An agent-based industrial robot task dynamic scheduling control method

Abstract

The application relates to the field of industrial automation control, in particular to an Agent-based industrial robot task dynamic scheduling control method. The method constructs a multi-dimensional resource space containing space, time, function mode and physical field intensity, uses dimensional mask to quantize the perspective difference of heterogeneous Agents to the same resource; uses a physical field propagation model to organize source and sink nodes into hyperedges, constructs a resource coupling hypergraph to explicitly express implicit cross-dimensional conflicts; extracts a transfer coefficient from the hypergraph to construct a coupling matrix and performs low-rank decomposition, converts task requirements into a unified resource request vector, superimposes the total occupation vector to calculate a global conflict index, and finds coupling conflicts in the allocation decision stage; when the index is over the limit, continuously adjust the Agent running parameters to actively eliminate conflicts and issue scheduling instructions without implicit conflicts. Thus, the detection is moved to the decision layer, conflicts are actively eliminated, the heterogeneous Agent mixing density is improved, and dead lock and shutdown are avoided.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation control, and in particular relates to an agent-based method for dynamic scheduling and control of industrial robot tasks. Background Technology

[0002] In the field of intelligent manufacturing, automated production lines widely employ heterogeneous autonomous agents (such as automated guided vehicles, multi-joint robotic arms, and fixed vision inspection stations) to collaboratively complete logistics, processing, and quality inspection tasks. These agents share limited physical space and resources such as machine tools and aisles, but due to differences in their structure and task nature, their demands for the same resources have fundamentally different dimensional perspectives: AGVs express a single transport as a two-dimensional occupancy vector of spatial location and time window; robotic arms express operations as a three-dimensional demand vector of workstation coordinates, functional modes, and end effector torque; and inspection equipment expresses imaging tasks as a three-dimensional constraint vector of trigger time, resolution level, and tolerable vibration limit. This dimensional heterogeneity results in their respective resource requests belonging to different mathematical subspaces, making direct unified conflict judgment impossible.

[0003] Existing scheduling systems generally rely on single-dimensional mutual exclusion detection, such as checking only spatial geometric interference or temporal overlap, assuming safety as long as there is no spatial or temporal conflict. However, significant cross-dimensional physical field coupling effects exist in real manufacturing environments: when a robotic arm operates in high-torque mode, the broadband vibration generated by its joint reducer propagates through the machine tool bed and foundation to a distant inspection station. Even if there is no overlap in time and space, the amplitude of environmental vibration at the moment of inspection may still exceed the threshold allowed by precision measurement, causing image blurring and misjudgment of dimensions. Similarly, when an AGV travels along aisles, its navigation safety envelope may intrude into the robotic arm's dynamic working envelope. Even if the nominal paths do not intersect, aerodynamic disturbances and emergency braking can induce collision risks. These conflicts are not direct preemption in any single dimension, but rather implicit conflicts caused by cross-dimensional transmission of physical fields. Existing scheduling methods cannot identify these conflicts during the resource allocation phase, creating a blind spot of implicit conflicts. Such conflicts also exhibit time-varying transitivity. The resource consumption of a certain agent affects neighboring agents through physical field coupling. The latter adjusts its own state and then transmits the interference to downstream devices, ultimately leading to global scheduling instability and unplanned shutdowns.

[0004] Therefore, there is an urgent need for a dynamic scheduling and control method that can proactively detect and resolve cross-dimensional implicit conflicts during the resource allocation decision-making stage. Summary of the Invention

[0005] Therefore, it is necessary to provide an agent-based dynamic scheduling and control method for industrial robot tasks to address the aforementioned technical problems.

[0006] This application provides an agent-based method for dynamic scheduling and control of industrial robot tasks, including:

[0007] S1. Based on the independently allocable resource units in the production line, construct a global resource space including spatial dimension, time dimension, functional mode dimension and physical field intensity dimension, and instantiate the resource units as resource nodes in the global resource space.

[0008] S2. Based on the resource usage characteristics of each type of Agent, generate a dimension mask to identify the dimensions that the corresponding Agent needs to specify when generating resource requests.

[0009] S3. Based on the physical field propagation model, by determining the cross-dimensional coupling relationship between resource nodes caused by the propagation of the physical field, the source nodes that are the radiation sources of the physical field and the sink nodes affected by the physical field are organized into hyperedges to obtain the resource coupling hypergraph.

[0010] S4. Based on the transmission relationship from the source node to the sink node recorded by each hyperedge in the resource coupling hypergraph, generate a coupling matrix whose matrix elements represent the physical field transmission coefficients from the source node to the sink node; perform low-rank decomposition on the coupling matrix to obtain multiple sets of low-rank factor vectors, and combine the multiple sets of low-rank factor vectors into a low-rank factor matrix.

[0011] S5. Receive the task to be scheduled, and generate the resource request vector of each target agent using a dimension mask according to the Agent type and task parameters required by the task to be scheduled; wherein, the elements of the resource request vector represent the demand for the corresponding resource node.

[0012] S6. Combine the target agents to obtain the target agent scheme, and superimpose the resource request vectors of each agent in the target agent scheme to form the total occupancy vector; call the low-rank factor matrix to calculate the global conflict index of the target agent scheme based on the total occupancy vector; where the global conflict index measures the degree of conflict caused by cross-dimensional coupling in the target agent scheme.

[0013] S7. When the global conflict index exceeds the index threshold, adjust the adjustable running parameters corresponding to the elements of the resource request vector of at least one Agent in the target Agent scheme until the recalculated global conflict index does not exceed the index threshold, and obtain the final total occupancy vector; generate scheduling instructions based on the target Agent scheme and the corresponding final total occupancy vector and issue them to the corresponding Agent.

[0014] The aforementioned agent-based dynamic scheduling and control method for industrial robots constructs a global resource space encompassing spatial, temporal, functional mode, and physical field strength dimensions, instantiating production line resource units as resource nodes within this space. Based on this, dimensional masks are generated for various heterogeneous agents to identify the dimensions that impose substantial constraints when generating resource requests. This quantifies and distinguishes the essential differences in dimensional perspectives between different agents such as AGVs, robotic arms, and vision inspection devices regarding the same physical resource. Subsequently, a physical field propagation model is used to determine the cross-dimensional coupling relationships between resource nodes caused by the propagation of physical fields such as vibration and electromagnetic fields. Nodes acting as radiation sources and their affected sink nodes are organized into hyperedges to construct a resource coupling hypergraph. This allows implicit conflicts that cannot be detected by single-dimensional detection (such as robotic arm vibration leading to loss of detection equipment accuracy, AGV passage intrusion into the robotic arm's working envelope, etc.) to be explicitly expressed as structured relationships within the hypergraph. Furthermore, physical field transmission coefficients are extracted from the hypergraph to construct a coupling matrix, which is then decomposed into a low-rank factor matrix, providing pre-compressed coupling information for the online scheduling phase. Upon receiving tasks to be scheduled online, the task requirements of each target agent are uniformly transformed into a one-dimensional resource request vector based on the same resource node space using a dimensionality mask. Then, by combining target agents and superimposing resource request vectors into a total occupancy vector, a low-rank factor matrix is ​​used to quickly calculate a global conflict index. This allows for real-time quantification of the degree of cross-dimensional coupling conflict in the resource allocation decision-making stage. If the global conflict index exceeds a threshold, the allocation scheme is not directly rejected. Instead, the adjustable operating parameters of the relevant agents are continuously adjusted to proactively resolve latent conflicts until the conflict index meets the constraints, at which point a scheduling instruction is generated and executed. This technical solution moves the detection of latent conflicts from the traditional post-deadlock formation stage to the task allocation decision-making stage. By continuously adjusting parameters, it empowers the scheduling system with the ability to proactively resolve cross-dimensional conflicts. This allows for high-density, secure hybridization of heterogeneous agents within a more compact spatiotemporal window without altering task semantics. This fundamentally avoids deadlocks, quality incidents, and unplanned downtime caused by cross-dimensional coupling in the physical field, significantly improving the stability of production line scheduling and equipment utilization. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1A flowchart illustrating an agent-based dynamic scheduling and control method for industrial robot tasks provided by this invention;

[0017] Figure 2 This is a schematic diagram of the process of constructing a resource coupling hypergraph in an optional embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram illustrating the process of calculating the global conflict index of the target Agent scheme in an optional embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] The implementing entity of this invention is a scheduling controller deployed at the edge of a smart manufacturing production line. This scheduling controller can be an industrial edge computing server with real-time data processing capabilities, or an edge cloud platform composed of multiple distributed computing nodes. The scheduling controller establishes communication connections with the local controllers of each agent on the production line via an industrial real-time Ethernet (e.g., EtherCAT, Profinet, or Time-Sensitive Networking (TSN)). Specifically, the AGV's onboard embedded controller, the robotic arm's servo control cabinet, and the industrial control computer of the vision inspection equipment all act as communication nodes connected to this industrial Ethernet, reporting real-time status data to the scheduling controller and receiving scheduling instructions. Furthermore, the scheduling controller is also connected to the production line's digital twin server via a high-speed data bus (e.g., OPC UA over TSN) to subscribe to real-time physical field data collected by accelerometers, current transformers, and visual positioning systems distributed on the production line. The digital twin server runs a real-time simulation model of the production line, responsible for fusing and processing the raw sensor data into estimated physical field strength values ​​for each monitoring point, and publishing these estimates to the scheduling controller in a standardized data structure. The scheduling controller internally runs a multi-threaded real-time task scheduling engine. The main thread is responsible for executing the dynamic scheduling control method of this invention within each scheduling cycle, while another thread is responsible for asynchronously receiving and processing the physical field update data stream from the digital twin server. The Manufacturing Execution System (MES) issues job tasks to the scheduling controller through the Enterprise Service Bus (ESB). The task message includes information such as task type, target workstation, process parameters, and expected completion time window.

[0021] In this application environment, the data flow is as follows: MES sends the tasks to be scheduled to the scheduling controller; the scheduling controller loads static data such as the global resource node table, dimension mask configuration table, resource coupling hypergraph, and low-rank factor matrix from its internally maintained offline database; the scheduling controller generates resource request vectors for each candidate agent based on the task parameters, combines candidate schemes and calculates the global conflict index, performs parameter adjustment when necessary, and finally generates scheduling instructions without implicit conflicts and sends them to the local controllers of each agent; the local controllers execute the instructions, and at the same time, sensor data is fed back to the scheduling controller via the digital twin server, forming a closed-loop monitoring and rescheduling.

[0022] The following seven embodiments, each corresponding to a different combination of steps at different levels in the claimed method, provide a detailed and progressively in-depth explanation of the technical solution of the present invention.

[0023] Example 1:

[0024] refer to Figure 1 The document presents a flowchart illustrating an agent-based dynamic scheduling and control method for industrial robot tasks, which includes the following steps:

[0025] S1. Based on the independently allocable resource units in the production line, construct a global resource space including spatial dimension, time dimension, functional mode dimension and physical field strength dimension, and instantiate the resource units as resource nodes in the global resource space.

[0026] Specifically, during the offline preparation phase, the scheduling controller first abstracts and models all physical resource units within the production line that can be independently occupied or manipulated by an agent. Let the global resource space be a six-dimensional mathematical space, with the total number of dimensions denoted as... ,Pick The dimensions are, in order: three spatial rectangular coordinate dimensions. , , Used to describe the location of resource units in the physical space of the production line; a time dimension Based on a unified scheduling time axis, it is discretized into a sequence of time slots of equal length. The span of each time slot is determined according to the minimum time scale of the production line cycle time and the transient response of the physical field; a functional mode dimension This is used to encode the types of operations that different devices can perform, such as the gripping mode and welding mode of a robotic arm, and the full-width inspection mode and local precision inspection mode of a vision inspection device. The specific encoding is uniformly maintained by the ontology knowledge base; one physical field strength dimension... , is a continuous quantity directly related to the coupling of physical fields. In this embodiment, it is discretized and divided into several intensity intervals according to the magnitude. Each interval corresponds to a nominal intensity value, which is used to represent the level of physical field that the resource node radiates outward or can tolerate when it is occupied.

[0027] The aforementioned six-dimensional space is not a dense, complete grid, but rather a sparse instantiation based on the actual resource units that exist within the production line and can be independently allocated.

[0028] Spatially, only entities with independent resource identifiers are registered as spatial nodes. For example, the clamping station of each machine tool is represented by the center point of its chuck, each independent tunnel is represented by its midline sampling point sequence, and the center point of the camera's field of view at each inspection station is represented. Temporally, time nodes are generated only for future time windows that have been opened for scheduling. The set of time slot numbers is denoted as... ,in This represents the total number of time slots. The span of each time slot physically corresponds to one decision cycle of the scheduling system or the transient response time of the physical field propagation. It is usually taken as the least common multiple of the production line cycle time to ensure that the start and end times of any task can be aligned with the time slot boundaries.

[0029] The functional mode dimension and the physical field strength dimension are instantiated based on the actual capabilities of each device. All valid nodes constitute a node set, and the total number of nodes is denoted as . Each node is composed of a unique six-tuple. The sign, among which , , Representing nodes respectively Coordinates in the spatial dimension Represents a node The time slot number in the time dimension. This indicates the operation type encoding of the section on the functional mode dimension. This represents the nominal intensity value of a node in the physical field intensity dimension, and assigns each node a linear index number from 1 to 6. A bidirectional mapping table is also established to store the correspondence between node indices and six-dimensional coordinates. This bidirectional mapping table supports querying node indices via six-tuples and also supports retrieving their six-dimensional coordinates from node indices, providing an efficient data access interface for the generation and occupancy query of resource request vectors in subsequent steps.

[0030] In addition, to support fast spatial range queries, a three-dimensional KD-tree index can be built for spatial nodes, so that given a spatial location and neighborhood radius, all nearby resource nodes can be quickly retrieved. This is very useful for calculating the distance from the source node to the sink node and the residual physical field intensity in subsequent hypergraph construction.

[0031] S2. Based on the resource usage characteristics of each type of Agent, generate a dimension mask to identify the dimension that the corresponding Agent needs to specify when generating a resource request.

[0032] Specifically, the scheduling controller generates a dimensional mask for each type of heterogeneous agent (e.g., AGV mobile robots, six-axis robotic arms, vision inspection equipment) participating in the scheduling on the production line, based on their resource usage characteristics. The dimensional mask is essentially a six-dimensional binary vector, where the first... When each component is logically true, it indicates that the Agent class takes the dimension as logically true when generating resource requests. It has a substantial constraint, meaning that the specific value of this dimension is determined by the task parameters and the current state of the Agent; its third When a component is set to "false", it means that the dimension is not within the decision-making perspective of this type of Agent, and a preset default value will be uniformly filled in the resource request expression.

[0033] In this embodiment, the dimension mask is generated by pre-setting based on the equipment technical specifications and empirical rules. During transport tasks, AGVs primarily focus on the spatial occupancy of the passageway and the time window for passage. They do not generate specific functional operation modes, and the physical field radiation caused by their operation is determined by their motion state rather than the task specification. Therefore, the spatial dimension in the AGV's dimension mask is... , , and time dimension The identifier is "Requires specification", and the functional mode dimension is... and physical field strength dimension The identifier is "No specification required". When a robotic arm performs tasks such as assembly, welding, and grinding, the spatial coordinates of the workstation, the functional mode to be executed, and the resulting torque or power level (directly affecting the physical field radiation intensity) need to be clearly specified. However, its time dimension requirement is specified by the task time window. Therefore, the spatial dimension in the dimensional mask of the robotic arm is not specified. , , Functional mode dimension and physical field strength dimension The identifier is "Requires specification", time dimension The identifier is "No need to specify". The spatial requirements of visual inspection equipment are determined by its fixed workstation coordinates. The time dimension requirement is reflected in the shooting window, the functional mode requirement in the detection resolution level, and the physical field strength requirement in the setting of the upper limit of vibration tolerance. However, since it typically does not actively generate physical field radiation, the time dimension in its dimensional mask is... Functional mode dimension and physical field strength dimension The identifier is "Requires specification", spatial dimension , , The identifier is set to "No need to specify".

[0034] For dimensions marked "No need to specify", a default value needs to be set for the Agent in that dimension. For example, the default value for an AGV in the functional mode dimension is the code corresponding to "transportation", and the default value in the physical field strength dimension is the nominal radiation intensity value calibrated according to its full-load motion state; the default value for a robotic arm in the time dimension is 0, indicating that its time constraint is entirely determined by the time window specified by the task; the default value for a vision inspection device in the spatial dimension is the coordinates of its fixed installation position. The dimension mask and its corresponding default value are associated and stored as a structured dimension mask configuration record. Specifically, a key-value pair data structure can be used, with the Agent type code as the key and the mask vector and default value vector as the values, for use when generating resource request vectors during the online scheduling phase.

[0035] S3. Based on the physical field propagation model, by determining the cross-dimensional coupling relationship between resource nodes caused by the propagation of the physical field, the source nodes that are the radiation sources of the physical field and the sink nodes affected by the physical field are organized into hyperedges to obtain the resource coupling hypergraph.

[0036] Specifically, the scheduling controller constructs a physical field propagation model to describe how the physical field (primarily represented by mechanical vibration field, but also applicable to electromagnetic fields, temperature fields, etc.) generated when an agent runs on a resource node propagates through media such as the machine tool bed, foundation, and structural components to other resource nodes in the production line, and constitutes implicit interference to agents working on those nodes. This embodiment uses a simplified point source attenuation model, taking the vibration field as an example, assuming the source node... It has a nominal physical field radiation intensity during operation. ,in Indicates the source node The physical field radiation intensity, expressed as the effective value of vibration velocity or acceleration level, is normalized to a dimensionless nominal value. This radiation intensity propagates through space to the sink node. Residual physical field strength at the location It can be expressed by the following formula:

[0037]

[0038] in, Indicates the source node At the remittance node The residual physical field strength generated at the location, with dimensions of Consistent, both being dimensionless values ​​after normalization; This represents the vibration attenuation coefficient of the building materials and structure of the production line, in units of... The values ​​were obtained through offline tapping experiments and transfer function measurements. Indicates the source node Spatial position vector, Indicates sink node Spatial location vector; Indicates the source node With sink node The Euclidean spatial distance between them.

[0039] Based on this propagation model, the scheduling controller first classifies all resource nodes by role. The classification rules are as follows: nodes belonging to the motion actuator class (such as robotic arms and AGV walking parts) and whose physical field strength dimension value is greater than zero are marked as source nodes; nodes belonging to the agent class (such as vision inspection stations and precision measurement stations) with explicit sensitivity requirements to the environmental physical field are marked as sink nodes. Each source node and sink node retains its spatial location coordinates and physical field-related attributes in the global resource space.

[0040] Then, the scheduling controller traverses each source node. The residual physics intensity is calculated for each of all sink nodes. For a given source node... and a certain sink node Calculated With sink node The corresponding physical field tolerance threshold Compare the tolerance threshold. Indicates sink node The upper limit of vibration allowed for normal performance of detection or measurement tasks is determined based on the accuracy class of the detection equipment and the vibration resistance parameters of the optical system, in accordance with... Normalized values ​​of the same dimensions. When Greater than or equal to When multiplied by a safety factor close to 1 (this safety factor can be a positive real number less than but close to 1, used to capture potential interference near the threshold), the source node is determined. With sink node There is a significant cross-dimensional coupling relationship between the sink node. Included in the source node This is the set of coupling effects centered on the sink node. After traversing all sink nodes, if this set of coupling effects is not empty, then the source node... All sink nodes in this set are organized into a hyperedge. This hyperedge is stored as a record in the data structure, containing a source node index and a list of sink node indices, along with the residual physics intensity value corresponding to each sink node. This allows for the subsequent extraction of transmission coefficients. All hyperedges form a set, resulting in a multi-dimensional resource coupling hypergraph. This hypergraph encodes traditional space-time conflicts and cross-dimensional coupling conflicts of physical fields with a unified data structure, enabling implicit conflicts to have an explicit topological expression and laying the data foundation for the subsequent construction of the coupling matrix.

[0041] S4. Based on the transmission relationship from the source node to the sink node recorded by each hyperedge in the resource coupling hypergraph, generate a coupling matrix whose matrix elements represent the physical field transmission coefficients from the source node to the sink node; perform low-rank decomposition on the coupling matrix to obtain multiple sets of low-rank factor vectors, and combine the multiple sets of low-rank factor vectors into a low-rank factor matrix.

[0042] Specifically, the scheduling controller extracts quantified coupling relationships from the constructed resource coupling hypergraph and organizes them into a matrix form that facilitates online algebraic operations. First, a mapping relationship is established from hyperedge records to the coupling matrix. Let the total number of source nodes after role classification be... The total number of sink nodes is For both the source and sink nodes, establish bidirectional mapping tables from their original indexes in the global resource node table to consecutive local indexes starting from 1: For the source node, establish the mapping... ,in For the sink node, establish a mapping. ,in .

[0043] Construct a OK Column coupling matrix .matrix The line, number Column elements Define as source node (Local index) to sink node The physical field transfer coefficient of (local index). The physical meaning of this physical field transfer coefficient is: source node When a unit radiation intensity is generated at a sink node, The attenuation ratio of the residual physical field intensity caused by the location. It is calculated by finding the source node in the hypergraph record. Corresponding superedge sink node residual physical field strength Combined with the source node Its own nominal radiation intensity ,calculate .because and For normalized values ​​of the same dimensions, This is a dimensionless pure attenuation coefficient, ranging from 0 to 1. For source-sink pairs without hyperedge relationships, The resulting coupling matrix It is a highly sparse non-negative matrix whose non-zero elements collectively reflect the intensity relationships of the physical field propagating from each radiation source to each sensitive point in the production line. In actual storage, Using Compressed Sparse Row (CSR) or Coordinate List (COO) format, only the row index, column index, and value of non-zero elements are stored, which greatly saves memory.

[0044] To meet the real-time requirement of millisecond-level conflict detection during the online scheduling phase, the coupling matrix needs to be... To significantly compress the size, this embodiment employs a low-rank approximation method based on eigenvalue decomposition to achieve this goal. First, a real symmetric positive semi-definite matrix is ​​constructed. The dimension of this matrix is Its elements represent the second-order coupling relationship between source nodes. Performing eigenvalue decomposition yields a set of nonnegative eigenvalues ​​and their corresponding orthogonal eigenvectors. Let the eigenvalues ​​be arranged in descending order as follows: ,in , representing the component index of SVD (Singular Value Decomposition) and eigenvalue decomposition; Indicates the first Each eigenvalue represents a coupling energy along a corresponding eigendirection. Let and The corresponding feature vector is ,in Indicates the first These eigenvectors constitute eigenvectors. A set of orthonormal bases in space. Based on the mathematical relationship between singular value decomposition and eigenvalue decomposition of a matrix, the index of the same component... Corresponding eigenvalues Singular values Right singular vector and left singular vector Components belonging to the same SVD are related to each other by a defined mathematical relationship: specifically, the original coupling matrix... The A singular value It can be derived from eigenvalues Export directly as ,in Indicates the first There are singular values ​​that satisfy... Dimensions and Consistent. (Number) The right singular vectors are the eigenvectors. , No. left singular vectors Then through Calculated (when) (time), among which Indicates the first A left singular vector.

[0045] To achieve data compression while preserving the main coupling characteristics, only the first part is retained. The left and right singular vectors corresponding to the maximal singular values, The singular vectors form the original coupling matrix. A rank Approximate basis. Let the retained rank be... Using this approximate basis, the original coupling matrix can be approximated as follows: The sum of the outer products of the group of low-rank factor vectors. Specifically, for the ... Group( ,in (This represents the index of the factor vector group retained after truncation). The square root of the singular value is combined with the corresponding singular vector to construct the sink factor vector. Source-side factor vector Sink factor vector The The elements are denoted as , indicating the first Group factor vectors at sink node Components at the location; source-side factor vector The The elements are denoted as , indicating the first Group factor vector at source node The portion at that location. Group factor vectors satisfy That is, the original The dimensional coupling matrix is ​​compressed into a form consisting of... An approximate form of a low-dimensional factor vector representation. Rank The selection is based on the cumulative energy percentage threshold of the eigenvalues: the cumulative energy percentage is defined as... When the ratio exceeds a preset threshold (e.g.) or When ) corresponds This refers to the rank retention. Due to the locality and attenuation properties of physical field propagation, the spectral energy of the coupling matrix is ​​usually concentrated on a few dominant modes, hence the rank retention. much smaller In practical applications, the value is usually between several tens and one hundred.

[0046] This Group of low-rank factor vectors and Index by Group Combined storage forms a low-rank factor matrix data structure for use during the online scheduling phase. During online computation, access to the original [database] is no longer required. A dimensional coupling matrix, and only this... By performing an inner product operation between the group factor vector and the resource occupancy sub-vector, an approximate result of the coupling relationship can be efficiently recovered. Simultaneously, the sink node tolerance threshold vector... (its first) element For sink node The tolerance limit), the local index mapping tables of the source node and sink node, and the global resource node table are all stored in the real-time database of the scheduling controller.

[0047] S5. Receive the task to be scheduled, and generate the resource request vector of each target agent using a dimension mask based on the Agent type and task parameters required by the task to be scheduled; wherein, the elements of the resource request vector represent the occupation requirements of the corresponding resource nodes.

[0048] Specifically, after the online scheduling phase begins, the scheduling controller receives a batch of tasks to be scheduled from the MES via industrial Ethernet. Each task is broken down into several sub-tasks by the process analysis module. Each sub-task clearly specifies the required Agent type, the operation object (such as the start and end points of material handling, assembly stations, and inspection stations), the latest completion time window, and key performance requirements (such as the robotic arm torque level and the detection resolution level). Based on the sub-task requirements, the task scheduling engine within the scheduling controller selects the currently available target Agents from the Agent resource pool and generates resource request vectors for each target Agent.

[0049] For each target Agent, the resource request vector generation process is as follows: First, the mask vector and preset default values ​​corresponding to the Agent class to which the Agent belongs are read from the dimension mask configuration library. Based on the identifiers of each dimension in the mask vector, dimensions marked as requiring specification are determined as sensitive dimensions, and dimensions marked as not requiring specification are determined as irrelevant dimensions. Then, according to the task parameters of the subtask, the task-specified values ​​are filled in for each sensitive dimension: the coordinates of the target workstation or channel segment are filled in for the spatial dimension, the sequence numbers of each time slot involved in the task time window are filled in for the time dimension, the function mode dimension is filled in with the function code corresponding to this operation, and the physical field intensity dimension determines the corresponding nominal radiation intensity value by looking up the table according to the torque or power level required by the task. For each irrelevant dimension, the preset default value of the Agent class is uniformly filled in. This constitutes a six-dimensional coordinate set of the resource nodes that the target Agent needs to occupy under this subtask.

[0050] Next, initialize a node with a length equal to the total number of resource nodes. A one-dimensional vector of all zeros is used as the initial resource request vector. Each six-dimensional tuple in the six-dimensional coordinate set is traversed, and the global resource node bidirectional mapping table is queried to obtain the linear index number of the matching resource node. The element position corresponding to this index number in the initial resource request vector is set to a flag value representing "occupancy demand". For most Agents, this flag value is 1, indicating that the resource node is activated and occupied. For Agents with physical field radiation attributes, further processing is required based on their operating mode: at the index position of the relevant node in the physical field intensity dimension, the element is updated to the nominal radiation intensity value of the Agent in that operating mode (for example, a fully torque-operated robotic arm fills in its normalized radiation intensity coefficient on the corresponding source node; this coefficient is greater than 0 and does not exceed 1, where 1 represents full radiation and less than 1 represents degraded use). The final result is... The dimension vector represents the resource request vector of the target agent. The positions of non-zero elements in this vector indicate the indices of occupied resource nodes, and their values ​​represent the intensity or flag of the occupation. Because... Typically large, but each agent occupies very few nodes, the vector uses a sparse storage format (such as compressed sparse line CSR or key-value pair list) to save memory and accelerate subsequent vector addition and inner product operations.

[0051] S6. Combine the target agents to obtain the target agent scheme, and superimpose the resource request vectors of each agent in the target agent scheme to form the total occupancy vector; call the low-rank factor matrix to calculate the global conflict index of the target agent scheme based on the total occupancy vector; where the global conflict index measures the degree of conflict caused by cross-dimensional coupling in the target agent scheme.

[0052] Specifically, after generating the resource request vectors for each target agent, the scheduling controller combines agents to form candidate scheduling allocation schemes (i.e., target agent schemes). The combination method involves selecting a subset of agents from the target agent set that can collectively complete the current task batch. This combination can be generated by a heuristic search algorithm within the scheduling engine (such as beam search) to produce several high-scoring candidate combinations, or it can be generated using the manager-bidder negotiation mechanism in the Contract Network protocol. After selecting a target agent scheme, the resource request vectors of all agents within that scheme are summed element-wise to obtain the total occupancy vector. Since each resource request vector is stored in a sparse format, the addition operation only requires merging the non-zero element lists of each vector and summing the values ​​of repeated indices. The computational complexity is proportional to the total number of non-zero elements.

[0053] Next, the global conflict index of the target agent scheme is calculated. First, using the source node local index mapping table and sink node local index mapping table established in S4, the global conflict index is calculated from the total occupancy vector. Extract the elements corresponding to the source node and the elements corresponding to the sink node respectively, and arrange them in order of local index of the source node to form the source node occupancy subvector. The sink node occupancy subvector is formed by arranging the sink nodes in order of their local indices. .in, The element Indicates the source node The sum of radiation intensity coefficients experienced under the current scheme (if multiple agents occupy the same source node, their radiation intensity is linearly superimposed). The element Indicates sink node Whether an element is active and occupied under the current scheme is determined by a value of 0 or 1 (or an occupancy flag greater than 0). The extraction operation can also be efficiently performed using sparsity: traverse the non-zero elements of the total occupied vector, find the corresponding source node local index or sink node local index in the mapping table based on its global index, and then accumulate the value to the corresponding position of the corresponding subvector.

[0054] Then, the low-rank factor matrix stored in the offline phase is called to calculate the global conflict index. Factor vectors are read group by group from the low-rank factor matrix. For the... Group factor vector, its source side vector The element Indicates the source node Contribution weights in this set of factors, sink vector The element Indicates sink node The received weights in this set of factors. Calculate the source-side vector. Occupied subvectors of source node inner product To obtain the intermediate variables corresponding to this set of factors. The physical meaning of this inner product is to pass the radiation contribution of all source nodes through the first... The group factor is mapped to a compressed equivalent radiation quantity. Because It is also sparse, and the inner product calculation only requires traversing the corresponding non-zero elements. Index, multiplication, accumulation, and corresponding addition This further reduces the computational load.

[0055] For each sink node ( The total radiation estimate is obtained by summing the products of the sink vector elements of each group of factors and the corresponding intermediate variables. The calculation formula is as follows:

[0056]

[0057] in, Indicates sink node The total radiation estimate is given, which approximates the total radiation propagation from all source nodes to the sink node under the current scheme via physical fields. The sum of residual physical field intensities generated at that location. Due to the coupling matrix... Already by Approximately, the above calculation is equivalent to However, the computational complexity is from Reduced to ,in for The number of non-zero elements. When When it is very small, the acceleration effect is extremely significant.

[0058] Then from the sink node tolerance threshold vector Read the tolerance threshold of each sink node Based on the subvector occupied by the sink node. Determine which sink nodes are activated in the current scheme (i.e. The conflict penalty is calculated only for active sink nodes. For each active sink node... Calculate the portion of total radiation estimated to exceed the tolerance threshold. Take the larger of the two values ​​and 0, then square it to get the penalty amount. The physical basis for using the square form is that no conflict penalty is generated when the residual physical field does not exceed the tolerance limit; once it does, the degree of conflict is measured by the square of the excess, which is consistent with the overshoot measurement method for energy or power. Furthermore, the gradient of the square function is zero at the zero point, ensuring good smoothness and numerical stability of the subsequent optimization process near the constraint boundary. The sum of the penalties for all activated sink nodes is the global conflict index of the target agent scheme. The complete expression for calculating the global conflict index is:

[0059]

[0060] in, The global conflict index is a non-negative real number with the dimension of the square of the physical field strength (proportional to the vibration power or energy). A value of 0 indicates that there are no cross-dimensional coupling conflicts. The larger the value, the more severe the global coupling conflict. The total number of sink nodes. The total number of source nodes. The rank retained for low-rank decomposition. The sink node occupies the first subvector. The occupancy value of each sink node (a value of 0 or 1 indicates whether the tolerance constraint is activated). The source node occupies the first subvector. The occupancy value of each source node (representing the normalized radiation intensity coefficient). For the first The sink lateral vector in the group factor vector at the sink node The element at that location, For the first In the group factor vector, the source side vector is at the source node The element at that location, For sink node The physical field tolerance threshold, and They have the same normalized dimensions. To maximize the function, a non-zero penalty is only generated when the total radiation estimate exceeds the tolerance threshold.

[0061] In this formula, the summation term Calculate intermediate variables Summation term Convergence The contribution of the group factors is used to estimate the total radiation. Subsequently and The difference is truncated by the max function, the negative values ​​are then squared, and the sink node is obtained. The amount of punishment; multiplied by Ensure that only active sink nodes contribute the penalty; ultimately for all The nested structure sums the values ​​of each sink node. This nested structure reflects the complete process from "radiation source - source-side projection - sink-side reconstruction - tolerance alignment - conflict quantification" layer by layer.

[0062] The aforementioned conflict index calculation process essentially transforms the resource request tensor and coupling tensor contraction operations, which originally needed to be performed in a six-dimensional tensor space, into a series of sparse matrix-vector multiplications and inner product operations. This process compresses the coupling information to a smaller value using the low-rank approximation of the coupling matrix. In the group factor vector, global quantitative evaluation of cross-dimensional conflicts is completed with extremely low dimensionality. This approach is the key to the real-time scheduling capability of this invention, as it utilizes the locality and low rank of physical field coupling to simplify the exponentially complex conflict detection problem into algebraic operations with linear time complexity.

[0063] S7. When the global conflict index exceeds the index threshold, adjust the adjustable running parameters corresponding to the elements of the resource request vector of at least one Agent in the target Agent scheme until the recalculated global conflict index does not exceed the index threshold, and obtain the final total occupancy vector; generate scheduling instructions based on the target Agent scheme and the corresponding final total occupancy vector and issue them to the corresponding Agent.

[0064] Specifically, the scheduling controller will use the global conflict index calculated in step S6. With the preset exponential threshold Compare. Exponential threshold. It is a very small positive real number close to zero, and its value is set according to the production line's tolerance for implicit conflicts (e.g., taking 0 for complete conflict-free operation, or taking a safe upper limit that allows for weak coupling based on engineering experience). When When this occurs, it indicates that the current target agent solution does not have unacceptable cross-dimensional coupling conflicts, and the solution is directly considered a feasible solution. When this occurs, it indicates that there is a hidden conflict in the current scheme, and a continuous parameter adjustment mechanism needs to be activated to resolve the conflict.

[0065] In this embodiment, parameter tuning employs a gradient-based iterative optimization method. Subscripts are used below. This indicates the sequence number of the Agent in the target Agent scheme. , This represents the total number of Agents in the scheme; agents with the same serial number... All the corresponding parameters point to the same Agent entity; subscript This represents the position index of an element in the resource request vector, corresponding to a specific resource node. First, a pre-defined mapping relationship is obtained between the adjustable operating parameters of each Agent in the target Agent scheme and the element values ​​of the resource request vector. For each type of Agent, this mapping relationship describes how the Agent's continuously adjustable operating parameters (such as the end-effector speed factor, torque limit factor, and trajectory smoothness parameter of the robotic arm; the travel speed factor and acceleration limit of the AGV; and the resolution level of the vision inspection device) affect the value of a specific element (mainly the occupancy value of the physical field strength dimension node) in its resource request vector. This mapping relationship is pre-stored in the form of an analytical function or interpolation table; for example, for the robotic arm (the... (the Agent), which is in the first... Occupancy value on each source node With torque limiting factor The relationship can be represented as .in, Indicates the first The resource request vector of the agent corresponds to the first... The element value of the source node, i.e., the element value of the source node. The normalized radiation intensity of each agent at the source node; As the torque limiting factor, Corresponding to full torque operation, Corresponding to reduced torque operation; The nominal radiation intensity of the source node under full torque is a normalized constant calibrated experimentally. The exponent, calibrated by dynamics, reflects the nonlinear relationship between radiation intensity and torque, obtained by fitting measured data. For non-radiative source nodes, their occupancy value is not affected by adjustable operating parameters, and the mapping relationship is a constant function.

[0066] Construct a constrained optimization problem: the optimization variables are the adjustable operating parameters of each Agent, the objective function is the sum of squared Euclidean distances between the resource request vectors of each Agent before and after adjustment (representing the degree of performance degradation, the smaller the better), and the constraint is that the global conflict index of the adjusted solution does not exceed [a certain value]. Then, the gradients of the objective function with respect to each adjustable operating parameter (first gradient) and the gradients of the constraint function (i.e., the global conflict index) with respect to each adjustable operating parameter (second gradient) are calculated. The two gradients are then weighted and summed using preset penalty weights or Lagrange multipliers to obtain the weighted gradient, the negative direction of which is the weighted search direction. The adjustable operating parameters of each Agent are updated along this search direction, and the corresponding element values ​​in their respective resource request vectors are updated synchronously according to the mapping relationship. After each round of updates, the total occupancy vector is re-stacked and the global conflict index is recalculated to check if the constraints are satisfied. This update process is iterated and repeated until... Or the number of iterations reaches a preset limit. The updated resource request vectors of each Agent after satisfying the constraints are then superimposed to obtain the final total occupancy vector.

[0067] The resource request vector of each Agent in the target Agent scheme is parsed into specific execution instructions: spatial nodes correspond to the target location coordinates, time nodes correspond to the time slot for action initiation, functional mode nodes correspond to the operation type and parameters, and the adjustable operating parameters corresponding to the physical field intensity nodes are written into the scheduling instruction message as operational constraint suggestions (such as the upper limit of the robot arm's torque). The scheduling controller sends the scheduling instructions to the local controller of each Agent via industrial Ethernet. After receiving the instructions, each Agent executes the corresponding task within the specified time window.

[0068] This embodiment breaks through the limitation of traditional scheduling, which only performs conflict detection in discrete dimensions, by explicitly modeling the physical field coupling relationship and deeply integrating it with scheduling optimization. It realizes dynamic scheduling control that detects and resolves cross-dimensional implicit conflicts in advance during the resource allocation decision-making stage, thereby avoiding deadlock and quality accidents caused by physical field coupling from the root. This allows heterogeneous agents to be safely mixed and operated in parallel within a more compact spatiotemporal window, improving the overall utilization rate and scheduling stability of the production line.

[0069] The aforementioned agent-based dynamic scheduling and control method for industrial robots constructs a global resource space encompassing spatial, temporal, functional mode, and physical field strength dimensions, instantiating production line resource units as resource nodes within this space. Based on this, dimensional masks are generated for various heterogeneous agents to identify the dimensions that impose substantial constraints when generating resource requests. This quantifies and distinguishes the essential differences in dimensional perspectives between different agents such as AGVs, robotic arms, and vision inspection devices regarding the same physical resource. Subsequently, a physical field propagation model is used to determine the cross-dimensional coupling relationships between resource nodes caused by the propagation of physical fields such as vibration and electromagnetic fields. Nodes acting as radiation sources and their affected sink nodes are organized into hyperedges to construct a resource coupling hypergraph. This allows implicit conflicts that cannot be detected by single-dimensional detection (such as robotic arm vibration leading to loss of detection equipment accuracy, AGV passage intrusion into the robotic arm's working envelope, etc.) to be explicitly expressed as structured relationships within the hypergraph. Furthermore, physical field transmission coefficients are extracted from the hypergraph to construct a coupling matrix, which is then decomposed into a low-rank factor matrix, providing pre-compressed coupling information for the online scheduling phase. Upon receiving tasks to be scheduled online, the task requirements of each target agent are uniformly transformed into a one-dimensional resource request vector based on the same resource node space using a dimensionality mask. Then, by combining target agents and superimposing resource request vectors into a total occupancy vector, a low-rank factor matrix is ​​used to quickly calculate a global conflict index. This allows for real-time quantification of the degree of cross-dimensional coupling conflict in the resource allocation decision-making stage. If the global conflict index exceeds a threshold, the allocation scheme is not directly rejected. Instead, the adjustable operating parameters of the relevant agents are continuously adjusted to proactively resolve latent conflicts until the conflict index meets the constraints, at which point a scheduling instruction is generated and executed. This technical solution moves the detection of latent conflicts from the traditional post-deadlock formation stage to the task allocation decision-making stage. By continuously adjusting parameters, it empowers the scheduling system with the ability to proactively resolve cross-dimensional conflicts. This allows for high-density, secure hybridization of heterogeneous agents within a more compact spatiotemporal window without altering task semantics. This fundamentally avoids deadlocks, quality incidents, and unplanned downtime caused by cross-dimensional coupling in the physical field, significantly improving the stability of production line scheduling and equipment utilization.

[0070] Example 2:

[0071] Based on the resource usage characteristics of each type of Agent, a dimension mask is generated to identify the dimensions that the corresponding Agent needs to specify when generating resource requests, including the following steps:

[0072] S21. Based on the pre-built ontology knowledge base, obtain the functional attributes and resource requirement descriptions of various Agents.

[0073] S22. For each type of Agent, extract the Agent's requirements for spatial location, time window, functional mode, and physical field intensity when performing tasks, based on the functional attributes and resource requirements description.

[0074] S23. Based on the requirements information, mark the dimension in the global resource space that the Agent needs to specify the value as the first identifier, and mark the other dimensions as the second identifier, and generate a mask vector.

[0075] S24. For the dimension marked as the second identifier in the mask vector, set the default value of the corresponding Agent in the corresponding dimension; associate and store the mask vector and the default value to obtain the dimension mask.

[0076] Specifically, the scheduling controller accesses a pre-built ontology knowledge base during the offline preparation phase. This ontology knowledge base can adopt existing RMFS (Robotic Mobile Fulfillment System) or ACNBP (Agent-based Collaborative Network Business Process) architectures, and use the Web Ontology Language (OWL) to describe equipment classes, resource classes, operation classes, and their attribute relationships in the production line. The ontology knowledge base has already defined functional attributes (such as the list of executable operations, workspace range, and payload capacity) and resource requirement descriptions (such as resource types and operating condition constraints) for Agent classes such as "AGV", "SixAxisArm", and "VisionInspector".

[0077] The scheduling controller retrieves the functional attributes and resource requirement descriptions of Agents class by class through the ontology query interface. For each Agent class, based on its functional attributes and resource requirement descriptions, it extracts the Agent's requirements for four aspects during task execution: spatial location, time window, functional mode, and physical field intensity. Specifically, if the ontology describes the Agent's workspace coordinates and channel occupancy characteristics, the Agent is considered to have substantial constraints in the spatial dimension; if the ontology describes the Agent's operation sequence and deadline requirements, it is considered to have substantial constraints in the time dimension; if the ontology defines a list of executable functional operation types for the Agent, it is considered to have substantial constraints in the functional mode dimension; if the ontology records the Agent's vibration characteristics or sensitivity parameters to environmental vibrations during operation, it is considered to have substantial constraints in the physical field intensity dimension. If the ontology does not describe a certain requirement, or if that requirement is entirely determined by other factors (such as external process constraints) and not controlled by the Agent itself, the Agent is considered to have no substantial constraints in that dimension.

[0078] Based on the extracted requirements information, for the six dimensions in the global resource space, the dimension for which the Agent needs to specify values ​​is marked as the first identifier (i.e., sensitive dimension identifier), and the remaining dimensions are marked as the second identifier (i.e., irrelevant dimension identifier), generating a six-bit mask vector. For the dimensions marked as the second identifier in the mask vector, corresponding preset default values ​​are set according to the typical values ​​or default states of that type of Agent recorded in the ontology for that dimension. For example, the preset default value for the AGV in the functional mode dimension comes from the "transportation" operation code defined in its ontology, and the preset default value in the physical field strength dimension comes from the nominal radiation value corresponding to its "full load operation" state. Finally, this mask vector is associated with the preset default values ​​of each irrelevant dimension and stored as a structured dimension mask record.

[0079] Compared to the preset method in Embodiment 1, this ontology-based automatic generation method not only ensures the consistency and accuracy of the mask, but also automatically generates the dimension mask simply by adding the corresponding description to the ontology library when a new Agent is added to the production line, thus improving the system's scalability and maintenance efficiency.

[0080] Example 3:

[0081] refer to Figure 2 Based on the physical field propagation model, by determining the cross-dimensional coupling relationships caused by physical field propagation among resource nodes, the source nodes that act as the radiation source of the physical field and the sink nodes affected by the physical field are organized into hyperedges to obtain a resource coupling hypergraph, including the following steps:

[0082] S31. Based on the operational attributes of various Agents, resource nodes are divided into source nodes and sink nodes.

[0083] S32. For each source node, based on the physical field propagation model, the residual physical field intensity of the source node to each sink node is calculated using the physical field radiation intensity of the source node and the spatial distance between the source node and each sink node.

[0084] S33. Traverse all source nodes. When the residual physical field strength exceeds the physical field tolerance threshold of the corresponding sink node, organize the source node and the sink node that exceeds the physical field tolerance threshold into a hyperedge to obtain the resource coupling hypergraph.

[0085] Specifically, based on the operational attributes of various agents, all resource nodes are first divided into two categories: source nodes and sink nodes. The division rule is as follows: analyze the agent type of each resource node and its operational characteristics on that node. If a node belongs to the motion actuator type agent (such as the workspace of a robotic arm joint, the walking area of ​​an AGV), and the agent's operation on that node generates outward-propagating physical fields (vibration, electromagnetic noise, etc.), then that node is marked as a source node. If a node belongs to the perception or precision operation type agent with a clear tolerance threshold to environmental physical fields (such as a vision inspection station, a precision measurement station, a vibration-sensitive processing station), and the normal operation of the agent on that node is affected by external physical fields, then that node is marked as a sink node. If a node simultaneously satisfies two characteristics (for example, some equipment workstations both generate vibration and are sensitive to vibration), then it is determined according to its main role in the scheduling scenario. The division results are used to establish source node index lists and sink node index lists, respectively.

[0086] Secondly, for each source node, the residual physical field intensity at each sink node is calculated based on the physical field propagation model (taking a vibration field as an example). This embodiment uses an exponential decay model with directional correction, and the calculation formula is as follows:

[0087]

[0088] in, Indicates the source node At the remittance node The residual physical field intensity generated at that location, Indicates the source node The physical field radiation intensity is expressed as a normalized dimensionless value, with the magnitude determined according to the Agent's operating mode. The material vibration attenuation coefficient, in units of The transfer rate of each typical path was obtained by least-squares fitting through offline tapping experiments. For the source node With sink node The spatial Euclidean distance between them. This is a directional correction factor used to describe the anisotropic characteristics of vibration propagating further along the dominant direction of the machine tool (such as the long axis of the guideway). In this embodiment, The directional expression is ,in For the source node The dominant vibration direction of the equipment and the source node Point to sink node The angle between the directions of the lines connecting them. The anisotropy degree is obtained by fitting measurement data using multi-directional transfer functions. This applies when the source and sink nodes are located in different structural elements and there is a connecting seam in between. The expression is additionally multiplied by a connector attenuation factor related to the connector type. (welding connection) near Bolted connection Slightly smaller ).

[0089] For each source node, the residual physical field intensity is calculated by traversing all sink nodes. The calculated Physical tolerance threshold of the corresponding sink node Compare. Tolerance threshold. The tolerance limit is determined based on the agent's operational requirements at the sink node: for example, the high-precision detection mode of a visual inspection device corresponds to an extremely low tolerance limit, while the low-precision mode corresponds to a higher tolerance limit. Greater than or equal to With a safety factor ( When determining the source node by multiplying the product of (used to capture potential interference close to the threshold)... With sink node There is a cross-dimensional coupling relationship between the sink node. Join with source node The coupling between the source node and the sink node set is the central factor. After traversal, if the set is not empty, the source node will be... Organize all sink nodes in this set into a hyperedge and record the hyperedge data. Traverse all source nodes and collect all hyperedges to obtain the complete resource coupling hypergraph.

[0090] This embodiment introduces a directional correction factor and a more refined node role division, making the resource coupling hypergraph more accurate in depicting the propagation of the real physical field. Compared with the isotropic simplified model in Embodiment 1, it can effectively reduce the omission and false judgment of coupling relationships, thereby improving the accuracy of subsequent conflict detection.

[0091] Example 4:

[0092] Based on the transmission relationships from source nodes to sink nodes recorded by each hyperedge in the resource coupling hypergraph, a coupling matrix is ​​generated, with matrix elements representing the physical field transmission coefficients from source nodes to sink nodes. The coupling matrix is ​​then decomposed into multiple sets of low-rank factor vectors, which are combined into a low-rank factor matrix. This process includes the following steps:

[0093] S41. Based on the residual physical field intensity of each sink node recorded in each hyperedge in the resource coupling hypergraph, and combined with the physical field radiation intensity of the source node corresponding to the hyperedge, calculate the physical field transfer coefficient from the source node to each sink node under unit radiation intensity.

[0094] S42. Construct a source-sink transfer matrix based on the physical field transfer coefficients; wherein the source-sink transfer matrix has sink nodes as rows and source nodes as columns, and the elements of the source-sink transfer matrix are the physical field transfer coefficients from the corresponding source node to the sink node.

[0095] S43. Perform truncated singular value decomposition on the source-sink transitive matrix to obtain the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix.

[0096] S44. Determine the rank K to be retained based on the cumulative energy ratio of the singular values ​​in the singular value diagonal matrix. Select the first K largest singular values ​​from the singular value diagonal matrix and extract the first K column vectors from the left singular vector matrix and the right singular vector matrix respectively to obtain the sequence of retained singular values, the group of retained left singular vectors, and the group of retained right singular vectors.

[0097] S45. Multiply the square root of each singular value in the sequence of retained singular values ​​by the corresponding left and right singular vectors to obtain K sets of low-rank factor vectors; combine the K sets of low-rank factor vectors to obtain a low-rank factor matrix.

[0098] Specifically, the data required to calculate the physics transfer coefficients is first extracted from the record information of each hyperedge in the constructed resource coupling hypergraph. For the source node corresponding to the hyperedge... and each sink node in the superedge The calculated residual physical field intensity has been stored in the hyperedge data structure. Source node Its own nominal physical field radiation intensity Read from the resource node attribute table. Calculate the physical field transfer coefficient from the source node to the sink node per unit radiation intensity. The physical meaning of this coefficient is: when the source node... When generating one unit of normalized radiation intensity, at the sink node The residual physical field intensity caused at the location; dimensionally, due to and Using the same normalization benchmark, It is a dimensionless pure proportionality coefficient, with a value between 0 and 1.

[0099] Secondly, a source-sink transfer matrix is ​​constructed based on the physical field transfer coefficients of all source-sink node pairs. This matrix is ​​arranged with sink nodes as rows (row indices correspond to the local indices of the sink nodes). ), with the source node as the column (the column index corresponds to a local index of the source node). ). The first of the matrix Line number The column elements are assigned the corresponding transit coefficients. For source-sink node pairs that do not appear in any hyperedge (i.e., there is no coupling relationship), the corresponding matrix elements are set to 0. Due to the locality of physical field propagation and the sparsity of device layout, this matrix is ​​a highly sparse asymmetric non-negative real matrix.

[0100] Then, truncated singular value decomposition is performed on the source-sink transitive matrix. The mathematical expression for truncated singular value decomposition is: ,in, It is a coupling matrix. for OK A column orthogonal matrix (left singular vector matrix) in which each column is a left singular vector; for OK A diagonal matrix of columns (singular value diagonal matrix), diagonal elements Singular values ​​to be preserved; for OK A column orthogonal matrix (right singular vector matrix) in which each column is a right singular vector. This refers to the rank retained after truncation. Rank The method for determining this is as follows: calculate the cumulative energy percentage of the sum of squares of all singular values, where the cumulative energy percentage is defined as... Starting with the largest singular value, the values ​​are accumulated sequentially. When the cumulative energy percentage first exceeds a preset threshold (e.g., 0.99 or higher), the number of singular values ​​already included in the accumulation is the total energy percentage. ;in, This represents the i-th element after sorting by singular value in descending order. There are several singular values. The principle for selecting this threshold is: to compress as much as possible while ensuring that the approximate accuracy is sufficient to support correct collision detection. This reduces the computational burden on online systems. In practical applications, since physical field coupling typically occurs only between a finite number of dimensional combinations, The value is much smaller than the minimum dimension of the matrix, usually within tens.

[0101] Next, the truncated singular value decomposition results are further processed and converted into a low-rank factor vector form that facilitates online inner product calculation. Specifically, the retained... The singular values ​​are ordered by number. Perform the following transformation: For the th A singular value Take its square root , respectively with the corresponding left singular vector (Right now The (column) and right singular vector (Right now The Multiplying the columns yields the sink factor vector. Source-side factor vector After this treatment, , And satisfy . and Each matrix is ​​not orthogonal, but the sum of their products constitutes the optimal rank of the original matrix. Approximate. For each set of factor vectors, record the values ​​of its sink side vector and source side vector, as well as the corresponding local index mapping relationship.

[0102] Finally, all Group low-rank factor vector, sink node tolerance threshold vector The local index mapping tables of the source node and sink node are stored together to form a low-rank factor matrix data structure for use in the online phase. This embodiment uses truncated singular value decomposition to compress coupled matrix information using a theoretically optimal least-squares approximation method, achieving the same rank... The following method can achieve a more accurate approximation than the feature decomposition in Example 1, thereby further ensuring the computational accuracy and confidence of online conflict detection.

[0103] Example 5:

[0104] Based on the Agent type and task parameters required by the task to be scheduled, resource request vectors for each target Agent are generated using dimension masks, including the following steps:

[0105] S51. For each target Agent, read the mask vector and preset default value corresponding to the Agent class to which the target Agent belongs from the dimension mask.

[0106] S52. Determine the dimension marked as the first identifier in the mask vector as the sensitive dimension, and the dimension marked as the second identifier in the mask vector as the irrelevant dimension; according to the task parameters, take the task-specified value for the sensitive dimension and the preset default value for the irrelevant dimension to form the multi-dimensional coordinates of the resource nodes that the target Agent needs to occupy.

[0107] S53. Query the resource node index that matches the multidimensional coordinates of the resource node in the global resource space, obtain the set of occupied node indexes, and initialize a one-dimensional zero vector with a length equal to the total number of resource nodes as the initial resource request vector.

[0108] S54. Based on the set of occupied node indices, configure the element at the corresponding index position in the initial resource request vector to represent the value of the occupied requirement, and obtain the intermediate resource request vector.

[0109] S55. For an Agent with physical field radiation attributes, determine the physical field radiation source node index according to the target Agent's operating mode; update the position of the physical field radiation source node index in the intermediate resource request vector to the nominal radiation intensity value to obtain the resource request vector.

[0110] Specifically, for each candidate instance selected as a target Agent, the scheduling controller first reads the mask vector and preset default values ​​corresponding to the Agent class to which the target Agent belongs from the dimension mask configuration library. The mask vector is a six-bit binary identifier sequence, with each bit corresponding to a dimension in the global resource space. The first identifier indicates a sensitive dimension (a value must be specified), and the second identifier indicates an irrelevant dimension (a value does not need to be specified). The preset default value record contains the default values ​​pre-set for each dimension with the second identifier.

[0111] Then, the dimension identified by the first identifier in the mask vector is determined as the sensitive dimension, and the dimension identified by the second identifier is determined as the irrelevant dimension. Based on the task parameters of the task to be scheduled, specific values ​​specified by the task are filled in for each sensitive dimension: for the spatial dimension, the spatial coordinates of the target workstation or the coordinate sequence of the path channel are filled in; for the temporal dimension, the sequence numbers of all time slots covered by the time window allowed for task execution are filled in; for the functional mode dimension, the operation type code required by the task is filled in; for the physical field strength dimension, the normalized nominal value of radiation intensity corresponding to the torque, speed, or power level required by the task is filled in. For each irrelevant dimension, the corresponding value from the preset default values ​​is uniformly filled in. The values ​​of these six dimensions constitute the multidimensional coordinates of one or more resource nodes that the target Agent needs to occupy. When the task involves consecutive time slots or consecutive paths, multiple corresponding multidimensional coordinate tuples are generated.

[0112] Next, a resource node index query is performed. All resource node multidimensional coordinate tuples generated in the previous step are traversed. For each multidimensional coordinate tuple, the values ​​of its six dimensions are used as the query key to retrieve the matching linear index number of the resource node in the bidirectional mapping table of the global resource space. All retrieved valid index numbers constitute the set of occupying node indexes. Simultaneously, a table with a length equal to the total number of global resource nodes is initialized. A one-dimensional zero vector is used as the initial resource request vector.

[0113] Based on the set of occupying node indices, the element value at the corresponding index position in the initial resource request vector is modified from zero to a value representing the occupation requirement. For general sensitive dimension occupation, this value is set to 1, indicating that the node is activated and occupied by the current Agent. This yields the intermediate resource request vector.

[0114] For agents with physical field radiation attributes (such as robotic arms and AGVs), further processing of the source nodes in the physical field intensity dimension is required. Based on the target agent's operating mode in this task (such as torque level and speed mode), the corresponding physical field radiation source node index and nominal radiation intensity value are retrieved from the equipment physical characteristic mapping table. The element positions in the intermediate resource request vector corresponding to these physical field radiation source node indices are updated from the original occupancy flag value of 1 to the nominal radiation intensity value. The final vector obtained is the complete resource request vector for the target agent.

[0115] The elements in this vector have specific technical meanings: a value of 1 indicates that the agent occupies the corresponding resource node but does not generate physical field radiation, or that the agent's tolerance constraint is activated; a value of a positive real number between 0 and 1 represents the normalized radiation intensity of the agent on the corresponding source node; a value of 0 indicates that the agent does not occupy the corresponding resource node. The positions of all non-zero elements represent the indices of all resource nodes involved in the agent's current task. This vector is stored using a sparse data structure, retaining only the index-value pairs of non-zero elements for efficient online computation.

[0116] Example 6:

[0117] refer to Figure 3 The process involves calling the low-rank factor matrix and calculating the global conflict index of the target agent scheme based on the total occupancy vector, including the following steps:

[0118] S61. Based on the source nodes and sink nodes partitioned in the resource coupling hypergraph, according to the elements corresponding to each source node in the total occupancy vector, form a source node occupancy subvector according to the arrangement order of the source nodes, and according to the elements corresponding to each sink node in the total occupancy vector, form a sink node occupancy subvector according to the arrangement order of the sink nodes.

[0119] S62. Extract each set of factor vectors from the low-rank factor matrix; where each set of factor vectors consists of the source side vector corresponding to the source node and the sink side vector corresponding to the sink node.

[0120] S63. For each set of factor vectors, calculate the inner product of the source side vector and the source node occupied sub-vector to obtain the intermediate variable corresponding to the corresponding factor vector.

[0121] S64. For each sink node, calculate the total radiation estimate of the corresponding sink node by using the sum of the products of the elements of the sink side vectors in each group of factor vectors at the corresponding sink node and the corresponding intermediate variables.

[0122] S65. Based on the resource coupling hypergraph, obtain the physical field tolerance threshold of each sink node; determine the target sink nodes to be activated based on the sink node occupancy subvector; for each target sink node, calculate the penalty amount for the total radiation estimate exceeding the physical field tolerance threshold; sum the penalties of all target sink nodes to obtain the global conflict index; wherein, the formula for calculating the global conflict index is:

[0123]

[0124] in, This indicates the global conflict index. The total number of sink nodes. The total number of source nodes. The rank retained by the low-rank decomposition of the coupling matrix. Indicates the sink node occupies the first subvector. The occupancy value of each sink node. Indicates the source node occupies the first subvector. The occupancy value of each source node. The th in the low-rank factor matrix Group factor vectors at sink node The element at that location, The th in the low-rank factor matrix Group factor vector at source node The element at that location, For sink node The physical field tolerance threshold.

[0125] Specifically, firstly, based on the source and sink node information already partitioned in the resource coupling hypergraph, the total occupancy vector is used... Extract two sub-vectors from the given vector. Total occupied vector For one A dimensional vector, whose elements are arranged in linear index order of global resource nodes. Based on the source node local index mapping table established in S4, determine... The elements corresponding to each source node (global index) are arranged in the order of their local indexes within those source nodes. Take them out one by one and arrange them to form the source node occupied subvector. Similarly, based on the sink node local index mapping table, determine... The elements corresponding to each sink node are arranged in order of their local indexes within the sink node. Take them out one by one and arrange them to form the sink node occupied subvector. This extraction operation is efficiently performed based on the correspondence between two sparse vectors and an index mapping table, without needing to traverse the entire table. Each element.

[0126] Secondly, factor vectors are extracted group by group from the low-rank factor matrix. For the first... Group( ), its source side vector It is a length of A real vector whose elements ( ) represents the source node In the The weights in the group factors; their sink vectors It is a length of A real vector whose elements ( ) represents the sink node In the The weights in the group factors. The group factor vectors were obtained offline through truncated singular value decomposition and stored in the real-time database.

[0127] For each set of factor vectors, calculate the source-side vector. Occupied subvectors of source node The inner product of these factors yields the intermediate variables corresponding to this set of factor vectors. The formula for calculating the inner product is: The physical meaning of this intermediate variable is: to pass the radiation contribution of all source nodes under the current scheme through the first... The source-side weights of the group factor vectors are projected and compressed to obtain the first group. The overall excitation intensity under each equivalent radiation mode. intermediate variables constitute a A dimensional intermediate variable vector realizes the transformation from... Source node occupancy information Dimensionality reduction mapping of dimensionality compression representation, Typically much smaller .

[0128] Then, for each sink node Using the sink lateral vectors in each group of factor vectors at the sink node element at With corresponding intermediate variables The sum of the products of these products is used to calculate the total radiation estimate for the sink node. The calculation formula is:

[0129]

[0130] The total radiation estimate Approximately equal to the original coupling matrix Calculated sink node The sum of physical field strengths at all source nodes The approximate accuracy depends on the truncated rank. ; The larger the value, the smaller the approximation error.

[0131] Next, from the sink node tolerance threshold vector Obtain the physical tolerance threshold for each sink node. Based on the subvector occupied by the sink node. The value of each element determines which sink nodes are activated in the current scheme: when When, it indicates a sink node. The node is occupied (i.e., an agent is performing a sensitive task on this node, and its tolerance constraints are in effect). Conflict penalties are calculated only for active sink nodes. For each active sink node... The portion of the total radiation estimate that exceeds the tolerance threshold is calculated, and the larger of this portion and 0 is taken, then squared to obtain the penalty amount for that sink node. This penalty amount is the penalty amount for the sink node. Local contribution to the global conflict index.

[0132] Finally, the penalties for all activated sink nodes are summed to obtain the global conflict index of the target agent scheme. The complete calculation formula is as follows:

[0133]

[0134] in, The global conflict index is a non-negative real number with the dimension of the square of the physical field strength (proportional to the vibration power or energy). A value of 0 indicates that there are no cross-dimensional coupling conflicts. The larger the value, the more severe the global coupling conflict. The total number of sink nodes. The total number of source nodes. The rank retained for low-rank decomposition. The sink node occupies the first subvector. The occupancy value of each sink node (a value of 0 or 1 indicates whether the tolerance constraint is activated). The source node occupies the first subvector. The occupancy value of each source node (representing the normalized radiation intensity coefficient). For the first The sink lateral vector in the group factor vector at the sink node The element at that location, For the first In the group factor vector, the source side vector is at the source node The element at that location, For sink node The physical field tolerance threshold, and They have the same normalized dimensions. To maximize the function, a non-zero penalty is only generated when the total radiation estimate exceeds the tolerance threshold.

[0135] In this formula, the summation term Calculate intermediate variables Summation term Convergence The contribution of the group factors is used to estimate the total radiation. Subsequently and The difference is truncated by the max function, the negative values ​​are then squared, and the sink node is obtained. The amount of punishment; multiplied by Ensure that only active sink nodes contribute the penalty; ultimately for all The nested structure sums the values ​​of each sink node. This nested structure reflects the complete process from "radiation source - source-side projection - sink-side reconstruction - tolerance alignment - conflict quantification" layer by layer.

[0136] Example 7:

[0137] Adjust the adjustable runtime parameters of at least one Agent in the target Agent scheme corresponding to the elements of the resource request vector until the recalculated global conflict index does not exceed the index threshold, to obtain the final total occupancy vector, including the following steps:

[0138] S71. Based on the preset mapping relationship between the adjustable operating parameters of each Agent in the target Agent scheme and the element values ​​of the resource request vector, and using the adjustable operating parameters of each Agent as optimization variables, construct an optimization problem with the objective function of minimizing the sum of squared Euclidean distances between the resource request vectors before and after adjustment, and with the constraint that the global conflict index does not exceed the exponential threshold.

[0139] S72. Based on the optimization problem, calculate the gradient of the objective function with respect to the adjustable operating parameters as the first gradient; calculate the gradient of the global conflict index with respect to the adjustable operating parameters as the second gradient; and sum the first gradient and the second gradient with weights to obtain the weighted gradient.

[0140] S73. Update the adjustable operating parameters along the search direction corresponding to the weighted gradient to obtain the updated adjustable operating parameters; update the resource request vector according to the updated adjustable operating parameters and the preset mapping relationship to obtain the updated resource request vector; recalculate the global conflict index according to the updated resource request vector.

[0141] S74. Repeat S73 until the global conflict index does not exceed the index threshold, and then sum the updated resource request vectors of each Agent in the target Agent scheme to form the final total occupancy vector.

[0142] Specifically, the following uses subscripts. This indicates the sequence number of the Agent in the target Agent scheme. , This represents the total number of Agents in the scheme; agents with the same serial number... All the corresponding parameters point to the same Agent entity; subscript This represents the position index of an element in the resource request vector, corresponding to a specific resource node. First, a pre-defined mapping relationship is obtained between the adjustable operating parameters of each Agent in the target Agent scheme and the element values ​​of the resource request vector. For each Agent type, this mapping relationship describes how a set of adjustable operating parameters of the Agent (e.g., the end-effector speed factor, torque limit factor, and trajectory smoothness parameter of the robotic arm; the travel speed factor and acceleration limit of the AGV; the resolution level of the vision inspection device) affects the value of a specific element in its resource request vector (mainly the radiation intensity value of the source node in the physical field strength dimension). This mapping relationship is stored in the Agent configuration file in the form of a set of differentiable functions or piecewise linear interpolation functions. For example, for the first... The first robotic arm agent Occupancy value of each source node With torque limiting factor The relationship can be expressed as ,in This represents the full nominal radiation intensity of the source node. This is an empirically calibrated power exponent. For non-radiative source nodes, its occupancy value is independent of adjustable operating parameters, and the mapping derivative is zero.

[0143] Secondly, using the adjustable operating parameters of all agents in the target agent scheme as optimization variables, a constrained continuous optimization problem is constructed. Objective function Defined as the sum of squared Euclidean distances between the resource request vectors of each Agent before and after adjustment, its expression is:

[0144]

[0145] in, This represents the vector of adjustable runtime parameters for all agents in the target agent scheme. The concatenated vector, i.e. ,each A high-dimensional vector is formed by concatenating the first and last entries according to the Agent's serial number. For the first A vector of adjustable runtime parameters for each Agent; This represents the set of Agent indices in the target Agent scheme. Indicates the first The original resource request vector of each agent before adjustment (i.e., under optimal performance conditions). Indicates in the parameter The resource request vector is updated based on a preset mapping relationship. The physical meaning of this objective function is to minimize the total performance degradation, i.e., to make the adjusted operating parameters as close as possible to the original optimal parameters. The constraint is the global conflict index of the adjusted scheme. Not exceeding the exponential threshold The optimization problem takes the form of a nonlinear least squares problem with inequality constraints.

[0146] Then, the primal-dual method is used to solve the above constrained optimization problem. The Lagrangian function is constructed as follows: ,in These are Lagrange multipliers, corresponding to inequality constraints. In each iteration, regarding the parameters... Minimize the Lagrange function with respect to the multipliers Maximize the Lagrange function. This yields... The update direction is the Lagrange function pair Negative gradient: This is a weighted combination of the first and second gradients, where the weights are Lagrange multipliers. Therefore, it is necessary to calculate the gradient of the objective function with respect to each adjustable operating parameter (first gradient) and the gradient of the global conflict index with respect to each adjustable operating parameter (second gradient).

[0147] The gradient of the objective function is obtained using the standard derivative rule of the Euclidean distance with respect to a vector: for the The Agent's first Adjustable operating parameters , The partial derivative with respect to it is The summation covers the nth... All elements in the Agent resource request vector For vectors The One element, For vectors The Each element.

[0148] The global conflict index gradient is derived from the conflict index calculation formula using the chain rule. For ease of expression, the first... Resource request vectors of each Agent Its adjustable running parameter vector The Jacobian matrix is ,in For parameter vectors The dimension of the Jacobian matrix. Line number Column elements are Therefore, row vectors That is, the first Jacobian matrix. The row represents the source node. The partial derivatives of the occupancy value with respect to each parameter. Global conflict index. For the A vector of adjustable runtime parameters for each Agent The gradient calculation formula is:

[0149]

[0150] in, Indicates the global conflict index For the A vector of adjustable runtime parameters for each Agent The gradient vector, whose th... Each component is . For sink node The total radiation estimate is calculated using the formula defined in the foregoing embodiments. Indicates the first The set of local indexes of the source node in the resource request vector of an Agent that are affected by adjustable runtime parameters. That is, the first Jacobian matrix. The gradient is calculated using the derivative function of a predefined mapping relationship. The physical meaning of each symbol in this gradient formula is consistent with that in Example 6. During the gradient calculation process, due to the low-rank factor vector... and Pre-stored offline. It can be quickly obtained by differentiating the analytical function, and the entire gradient calculation can be completed in milliseconds.

[0151] The first gradient ( right The gradient of the first gradient and the gradient of the second gradient right (gradient) according to the current Lagrange multiplier We obtain the weighted gradient by weighted summation. Lagrange multipliers The update is dynamically updated based on the degree of constraint violation during the iteration process to guide the search towards the feasible region. The update rule adopts the multiplier update formula for inequality constraints in the augmented Lagrange method: ,in This is a penalty parameter, which can be increased proportionally in each outer iteration to accelerate convergence. Initially... It can be 0 or a small positive number. When hour, Gradually increase the weight of the constraint in the Lagrange function; when hour, Gradually reduce it to 0 to minimize the performance loss of the search focus.

[0152] The adjustable operating parameters of each Agent are updated along the negative direction of the weighted gradient to obtain the updated adjustable operating parameters. A suitable update step size is determined using a line search or adaptive step size strategy to ensure that the objective function decreases or the constraint violation amount is reduced, and the parameters are projected back into their respective physical feasible regions (e.g., the torque factor is limited to the (0,1] interval). Based on the updated adjustable operating parameters and the preset mapping relationship, the updated resource request vectors of each Agent are recalculated. The updated resource request vectors are then superimposed to obtain a new total occupancy vector. The global conflict index calculation process in S6 is then re-executed to obtain the updated... value.

[0153] Repeat the update and recalculation process described above until the termination condition is met. The termination condition includes the global conflict index. No more than (Successfully resolved conflict) or the number of iterations reaches the preset maximum limit (to avoid infinite loops, in which case the iteration process is selected). The solution that minimizes both the minimum and the minimum constraint violation is taken as the approximate solution. The final updated resource request vectors of each agent in the target agent scheme that satisfies or approximately satisfies the constraints are then superimposed to obtain the final total occupancy vector.

[0154] Based on the resource request vectors of each Agent in the final total resource occupancy vector, corresponding scheduling instructions are generated. The scheduling instructions include the Agent identifier, the start and end time slots of the task operation, the target spatial location or path coordinate sequence, the function mode code, and suggested adjustable operating parameter values. The scheduling controller sends the scheduling instructions one by one to the local controller of each Agent via industrial Ethernet. After receiving the instructions, each Agent executes the task according to the specified parameters within the specified time window.

[0155] This embodiment, through the gradient optimization mechanism based on the Lagrange multiplier method, endows the scheduling system with the ability to proactively resolve cross-dimensional implicit conflicts. This allows for fine-tuning of Agent operating parameters (using dimensionality reduction) to satisfy physical field coupling constraints during the task allocation decision-making stage, rather than simply rejecting or postponing tasks. This maximizes the parallel operation density of heterogeneous Agents while ensuring production line safety. The gradient calculation in the adjustment process reuses the low-rank factor matrix from the offline stage, resulting in extremely low online computational overhead, fully meeting the real-time requirements of industrial scheduling.

[0156] The aforementioned agent-based dynamic scheduling and control method for industrial robots constructs a global resource space encompassing spatial, temporal, functional mode, and physical field strength dimensions, instantiating production line resource units as resource nodes within this space. Based on this, dimensional masks are generated for various heterogeneous agents to identify the dimensions that impose substantial constraints when generating resource requests. This quantifies and distinguishes the essential differences in dimensional perspectives between different agents such as AGVs, robotic arms, and vision inspection devices regarding the same physical resource. Subsequently, a physical field propagation model is used to determine the cross-dimensional coupling relationships between resource nodes caused by the propagation of physical fields such as vibration and electromagnetic fields. Nodes acting as radiation sources and their affected sink nodes are organized into hyperedges to construct a resource coupling hypergraph. This allows implicit conflicts that cannot be detected by single-dimensional detection (such as robotic arm vibration leading to loss of detection equipment accuracy, AGV passage intrusion into the robotic arm's working envelope, etc.) to be explicitly expressed as structured relationships within the hypergraph. Furthermore, physical field transmission coefficients are extracted from the hypergraph to construct a coupling matrix, which is then decomposed into a low-rank factor matrix, providing pre-compressed coupling information for the online scheduling phase. Upon receiving tasks to be scheduled online, the task requirements of each target agent are uniformly transformed into a one-dimensional resource request vector based on the same resource node space using a dimensionality mask. Then, by combining target agents and superimposing resource request vectors into a total occupancy vector, a low-rank factor matrix is ​​used to quickly calculate a global conflict index. This allows for real-time quantification of the degree of cross-dimensional coupling conflict in the resource allocation decision-making stage. If the global conflict index exceeds a threshold, the allocation scheme is not directly rejected. Instead, the adjustable operating parameters of the relevant agents are continuously adjusted to proactively resolve latent conflicts until the conflict index meets the constraints, at which point a scheduling instruction is generated and executed. This technical solution moves the detection of latent conflicts from the traditional post-deadlock formation stage to the task allocation decision-making stage. By continuously adjusting parameters, it empowers the scheduling system with the ability to proactively resolve cross-dimensional conflicts. This allows for high-density, secure hybridization of heterogeneous agents within a more compact spatiotemporal window without altering task semantics. This fundamentally avoids deadlocks, quality incidents, and unplanned downtime caused by cross-dimensional coupling in the physical field, significantly improving the stability of production line scheduling and equipment utilization.

[0157] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0158] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An agent-based dynamic task scheduling control method for industrial robots, characterized in that, The method includes: S1. Based on the independently allocable resource units in the production line, construct a global resource space including spatial dimension, time dimension, functional mode dimension and physical field intensity dimension, and instantiate the resource units as resource nodes in the global resource space; S2. Based on the resource usage characteristics of each type of Agent, generate a dimension mask to identify the dimensions that the corresponding Agent needs to specify when generating resource requests. S3. Based on the physical field propagation model, by determining the cross-dimensional coupling relationship caused by the propagation of the physical field among the resource nodes, the source nodes that are the radiation sources of the physical field and the sink nodes affected by the physical field are organized into hyperedges to obtain a resource coupling hypergraph. S4. Based on the transmission relationship from the source node to the sink node recorded by each hyperedge in the resource coupling hypergraph, generate a coupling matrix whose matrix elements represent the physical field transmission coefficients from the source node to the sink node; perform low-rank decomposition on the coupling matrix to obtain multiple sets of low-rank factor vectors, and combine the multiple sets of low-rank factor vectors into a low-rank factor matrix. S5. Receive the task to be scheduled, and generate a resource request vector for each target agent using the dimension mask according to the Agent type and task parameters required by the task to be scheduled; wherein, the elements of the resource request vector represent the occupation requirements of the corresponding resource nodes. S6. Combine the target agents to obtain a target agent scheme, and superimpose the resource request vectors of each agent in the target agent scheme to form a total occupancy vector; call the low-rank factor matrix to calculate the global conflict index of the target agent scheme based on the total occupancy vector; wherein, the global conflict index measures the degree of conflict caused by cross-dimensional coupling in the target agent scheme; S7. When the global conflict index exceeds the index threshold, adjust the adjustable operating parameters of at least one Agent in the target Agent scheme corresponding to the element of the resource request vector until the recalculated global conflict index does not exceed the index threshold, and obtain the final total occupancy vector; generate scheduling instructions according to the target Agent scheme and the corresponding final total occupancy vector and send them to the corresponding Agent.

2. The method according to claim 1, characterized in that, The step of generating a dimension mask based on the resource usage characteristics of each type of Agent, used to identify the dimensions that the corresponding Agent needs to specify when generating resource requests, includes: S21. Based on the pre-built ontology knowledge base, obtain the functional attributes and resource requirement descriptions of various Agents; S22. For each type of Agent, extract the Agent's requirements for spatial location, time window, functional mode, and physical field intensity when performing tasks, based on the functional attributes and resource requirement description. S23. Based on the requirement information, mark the dimension in the global resource space that the Agent needs to specify the value as the first identifier, and mark the other dimensions as the second identifier to generate a mask vector; S24. For the dimension marked as the second identifier in the mask vector, set a preset default value for the corresponding Agent in the corresponding dimension; associate and store the mask vector and the preset default value to obtain the dimension mask.

3. The method according to claim 1, characterized in that, The physical field propagation model determines the cross-dimensional coupling relationships between resource nodes caused by physical field propagation. It then organizes source nodes (which act as sources of physical field radiation) and sink nodes affected by the physical field into hyperedges, resulting in a resource coupling hypergraph, including: S31. Based on the operational attributes of various Agents, the resource nodes are divided into source nodes and sink nodes; S32. For each source node, based on the physical field propagation model, using the physical field radiation intensity of the source node and the spatial distance between the source node and each sink node, calculate the residual physical field intensity of the source node on each sink node. S33. Traverse all the source nodes. When the residual physical field strength exceeds the physical field tolerance threshold of the corresponding sink node, organize the source node and the sink node that exceeds the physical field tolerance threshold into a hyperedge to obtain the resource coupling hypergraph.

4. The method according to claim 3, characterized in that, The coupling matrix is ​​generated based on the transmission relationship from the source node to the sink node recorded by each hyperedge in the resource coupling hypergraph, and the matrix elements represent the physical field transmission coefficients from the source node to the sink node. The coupling matrix is ​​decomposed into multiple sets of low-rank factor vectors, and these multiple sets of low-rank factor vectors are combined into a low-rank factor matrix, including: S41. Based on the residual physical field intensity of each sink node recorded by each hyperedge in the resource coupling hypergraph, and combined with the physical field radiation intensity of the source node corresponding to the hyperedge, calculate the physical field transfer coefficient from the source node to each sink node under unit radiation intensity. S42. Construct a source-sink transfer matrix based on the physical field transfer coefficients; wherein the source-sink transfer matrix has the sink node as the row and the source node as the column, and the elements of the source-sink transfer matrix are the physical field transfer coefficients corresponding to the source node to the sink node; S43. Perform truncated singular value decomposition on the source-sink transit matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix; S44. Determine the rank K to be retained based on the cumulative energy ratio of the singular values ​​in the singular value diagonal matrix. Select the first K largest singular values ​​from the singular value diagonal matrix and extract the first K column vectors from the left singular vector matrix and the right singular vector matrix respectively to obtain the sequence of retained singular values, the group of retained left singular vectors, and the group of retained right singular vectors. S45. Multiply the square root of each singular value in the retained singular value sequence by the corresponding retained left singular vector and the retained right singular vector to obtain K sets of low-rank factor vectors; combine the K sets of low-rank factor vectors to obtain the low-rank factor matrix.

5. The method according to claim 2, characterized in that, The step of generating resource request vectors for each target agent using the dimensional mask based on the Agent type and task parameters required by the task to be scheduled includes: S51. For each target Agent, read the mask vector corresponding to the Agent class to which the target Agent belongs and the preset default value from the dimension mask; S52. The dimension in the mask vector marked with the first identifier is determined as the sensitive dimension, and the dimension in the mask vector marked with the second identifier is determined as the irrelevant dimension; according to the task parameters, the sensitive dimension is given a task-specified value, and the irrelevant dimension is given a preset default value, thus forming the multi-dimensional coordinates of the resource nodes to be occupied by the target Agent. S53. Query the resource node index that matches the multidimensional coordinates of the resource node in the global resource space to obtain the set of occupied node indexes, and initialize a one-dimensional zero vector with a length equal to the total number of resource nodes as the initial resource request vector. S54. Based on the set of occupied node indices, configure the element at the corresponding index position in the initial resource request vector as a value representing the occupied demand to obtain the intermediate resource request vector; S55. For an Agent with physical field radiation attributes, determine the physical field radiation source node index according to the operating mode of the target Agent; update the position of the physical field radiation source node index in the intermediate resource request vector to the nominal radiation intensity value to obtain the resource request vector.

6. The method according to claim 1, characterized in that, The step of calling the low-rank factor matrix and calculating the global conflict index of the target agent scheme based on the total occupancy vector includes: S61. Based on the source nodes and sink nodes divided in the resource coupling hypergraph, according to the elements corresponding to each source node in the total occupancy vector, a source node occupancy subvector is formed according to the arrangement order of the source nodes, and according to the elements corresponding to each sink node in the total occupancy vector, a sink node occupancy subvector is formed according to the arrangement order of the sink nodes. S62. Extract each set of factor vectors from the low-rank factor matrix; wherein each set of factor vectors consists of a source side vector corresponding to the source node and a sink side vector corresponding to the sink node; S63. For each group of factor vectors, calculate the inner product of the source side vector and the source node occupied sub-vector to obtain the intermediate variable corresponding to the factor vector. S64. For each sink node, the total radiation estimate of the corresponding sink node is calculated by using the sum of the products of the elements of the sink side vectors in each group of factor vectors at the corresponding sink node and the corresponding intermediate variables. S65. Based on the resource coupling hypergraph, obtain the physical field tolerance threshold of each sink node; determine the target sink node to be activated based on the sink node occupancy subvector; for each target sink node, calculate the penalty amount for the total radiation estimate exceeding the physical field tolerance threshold; sum the penalty amounts of all target sink nodes to obtain the global conflict index; wherein, the formula for calculating the global conflict index is: ； in, This represents the global conflict index. The total number of sink nodes. The total number of source nodes. The rank retained by performing low-rank decomposition on the coupling matrix. Indicates the first subvector occupied by the sink node The occupancy value of each sink node. Indicates the source node occupying the sub-vector of the first... The occupancy value of each source node. The th in the low-rank factor matrix The group of factor vectors at the sink node The element at that location, The th in the low-rank factor matrix The group of factor vectors at the source node The element at that location, For sink node The physical field tolerance threshold.

7. The method according to claim 1, characterized in that, The adjustment of adjustable operating parameters corresponding to elements of the resource request vector for at least one Agent in the target Agent scheme, until the recalculated global conflict index does not exceed the index threshold, to obtain the final total occupancy vector, includes: S71. Based on the preset mapping relationship between the adjustable operating parameters of each Agent in the target Agent scheme and the element values ​​of the resource request vector, and using the adjustable operating parameters of each Agent as optimization variables, construct an optimization problem with the objective function of minimizing the sum of squared Euclidean distances of the resource request vectors before and after adjustment, and with the constraint that the global conflict index does not exceed the index threshold. S72. Based on the optimization problem, calculate the gradient of the objective function with respect to the adjustable operating parameters as the first gradient; calculate the gradient of the global conflict index with respect to the adjustable operating parameters as the second gradient; and sum the first gradient and the second gradient with weights to obtain the weighted gradient. S73. Update the adjustable operating parameters along the search direction corresponding to the weighted gradient to obtain the updated adjustable operating parameters; update the resource request vector according to the updated adjustable operating parameters and the preset mapping relationship to obtain the updated resource request vector; recalculate the global conflict index according to the updated resource request vector. S74. Repeat S73 until the global conflict index does not exceed the index threshold, and then sum the updated resource request vectors of each Agent in the target Agent scheme to form the final total occupancy vector.

Images