of the unit includes:

[0058] The system periodically reads all joint angles from the robotic arm controller, spindle speed and feed rate from the CNC machine tool system, and real-time coordinates from the automated guided vehicle positioning system.

[0059] All the acquired data are concatenated into a one-dimensional column vector according to the same order and dimensions as the theoretical state vector, which serves as the actual state vector of the unit.

[0060] Preferably, calculating the Mahalanobis distance between the actual state vector and the theoretical state vector includes:

[0061] The covariance matrix is ​​calculated based on the set of deviation vectors between multiple actual state vectors and their corresponding theoretical state vectors under historical normal operating conditions. ;

[0062] Using the inverse of the covariance matrix Calculate the Mahalanobis distance between the current actual state vector and the theoretical state vector. , calculate Compare with a confidence threshold set based on a statistical distribution.

[0063] Preferably, the step of non-linearly increasing the collaborative cost of the edges corresponding to the task sequence segment in the service graph through a preset weight update function, based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the occurrence of the anomaly, includes:

[0064] For each edge in the task sequence segment corresponding to the occurrence of the anomaly, the new collaborative cost is... From the original collaborative cost The updated formula is: ;

[0065] in, For the calculated Mahalanobis distance, The preset Mahalanobis distance threshold, and These are preset normal values, which control the basic magnitude and growth rate of the cost increase, respectively.

[0066] This invention addresses the problems of poor interoperability and difficult collaborative control among multiple heterogeneous devices within an intelligent manufacturing unit due to inconsistent interfaces and protocols. At the planning level, inputting a spatiotemporal tensor to assess the execution risk of candidate paths avoids operational conflicts between multiple devices, thereby selecting the optimal task sequence that balances efficiency and safety, improving the reliability of collaborative execution. At the execution level, calculating the Mahalanobis distance between theoretical and actual states accurately monitors execution deviations. When anomalies occur, a compensation path can be quickly planned from the current state to complete the task, ensuring production continuity and fault tolerance. Feedback on the degree of execution anomalies is used to update the collaborative cost in the service graph, enabling the system to avoid high-risk collaborative links in subsequent planning, achieving continuous improvement of the control strategy and enhancing the overall performance of the system under complex operating conditions. Attached Figure Description

[0067] Figure 1 A flowchart of the first embodiment;

[0068] Figure 2 A schematic diagram illustrating the generation of candidate collaborative paths for reverse search;

[0069] Figure 3 This diagram illustrates the generation and comparison of theoretical and actual state vectors.

[0070] Figure 4 This is a schematic diagram of anomaly detection based on Mahalanobis distance. Detailed Implementation

[0071] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0072] In the first embodiment, the present invention proposes an edge computing collaborative control method for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit, see [link to relevant documentation]. Figure 1 This includes the following steps:

[0073] S1, Obtain the functional metadata and real-time status data of each heterogeneous device in the intelligent manufacturing unit, and construct a weighted directed acyclic service graph representing the behavioral association between devices based on the unified resource representation framework. The weight of the edge in the graph is initialized as the collaboration cost.

[0074] Specifically, the system actively queries or subscribes to the functional metadata of each device through standard industrial protocols such as OPCUA or MT Connect. For example, it obtains the motion range, end effector type, and service interfaces for grasping and placing of a six-axis robot, and the machining capabilities, tool library information, and service interfaces for starting and stopping of a CNC machine tool. Simultaneously, it collects real-time sensor data and operation logs from each device via MQTT or DDS protocols to obtain real-time status data such as current coordinates, joint angles, spindle speed, and idle status. A unified resource representation framework is defined using JSON, formatting the heterogeneous data into standard services containing fields such as device ID, service name, preconditions, post-effects, and execution time. It iterates through all services; if the post-effect of service A satisfies the precondition of service B, a directed edge is created in the service graph from the node representing service A to the node representing service B. The initial weight of the edge, i.e., the collaboration cost, can be set as the sum of the historical average execution time of service B and the switching time between the two services.

[0075] In an optional embodiment, the construction of a weighted directed acyclic service graph representing the behavioral associations between devices based on the unified resource representation framework, wherein the weights of the edges in the graph are initialized as collaboration costs, including:

[0076] The robotic arms, CNC machine tools, and automated guided vehicles within the intelligent manufacturing unit are abstracted as nodes in the diagram;

[0077] The processing, handling, and testing services provided by the equipment are abstracted as directed edges connecting the nodes;

[0078] After normalizing the historical average execution time, rated energy consumption, and historical failure rate of each service, the collaborative cost C of the edge is initialized by weighted summation, and the calculation formula is as follows: ;

[0079] Wherein, T, E, and F are the historical average execution time, rated energy consumption, and historical failure rate of the service, respectively;

[0080] , , This is the corresponding normalized baseline value;

[0081] , , These are the preset weighting coefficients.

[0082] Physical devices are mapped to graphical nodes. For example, a six-axis robotic arm RA01, a five-axis CNC machine tool CNC01, and an automated guided vehicle (AGV01) become three independent nodes in the service graph. Their functions are transformed into directed edges. For instance, the complete action of robotic arm RA01 picking up a workpiece from the hopper and placing it onto machine tool CNC01 constitutes a directed edge from node RA01 to node CNC01, representing the flow of materials and the continuity of services.

[0083] Taking the service of placing a workpiece from the robotic arm RA01 to the CNC01 machine tool as an example, assuming the historical average execution time T is 5 seconds, the rated energy consumption E is 0.1 kW·h, and the historical failure rate F is 0.5%, the normalized baseline value for the entire unit is set to time. It's 20 seconds, energy consumption It is 1 kWh, and the failure rate is... It's 5%. Then set the time weight. The energy consumption weight is 0.5. The failure rate weight is 0.3. The initial coordination cost of this service edge is calculated as C = 0.175 according to the formula. This value of 0.175 is the comprehensive cost of performing this service, and each edge in the graph is initialized in this way.

[0084] S2, Receive production tasks and parse them into target states in the service graph. Use a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate collaborative paths. For each candidate collaborative path, construct a spatiotemporal tensor representing the resource occupancy of each device and calculate the collaborative execution risk coefficient. Select the path with the smallest collaborative execution risk coefficient as the optimal collaborative task sequence.

[0085] Specifically, the Manufacturing Execution System (MES) receives instructions such as "Produce part P001." These instructions are parsed into one or more state nodes in the service graph, for example, a service node representing that part P001 has been processed and placed on the finished product conveyor belt. Starting from this target state node, the A* search algorithm is applied for reverse backtracking. All parent nodes that can directly lead to the target state, i.e., preceding service nodes, are found, and the cost of reaching the target node from these parent nodes is evaluated. For each newly discovered parent node, this process is repeated, continuously backtracking upstream until a node representing the current idle or initial state of all relevant equipment is reached. During this process, if multiple service combinations can achieve a certain intermediate state, the algorithm explores these branches in parallel, generating multiple candidate cooperative paths composed of service sequences that can achieve the production task objective from the current equipment state.

[0086] For a candidate collaborative path, such as robot A picking up material, moving to machine tool B, placing the material, and machine tool B closing its door for processing, a four-dimensional spatiotemporal tensor is established. Three dimensions represent the physical space grid of the intelligent manufacturing unit, and one dimension represents time. Based on the estimated execution time of each service in the path and the geometric model of the equipment, it is calculated which spatial grids, such as the arm span of robot A and the protective door of machine tool B, will occupy at each time slice. The value of these grids is set to 1, and the rest to 0. After filling the entire path, the spatiotemporal tensor is checked. If the value of any spatiotemporal coordinate point is greater than 1, it indicates a spatiotemporal conflict, such as the robot's path overlapping with the path of another running AGV at a certain moment. The collaborative execution risk coefficient is weighted based on the number of conflict points, the importance of the equipment involved, and the estimated conflict duration. The risk coefficients of all candidate collaborative paths are compared, and the path with the coefficient value of zero or the smallest is selected as the optimal collaborative task sequence.

[0087] In an optional embodiment, the step of using a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate cooperative paths includes:

[0088] Using the target state as the starting point, the preceding service nodes that can reach the state are traversed in reverse in the service graph.

[0089] The reverse traversal process is continuously iterated until the starting node of the search path matches the current device state, thus obtaining a complete candidate cooperative path;

[0090] Repeat this process until all paths that meet the preset path length or cost limits are found.

[0091] Assuming the target state is that workpiece A has been processed and stored in the finished product area, we search the service graph for service nodes that can lead to this state, such as the AGV01's inbound service. Taking the preceding state of the AGV01 inbound service, i.e., workpiece A is on AGV01, as a new sub-target, we continue backtracking to find the service that can realize this state: the CNC machine tool CNC01 unloads the processed workpiece A onto AGV01.

[0092] The above iterative process continuously traces back, sequentially experiencing the machining of workpiece A on CNC01, and the subsequent services provided by robotic arm RA01, such as loading workpiece A from the blank area to CNC01. When the path's starting point is traced back to the initial state of the current intelligent manufacturing unit—that is, workpiece A in the blank area—and perfectly matches, a complete collaborative path from the initial state to the target state is generated. If multiple devices or process options exist within the unit, such as another machine tool CNC02, the algorithm explores all possible path combinations and filters them based on constraints such as total collaborative cost or total time, generating a candidate path set containing multiple feasible solutions. See [link to relevant documentation]. Figure 2 .

[0093] In an optional embodiment, the construction of a spatiotemporal tensor representing the resource occupancy of each device and the calculation of a cooperative execution risk coefficient include:

[0094] Construct a spatiotemporal tensor T with device ID as the first dimension, time step as the second dimension, and resource type as the third dimension;

[0095] For each task step in the candidate collaborative path, the estimated resource consumption of the step is filled into the corresponding position of the tensor T;

[0096] The collaborative execution risk coefficient R is calculated by weighted summation of all resource occupancy points exceeding a preset safety threshold, using the following formula:

[0097]

[0098] in, This represents the resource usage value of type k for device i at time step j. The rated capacity of resource type k for device i, For the risk weight of type k resources, The security threshold for type k resources. For indicator functions, when the condition is true, that is... The value is 1 if it is true, and 0 otherwise.

[0099] Construct a three-dimensional spatiotemporal tensor model. The first dimension is the device index, for example, 1 represents the robotic arm RA01, and 2 represents the machine tool CNC01. The second dimension is the discrete time step, for example, with a step size of 0.1s. The third dimension is the resource type index, for example, 1 represents CPU utilization, and 2 represents tool wear. For a machining step performed by machine tool CNC01 in a candidate path, assuming that the step lasts from the 50th time step to the 150th time step, and the estimated CPU utilization will remain at 85% and the tool wear will increase by 0.02 units during this period, then in the tensor T, set the value of the element with device index 2, time step from 50 to 150, and resource index 1 to 0.85, and update the value of the element with resource index 2 accordingly.

[0100] Assuming the rated capacity of the CNC01 machine tool's CPU A value of 1.0, or 100%, is the safety threshold. Set to 0.9, which is 90%, for risk weighting. The value is 0.4. At a certain time step j=100, if the estimated CPU usage value is... The value is 0.95, exceeding the safety threshold of 0.9, indicating the function... The value is set to 1. At this point, the risk value generated by this point is 0.38. The collaborative execution risk coefficient R is the sum of the risk values ​​generated by all devices at all time steps and at all resource occupancy points exceeding their respective safety thresholds throughout the entire path execution process.

[0101] S3, Distribute control commands to each device according to the optimal collaborative task sequence, and synchronously predict the theoretical state vector of the sequence execution; Collect the actual state data fed back by each device and fuse them into the actual state vector of the unit;

[0102] Specifically, the optimal collaborative task sequence, such as robot grasping, moving, and placing, is decomposed into atomic instructions that each device can directly execute. Grasping instructions are converted into native code for the robot controller and issued via dedicated device drivers or API interfaces. Moving instructions are decomposed into a series of path point posture data streams and sent to the robot. Simultaneously with instruction issuance, a built-in digital twin model or kinematic model begins synchronous simulation, predicting the robot's joint angles, end-effector coordinates, and machine tool spindle speeds millisecond by millisecond based on the instruction content and device dynamic parameters. These predicted values ​​are then combined into a high-dimensional theoretical state vector. Simultaneously, edge nodes rapidly acquire encoder readings from the robot and sensor data from the machine tool. A Kalman filter algorithm is used to process and align the multi-source heterogeneous actual measurements, fusing them into an actual state vector with the same dimension and structure as the theoretical state vector. (See [link to relevant documentation]). Figure 3 .

[0103] In an optional embodiment, the step of distributing control commands to each device according to the optimal cooperative task sequence and synchronously predicting the theoretical state vector of the sequence execution includes:

[0104] A digital twin model of the intelligent manufacturing unit is established, and simulation is performed in the model according to the control commands distributed to each device;

[0105] At each preset time sampling point, key state parameters of each virtual device in the model are extracted, and the parameters are arranged and combined in a predetermined order to obtain the theoretical state vector. .

[0106] A digital twin simulation environment, precisely corresponding to the physical units, is used. This environment contains accurate kinematic and dynamic models of each device; for example, the virtual model of the robotic arm RA01 has the same joint structure and motion constraints as the physical entity. Once a cooperative path is determined, the corresponding sequence of underlying control commands, such as angle control commands sent to the robotic arm and spindle speed commands sent to the machine tool, is synchronously input into the digital twin model. The model performs virtual operation according to the commands, simulating the expected response in the physical world. Snapshots of the virtual environment are taken at a fixed frequency, e.g., 10 times per second. At each sampling moment, the current angle values ​​of the six joints are read from the virtual robotic arm model, the spindle speed and tool feed rate are read from the virtual CNC machine tool model, and the two-dimensional plane coordinates are read from the virtual automated guided vehicle model. These sampled values—e.g., six angle values, two speed values, and two coordinate values—are concatenated in a predefined fixed order to form a 10-dimensional column vector, which is the theoretical state vector at that moment. , which represents the state of the system under ideal, undisturbed conditions.

[0107] In an optional embodiment, the step of collecting the actual state data fed back by each device and fusing it into the actual state vector of the unit includes:

[0108] The system periodically reads all joint angles from the robotic arm controller, spindle speed and feed rate from the CNC machine tool system, and real-time coordinates from the automated guided vehicle positioning system.

[0109] All acquired data are concatenated into a one-dimensional column vector according to the same order and dimensions as the theoretical state vector, which serves as the actual state vector of the unit. .

[0110] The real data acquisition process, executed synchronously within the physical manufacturing unit, sends data requests to the controllers of field devices at the same frequency as the digital twin sampling via the Industrial Ethernet protocol. For example, the controller connected to the robotic arm RA01 reads real-time angle data of the six joints fed back by the encoder, such as 1.57 radians and -0.8 radians. Simultaneously, it queries the current actual spindle speed, such as 2998 revolutions per minute, and the actual feed rate, such as 499 mm / min, from the CNC system of the CNC machine tool CNC01 via the OPCUA protocol.

[0111] The position of the AGV01 in the workshop coordinate system is obtained from the LiDAR or UWB positioning module mounted on it, such as X coordinate 12.34 meters and Y coordinate 5.67 meters. After obtaining all the raw data, data alignment and fusion are performed. This ensures that the units and order of all data match the aforementioned theoretical state vector. The structures are completely identical. For example, the read angle, velocity, coordinates, and other data are arranged in the same 10-dimensional order and combined into a single column vector, which is the actual state vector of the current unit. This reflects the operating status of the physical system.

[0112] S4. Calculate the Mahalanobis distance between the actual state vector and the theoretical state vector. When the distance exceeds a preset threshold, freeze the execution of the current task sequence. Using the current actual state vector as a new starting point, replan the compensation path to the target state. Based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, increase the collaborative cost of the edge corresponding to the task sequence segment in the service graph nonlinearly through a preset weight update function.

[0113] Specifically, based on historical normal operation data, the covariance matrix among the variables in the state vector is calculated. During task execution, the Mahalanobis distance between the current actual state vector and the theoretical state vector is calculated in real time. If this distance is greater than the threshold of the 99% confidence interval calculated based on historical data, it is judged as an anomaly, such as the robot's actual movement trajectory deviating from the theoretical trajectory due to excessive load. At this time, the edge controller immediately sends an emergency stop or pause command to the robot and related equipment, freezing the task. The currently fused actual state vector is mapped to a new starting node in the service graph, keeping the original task's target state node unchanged, and the inverse search algorithm is called again to plan a compensation path to recover from the current abnormal state and continue completing the task. At the same time, the service being executed when the anomaly occurred is located, i.e., a certain edge in the service graph. Based on the magnitude of the Mahalanobis distance exceeding the threshold, an exponential function is used as the weight update function to increase the collaboration cost of that edge. The greater the Mahalanobis distance exceeds the threshold, the greater the increase in collaboration cost, so that subsequent task planning automatically avoids this collaborative link that has been proven to have low reliability.

[0114] In an optional embodiment, calculating the Mahalanobis distance between the actual state vector and the theoretical state vector includes:

[0115] The covariance matrix is ​​calculated based on the set of deviation vectors between multiple actual state vectors and their corresponding theoretical state vectors under historical normal operating conditions. ;

[0116] Using the inverse of the covariance matrix Calculate the current actual state vector With the theoretical state vector Mahalanobis distance between , calculate Compare with a confidence threshold set based on a statistical distribution.

[0117] Under normal operating conditions, it continuously collects a large number of actual state vectors. and theoretical state vector Yes. For each pair of vectors, calculate the deviation vector. The deviation vector reflects the random fluctuations and measurement noise of the system during normal operation. Using a large number of collected deviation vector samples, a covariance matrix can be calculated. The matrix represents the statistical correlation between deviations of different state variables, such as whether a small angular deviation of one joint of a robotic arm is usually accompanied by a deviation of another joint. During the online monitoring phase, the current... and And calculate the current deviation vector. This is done using the inverse of the pre-calculated covariance matrix. Through formula

[0118]

[0119] The calculation is performed. Essentially, the formula transforms the deviation vector to a new coordinate space where all dimensions are uncorrelated and have a variance of 1, and then calculates the Euclidean distance. The resulting Mahalanobis distance... It is an anomaly measure that comprehensively considers the correlation between variables and the magnitude of their individual fluctuations. The calculated... A value, such as 5.8, is compared with a threshold determined based on the chi-square distribution and a preset confidence level, such as 99.9%. If the value exceeds the threshold, a system anomaly is determined. Figure 4 .

[0120] In an optional embodiment, the step of non-linearly increasing the collaborative cost of the edges corresponding to the task sequence segment in the service graph through a preset weight update function, based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the occurrence of the anomaly, includes:

[0121] For each edge in the task sequence segment corresponding to the occurrence of the anomaly, the new collaborative cost is... From the original collaborative cost The updated formula is: ;

[0122] in, For the calculated Mahalanobis distance, The preset Mahalanobis distance threshold, and These are preset normal values, which control the basic magnitude and growth rate of the cost increase, respectively.

[0123] When Mahalanobis distance is detected Exceeding the threshold When the anomaly occurs, the system locates the task step being executed. Assuming the anomaly occurs during the process of robotic arm RA01 placing a workpiece onto machine tool CNC01, the system will lock the service edge in the service graph pointing from node RA01 to node CNC01, preparing to perform a penalty update on the collaboration cost. The collaboration cost of this edge before the update is... For example, it is 0.175.

[0124] Apply the cost update formula to calculate the new collaborative cost. Assuming Mahalanobis distance threshold The value is 4.5, while the currently detected abnormal Mahalanobis distance is... Version 6.0. Preset update parameters. It is 0.1. The value is 0.8. Substituting this into the formula, the penalty term is calculated to be 0.232. The new collaboration cost... The cost will be the original cost plus the aforementioned penalty term, which is 0.407. Due to the characteristics of the exponential function, the greater the Mahalanobis distance exceeds the threshold, the greater the increase in cost. The update increases the execution cost of this service edge, and in future task planning, it will tend to avoid this path that has recently experienced anomalies, thereby achieving risk avoidance and self-optimization.

[0125] In a second embodiment, the present invention also provides an edge computing collaborative control system for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit, comprising the following modules:

[0126] The acquisition module is used to acquire the functional metadata and real-time status data of each heterogeneous device in the intelligent manufacturing unit, and to construct a weighted directed acyclic service graph representing the behavioral association between devices based on the unified resource representation framework. The weight of the edges in the graph is initialized as the collaboration cost.

[0127] The construction module is used to receive production tasks and parse them into target states in the service graph. It uses a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate collaborative paths. For each candidate collaborative path, it constructs a spatiotemporal tensor representing the resource occupancy of each device and calculates the collaborative execution risk coefficient. The path with the smallest collaborative execution risk coefficient is selected as the optimal collaborative task sequence.

[0128] The fusion module is used to distribute control commands to each device according to the optimal collaborative task sequence, and synchronously predict the theoretical state vector of the sequence execution; collect the actual state data fed back by each device, and fuse them into the actual state vector of the unit.

[0129] The calculation module is used to calculate the Mahalanobis distance between the actual state vector and the theoretical state vector. When the distance exceeds a preset threshold, the execution of the current task sequence is frozen, and a compensation path to the target state is replanned with the current actual state vector as the new starting point. Based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, the collaborative cost of the edge corresponding to the task sequence segment in the service graph is nonlinearly increased through a preset weight update function.

[0130] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0131] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0132] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An edge computing collaborative control method for intelligent manufacturing unit multi-device heterogeneous interconnection, characterized in that, Includes the following steps: Obtain the functional metadata and real-time status data of each heterogeneous device in the intelligent manufacturing unit, and construct a weighted directed acyclic service graph representing the behavioral association between devices based on the unified resource representation framework. The weight of the edge in the graph is initialized as the collaboration cost. The system receives production tasks and parses them into target states in the service graph. It then uses a reverse search algorithm to backtrack from the target state to the current device state, generating multiple candidate collaborative paths. For each candidate collaborative path, a spatiotemporal tensor representing the resource occupancy of each device is constructed and the collaborative execution risk coefficient is calculated. The path with the smallest collaborative execution risk coefficient is selected as the optimal collaborative task sequence. Control commands are distributed to each device according to the optimal collaborative task sequence, and the theoretical state vector of the sequence execution is predicted synchronously; the actual state data fed back by each device is collected and fused into the actual state vector of the unit. Calculate the Mahalanobis distance between the actual state vector and the theoretical state vector. When the distance exceeds a preset threshold, freeze the execution of the current task sequence. Using the current actual state vector as a new starting point, replan the compensation path to the target state. Based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, increase the collaborative cost of the edge corresponding to the task sequence segment in the service graph nonlinearly through a preset weight update function.

2. The method of claim 1, wherein, The weighted directed acyclic service graph representing the behavioral relationships between devices is constructed based on the unified resource representation framework. The weights of the edges in the graph are initialized as collaboration costs, including: The robotic arms, CNC machine tools, and automated guided vehicles within the intelligent manufacturing unit are abstracted as nodes in the diagram; The processing, handling, and testing services provided by the equipment are abstracted as directed edges connecting the nodes; After normalizing the historical average execution time, rated energy consumption, and historical failure rate of each service, the collaborative cost C of the edge is initialized by weighted summation.

3. The method of claim 1, wherein, The reverse search algorithm is used to backtrack from the target state to the current device state, generating multiple candidate cooperative paths, including: Using the target state as the starting point, the preceding service nodes that can reach the target state are traversed in reverse in the service graph. The reverse traversal process is continuously iterated until the starting node of the search path matches the current device state, thus obtaining a complete candidate cooperative path; Repeat this process until all paths that meet the preset path length or cost limits are found.

4. The method of claim 1, wherein, The construction of the spatiotemporal tensor representing the resource occupancy of each device and the calculation of the collaborative execution risk coefficient include: Construct a spatiotemporal tensor T with device ID as the first dimension, time step as the second dimension, and resource type as the third dimension; For each task step in the candidate collaborative path, the estimated resource consumption of the step is filled into the corresponding position of the tensor T; The collaborative execution risk coefficient R is calculated by weighted summation of all resource occupancy points exceeding a preset safety threshold, using the following formula: ； wherein, is the k-type resource occupation value of device i at time step j, is the k-type resource nominal capacity of device i, is the risk weight of k-type resource, is the safety threshold of k-type resource, I() is an indicator function that takes the value 1 when the condition holds, i.e. , and 0 otherwise.

5. The method of claim 1, wherein, The step of distributing control commands to each device according to the optimal collaborative task sequence and synchronously predicting the theoretical state vector for the execution of the sequence includes: A digital twin model of the intelligent manufacturing unit is established, and simulation is performed in the model according to the control commands distributed to each device; At each preset time sampling point, key state parameters of each virtual device in the model are extracted, and the parameters are arranged and combined in a predetermined order to obtain the theoretical state vector.

6. The method of claim 5, wherein, The actual state data collected from each device is fused into the actual state vector of the unit, including: The system periodically reads all joint angles from the robotic arm controller, spindle speed and feed rate from the CNC machine tool system, and real-time coordinates from the automated guided vehicle positioning system. All the acquired data are concatenated into a one-dimensional column vector according to the same order and dimensions as the theoretical state vector, which serves as the actual state vector of the unit.

7. The method of claim 6, wherein, The calculation of the Mahalanobis distance between the actual state vector and the theoretical state vector includes: The covariance matrix S is calculated based on the set of deviation vectors between multiple actual state vectors and their corresponding theoretical state vectors under historical normal operating conditions. using the inverse of the covariance matrix , calculating Mahalanobis distance D between the current actual state vector and the theoretical state vector M , comparing the calculated D M with a confidence threshold set based on the statistical distribution.

8. The method of claim 7, wherein, The step of non-linearly increasing the collaborative cost of the edges corresponding to the task sequence segment in the service graph through a preset weight update function, based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, includes: For each edge in the task sequence segment corresponding to the occurrence of the anomaly, the new collaborative cost is... From the original collaborative cost The updated formula is: ; Among them, D T The preset Mahalanobis distance threshold, and These are preset normal values, which control the basic magnitude and growth rate of the cost increase, respectively.

9. An edge computing collaborative control system for heterogeneous interconnection of multiple devices in an intelligent manufacturing unit, characterized in that, Includes the following modules: The acquisition module is used to acquire the functional metadata and real-time status data of each heterogeneous device in the intelligent manufacturing unit, and to construct a weighted directed acyclic service graph representing the behavioral associations between devices based on the unified resource representation framework. The weight of the edges in the graph is initialized as the collaboration cost. The construction module is used to receive production tasks and parse them into target states in the service graph. It uses a reverse search algorithm to backtrack from the target state to the current device state and generate multiple candidate collaborative paths. For each candidate collaborative path, a spatiotemporal tensor representing the resource occupancy of each device is constructed and the collaborative execution risk coefficient is calculated. The path with the smallest collaborative execution risk coefficient is selected as the optimal collaborative task sequence. The fusion module is used to distribute control commands to each device according to the optimal collaborative task sequence, and synchronously predict the theoretical state vector of the sequence execution; collect the actual state data fed back by each device, and fuse them into the actual state vector of the unit. The calculation module is used to calculate the Mahalanobis distance between the actual state vector and the theoretical state vector. When the distance exceeds a preset threshold, the execution of the current task sequence is frozen, and a compensation path to the target state is replanned with the current actual state vector as the new starting point. Based on the magnitude of the Mahalanobis distance and the task sequence segment corresponding to the anomaly, the collaborative cost of the edge corresponding to the task sequence segment in the service graph is nonlinearly increased through a preset weight update function.

10. The system according to claim 9, characterized in that, The weighted directed acyclic service graph representing the behavioral relationships between devices is constructed based on the unified resource representation framework. The weights of the edges in the graph are initialized as collaboration costs, including: The robotic arms, CNC machine tools, and automated guided vehicles within the intelligent manufacturing unit are abstracted as nodes in the diagram; The processing, handling, and testing services provided by the equipment are abstracted as directed edges connecting the nodes; After normalizing the historical average execution time, rated energy consumption, and historical failure rate of each service, the collaborative cost C of the edge is initialized by weighted summation.

Images