A robot complex task dynamic scheduling method

Abstract

The application provides a robot complex task dynamic arrangement method, belongs to the field of robot task arrangement, and is used for solving the problems of weak dynamic environment adaptation capability of traditional static task arrangement, poor multi-robot collaboration and insufficient risk prediction in the related art. The method collects robot state and event data, constructs a double matrix, fuses the double matrix into a tensor, performs low-rank decomposition, dynamically generates and optimizes a threshold, calculates a task priority based on multiple factors, combines resource constraints and federated learning to allocate resources, synchronously adapts across scenes, pre-plans and optimizes visual feedback in execution, and realizes dynamic adaptation, efficient collaboration and stable execution of complex tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robot task scheduling, in particular to a robot complex task dynamic scheduling method. BACKGROUND

[0002] With the development of industrial automation and service intelligence, the demand for robots to perform multiple tasks in complex scenarios is increasing, such as assembly of flexible production lines, sorting and distribution of intelligent logistics parks, etc. Robot task scheduling, as a core technology for coordinating robot actions and optimizing resource utilization, directly affects the efficiency of task execution and system adaptability.

[0003] Currently, robot complex task scheduling mostly adopts traditional preset process or fixed instruction set method, i.e. according to pre-set task logic and environmental parameters, fixed execution steps and resources are allocated to robots. Such methods can achieve basic functions in stable environment and single task scenarios, but in dynamic and complex scenarios, they still rely on manual preset rules to adjust the process.

[0004] However, the traditional method has significant defects: first, it is difficult to adapt to changes in dynamic environment and task requirements. When encountering sensor data anomalies, task priority changes or resource conflicts, it cannot adjust the task process in real time, which easily leads to low execution efficiency or task failure; second, it lacks deep quantitative mining of the correlation between state, event and task, which cannot accurately assess the co-occurrence correlation of state and event and the impact of event on state, resulting in unreasonable resource allocation and poor task coordination; third, it has weak cross-scene adaptation ability, and needs to rebuild the scheduling model in new scenarios, and lacks feedback optimization mechanism for execution results, making it difficult to continuously improve system performance. These defects directly limit the application effect of robots in complex dynamic scenarios. SUMMARY

[0005] The present application provides a robot complex task dynamic scheduling method, which can realize dynamic and efficient scheduling of complex tasks through quantitative correlation modeling, dynamic threshold adjustment and full-link collaboration, and improve the adaptability and execution efficiency of the system in dynamic environment.

[0006] The application provides a robot complex task dynamic scheduling method. It includes: data acquisition step, collecting robot running state data and event data; association modeling step, constructing state-event association matrix and event-state influence matrix, wherein the state-event association matrix quantifies the co-occurrence association strength of the state and the event by counting the number of simultaneous occurrences of a specific state and a specific event, the total number of occurrences of the state alone, and the total number of occurrences of the event alone, and using the intersection ratio correction method, the event-state influence matrix quantifies the influence amplitude of the event on the state by calculating the change amplitude of the specific state after the specific event is triggered and normalizing the change amplitude with the maximum change amplitude of all events on the state; threshold adjustment step, dynamically generating clustering threshold and filtering threshold based on the state-event association matrix and event-state influence matrix, wherein the clustering threshold is determined by superimposing the basic threshold and the maximum association weight of the event and all states, and the filtering threshold is determined by superimposing the minimum threshold and the sum of the influence degrees of all events on the state; task priority calculation step, combining the time sensitivity coefficient corresponding to the task deadline, the task resource demand proportion, the task urgency coefficient, and the product sum of the association degree of the task corresponding state and effective event and the influence degree of the effective event on the state, and calculating the task priority in a weighted manner; resource allocation step, resource allocation based on the task priority and resource constraints, the resource constraints being that the total resource consumption of each task in the aggregated state does not exceed the total system resource; process execution step, executing the task process according to the task priority and resource allocation result.

[0007] By adopting the above technical solution, real-time data acquisition, quantitative association modeling, dynamic threshold adjustment, multi-factor priority calculation, and resource constraint allocation are combined, breaking through the limitations of traditional preset processes, realizing precise quantification of state-event association and dynamic adjustment of task process, effectively solving the problems of poor task adaptability and low execution efficiency in dynamic environment, and improving the intelligence and accuracy of robot complex task scheduling.

[0008] Further, the association modeling step further includes fusing the state-event association matrix and the event-state influence matrix into a state-event-time three-dimensional tensor, wherein any "state-event-time" element value in the three-dimensional tensor is calculated by geometric average of the association degree of the state and the event and the influence degree of the event on the state, to uniformly quantify the dynamic association relationship among state, event and time.

[0009] By adopting the above technical solution, the originally independent association matrix and influence matrix are integrated into a unified space-time association carrier, eliminating the consistency conflict of double-matrix independent modeling, while retaining the dynamic association characteristics of state, event and time, providing more comprehensive association basis for subsequent scheduling optimization.

[0010] Further, the association modeling step further comprises decomposing the three-dimensional tensor into a product of a state core tensor, an association core tensor and a time core tensor using a low-rank decomposition algorithm, wherein the state core tensor represents low-dimensional characteristics of the state dimension, the association core tensor represents low-dimensional characteristics of the state-event interaction, and the time core tensor represents low-dimensional characteristics of the time dimension, so as to reduce the computational complexity of the three-dimensional tensor from a three-order complexity of the number of states, the number of events and the number of time steps to a linear complexity.

[0011] By adopting the above technical solution, the computational complexity in a high-dimensional scenario is greatly reduced, the computational delay caused by excessively high state and event dimensions is avoided, the real-time performance in a large-scale robot cluster or a high-dynamic scenario is ensured, and support is provided for efficient operation of the system.

[0012] Further, the association modeling step further comprises dynamically adjusting a rank parameter of the low-rank decomposition based on a task priority, for a high-priority task, increasing the rank parameter to increase the feature dimension of the core tensor, and retaining detailed information of the state-event association; and for a low-priority task, reducing the rank parameter to reduce the feature dimension of the core tensor, and improving the computational efficiency within an acceptable information loss range.

[0013] By adopting the above technical solution, the rank parameter is accurately adapted to the task demand, the association information integrity of high-priority tasks (such as high-precision assembly) is ensured, the computational efficiency of low-priority tasks (such as routine inspection) is improved, and the relationship between precision and efficiency is balanced.

[0014] Further, the threshold adjustment step further comprises autonomously optimizing the threshold coefficient by an intelligent optimization algorithm, the intelligent optimization algorithm constructs a weighted summation objective function with a task completion rate, a resource utilization rate and an information loss rate as targets, maximizes the objective function value by iteratively adjusting the threshold coefficient, and realizes multi-objective balance of task efficiency, resource utilization and information retention.

[0015] By adopting the above technical solution, the traditional way of manually setting the threshold coefficient is replaced, the autonomous optimization of the threshold coefficient is realized, the deviation caused by experience dependence is avoided, and the multi-objective demand is balanced, thereby improving the rationality of threshold adjustment and the overall performance of the system.

[0016] Further, the threshold adjustment step further comprises constructing a multi-robot interaction relationship graph, taking each robot as a graph node, the graph node features including a current threshold coefficient of the robot, a task list and a fault risk; calculating the similarity of the features of two robot nodes by a graph attention mechanism, and normalizing the similarity to obtain a task dependency degree of the two robots; and dynamically adjusting a threshold conflict correction strategy based on the task dependency degree, and using a smaller conflict correction factor for a robot with a high task dependency degree.

[0017] By adopting the technical scheme, the task dependency relationship among multiple robots is accurately captured, task coordination is prevented from being broken due to threshold conflict correction, threshold adjustment coordination in a heterogeneous scene of multiple robots is improved, and resource competition and task delay are reduced.

[0018] Further, the cross-scene adaptation step further comprises: extracting visual image features, laser point cloud features and task text features of a new scene; taking the task text features as anchor points, calculating similarities of the visual image features, the laser point cloud features and the task text features, normalizing the similarities to obtain attention weights of the modal features, aligning the multi-modal features based on the attention weights to generate a scene implicit representation; and calculating similarities of the scene implicit representation and historical scene implicit representations, and transferring associated model parameters of the historical scenes based on the similarities.

[0019] By adopting the technical scheme, the implicit association between scenes is mined, the new scene "cold start" problem is solved, the workload and data dependence of model construction in a new scene are reduced, and the cross-scene adaptation efficiency and accuracy are improved.

[0020] Further, the pre-planning step further comprises: extracting a historical information gain sequence, a state-event correlation degree change trend and a task complexity change feature, constructing a time series prediction model, separating long-term trend features and periodic features of the information gain, and focusing on key historical time steps through an attention mechanism to predict information gain at future time steps; collecting robot motor current, temperature and sensor signal strength features, giving high weights to key fault features through an attention mechanism, calculating errors between input features and reconstructed features, and deviations between encoding implicit features and normal sample implicit features using an encoding-decoding model, and mapping the errors and deviations to a robot failure probability after weighting; when the predicted information gain is lower than a set threshold or the failure probability is higher than a set threshold, generating an alternative process including a transition task, and similarities between state features of the transition task, a current execution task and a target task are all not lower than a set threshold.

[0021] By adopting the technical scheme, early warning and pre-planning of "information deficiency + failure risk" are realized, task interruption is avoided, and process smooth connection is ensured through a transition task, thereby improving stability and robustness of task execution.

[0022] Further, the resource allocation step further comprises aggregating local resource parameters of multiple robots using a federated learning framework, and the aggregation process comprises: first, removing extreme parameters that are more than ±2 times the standard deviation of the local parameters, and then performing weighted averaging on the remaining parameters with the robot task completion rate as the weight; before uploading the local parameters, adding Laplacian noise to the parameters, and the noise scale is determined by the maximum change range of the parameters, the privacy protection level and the number of robots.

[0023] By adopting the technical scheme, the robustness of resource aggregation in a heterogeneous data scene is improved, extreme parameter interference is avoided, data privacy and security are ensured, an industrial scene data security specification is met, and the dual goals of "robust aggregation + privacy protection" are achieved.

[0024] Further, a visual feedback step is further included, which maps the task nodes, event nodes and resource nodes into a dynamic flowchart, the task nodes are marked with the task priority calculated based on claim 1, the proportion of task resource demand in total system resources, and the associated influence product of the task corresponding state and effective event, the event nodes are marked with the influence degree of the event on the state, the proportion of task delay time in total task time length is taken as an increment of the dynamic rank adjustment error threshold in the association modeling step, and the proportion of task delay time in the deadline time is taken as a deduction item of the intelligent optimization algorithm reward function in the threshold adjustment step, so as to realize feedback optimization of the execution result on the model parameters.

[0025] By adopting the technical scheme, visual monitoring of the task flow is realized, manual intervention is facilitated, a closed loop of "execution-feedback-optimization" is constructed, the parameters of association modeling and threshold adjustment are continuously iteratively optimized, and the long-term running performance of the system is improved.

[0026] In summary, the present application at least has the following beneficial effects:

[0027] 1. A full-link cooperative robot complex task dynamic scheduling method is provided, which realizes efficient task scheduling in a dynamic environment.

[0028] 2. Tensor association modeling and dynamic rank adjustment are used to balance the precision and computational efficiency of association information.

[0029] 3. Multi-agent threshold optimization and federated resource aggregation are used to improve multi-robot cooperation and data privacy security.

[0030] 4. Cross-scene migration and double-risk pre-planning are used to enhance system adaptability and task execution robustness.

[0031] It should be understood that the content described in the summary section is not intended to limit the key or important features of the embodiments of the present application, nor to limit the scope of the present application. Other features of the present application will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0032] The above and other features, advantages and aspects of the embodiments of the present application will become more apparent by describing in detail preferred embodiments thereof with reference to the attached drawings in which:

[0033] Figure 1An exemplary operating environment in which embodiments of the present application can be implemented is shown.

[0034] Figure 2 A flowchart of a robot complex task dynamic scheduling method in embodiments of the present application is shown. DETAILED DESCRIPTION

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in detail with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0036] In addition, the term "and / or" herein merely describes an association relationship of associated objects, and indicates that there can be three relationships, for example, A and / or B can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects.

[0037] The present application provides a robot complex task dynamic scheduling method, which can quantify state-event association, dynamically adjust threshold and task priority, adapt to cross-scenario and multi-robot collaboration, and also can pre-control information deficiency and failure risk, effectively improving task execution efficiency, system robustness and adaptability in dynamic environment.

[0038] Figure 1 An exemplary operating environment in which embodiments of the present application can be implemented is shown.

[0039] Referring to Figure 1 The operating environment includes a hardware system supporting the full-link execution of the "robot complex task dynamic scheduling method", which takes "data acquisition-computing processing-collaborative communication-task execution-visualization monitoring" as the core logic, covers five components of perception layer devices, computing layer devices, communication layer devices, execution layer devices, and visualization and monitoring layer devices, and forms a closed loop through data transmission and instruction interaction, to meet the hardware requirements of the core functions such as dynamic association modeling, multi-robot collaboration, cross-scenario migration, and double risk pre-planning in the method.

[0040] The perception layer device is an input source for data collection, responsible for acquiring robot state, environmental information, and multi-modal scene features, including robot body sensors, environmental perception devices, and multi-modal data collection devices. Among them, the robot body sensors (such as GPS, motor current sensor, load sensor) are integrated on each robot, collecting real-time position, running state, and task execution progress; environmental perception devices (such as distributed laser radar, visual camera) are deployed in the scene, collecting global obstacle distribution, environmental anomalies, and other data; multi-modal data collection devices (such as high-definition industrial cameras, 3D laser scanners, text instruction receiving modules) are used to extract visual images, laser point clouds, task texts, and other multi-modal features. The perception layer device transmits the collected raw data to the computing layer device through the communication link, providing data support for subsequent association modeling (building state-event association matrix, event-state influence matrix), fault prediction (extracting motor current, temperature, and other fault features), and cross-scene adaptation (multi-modal feature migration).

[0041] The computing layer device is the power center of algorithm running, adopting a "local computing + cloud / edge computing" hierarchical architecture, including robot local computing unit, cloud / edge computing node, and storage device. The robot local computing unit (such as ARMCortex-A embedded processor, FPGA) is integrated on the robot, processing low-latency tasks (such as local dynamic threshold preliminary calculation, real-time action instruction analysis); the cloud / edge computing node deploys high-performance computing clusters (including multi-core CPU, GPU / TPU accelerator), undertakes core algorithm calculation, including large-scale tensor modeling (tensor train decomposition, dynamic rank adjustment), multi-objective optimization (multi-objective deep deterministic policy gradient), long-term prediction (time series fusion Transformer prediction information gain), and executes multi-robot local parameter robust aggregation and differential privacy protection (adding Laplace noise) as a federal learning server; the storage device (such as SSD array, distributed file system) stores historical scene data, model parameters (state-event association matrix, tensor decomposition core tensor), supporting contrast learning and variational Bayesian inference in cross-scene migration. The computing layer device receives the data transmitted by the perception layer, generates orchestration instructions after algorithm processing, and issues them to the execution layer device through the communication layer.

[0042] The communication layer device is a data link for multi-agent collaboration, ensuring low-latency and high-reliability data transmission, including an industrial communication network, an edge communication gateway, and a secure communication module. The industrial communication network (such as Profinet industrial Ethernet, 5G wireless network) enables millisecond-level data transmission between robots and the cloud / edge, and between robots, supporting real-time threshold conflict correction and global orchestration instruction issuance. The edge communication gateway is deployed in large-scale robot cluster scenarios to achieve local data preprocessing (filtering redundant data) and traffic scheduling, reducing the communication pressure on the cloud and ensuring the sparse upload of local parameters in federated learning. The secure communication module (such as SSL / TLS encryption protocol, national encryption algorithm) encrypts the transmission data to protect the privacy and security of federated learning parameters and sensitive task data (such as process paths). The communication layer device connects the perception layer, the computing layer, the execution layer, and the visualization and monitoring layer, enabling safe and efficient transmission of data and instructions across the entire link.

[0043] The execution layer device is a physical carrier for task implementation, including heterogeneous robots and auxiliary execution devices. Heterogeneous robots are configured according to scene requirements (such as assembly robots, logistics robots, sorting robots), each robot supports real-time reception of orchestration instructions (task execution order, resource allocation result) issued by the computing layer, and dynamically adjusts actions (such as switching tasks according to pre-planned alternative processes, quickly avoiding obstacles in case of failure); auxiliary execution devices (such as charging stations, material temporary storage tables, high-precision positioning landmarks) are deployed in the scene to provide energy supply, material storage, and position correction for robots, supporting smooth task transition (such as transition task execution in failure pre-planning). The execution layer device receives orchestration instructions transmitted by the communication layer, completes specific task execution, and feeds back execution results (task completion rate, delay time) to the computing layer and visualization layer through the communication layer.

[0044] The visualization and monitoring layer device is a feedback terminal for human-machine interaction and closed-loop optimization, including industrial monitoring terminals and data feedback interface devices. Industrial monitoring terminals (such as industrial display screens, industrial computers) display dynamic BPMN flowcharts in real time, labeling task priorities, failure risks, resource requirements, and other information, supporting manual viewing and intervention; data feedback interface devices convert task execution results (such as delay time proportion) into feedback signals, which are transmitted to the computing layer through the communication layer, driving associated modeling (dynamic rank adjustment error threshold correction) and threshold adjustment (multi-objective optimization reward function deduction) closed-loop optimization.

[0045] The above five types of devices form a complete hardware support system through the logical collaboration of "perception layer data collection → communication layer data transmission → computing layer data processing and instruction generation → communication layer instruction issuance → execution layer task execution → execution results feedback to the computing layer and visualization layer", ensuring that the "dynamic orchestration method for complex robot tasks" can be efficiently implemented in complex conditions such as dynamic environments, multi-robot heterogeneity, and cross-scenario.

[0046] The embodiment of the application discloses a robot complex task dynamic scheduling method.

[0047] Figure 2 The flow chart of the robot complex task dynamic scheduling method in the embodiment of the application is shown.

[0048] Referring to Figure 2 , the method specifically comprises the following steps:

[0049] S1: a data acquisition step of acquiring the running state data and event data of the robot.

[0050] In the method of the present step, only the operation of acquiring the running state data and event data of the robot needs to be performed, so as to provide original data input for the subsequent steps.

[0051] The running state data specifically comprises real-time position data of the robot , residual capacity data , load rate data and motor temperature data , which are respectively acquired through the GPS sensor (or laser SLAM sensor), capacity sensor, weight sensor and temperature sensor carried by the robot; the event data specifically comprises "obstacle detection event", "task completion event", "low capacity early warning event" and the like, which are generated by triggering the sensor data. In order to ensure the data precision, the original state data acquired needs to be subjected to noise modeling and preliminary preprocessing, and the mathematical model of the original state data is as follows:

[0052]

[0053] In the formula, is the position data collected at t (unit: m), is the real position of the robot at t (obtained by high-precision calibration equipment), is the Gaussian noise of the position sensor, satisfying is the noise variance (the value is determined by the model of the GPS sensor, such as of the industrial-grade GPS); is the residual capacity data collected at t (unit: %), is the real residual capacity of the robot at t, is the noise of the capacity sensor, satisfying (the value is determined by the precision parameter of the capacity sensor). The data acquisition frequency is set to (i.e. collect once every 0.1 second) to meet the sampling requirement of the "state-event-time" three-dimensional tensor time dimension in the subsequent association modeling step. For the original data with noise, a moving average filter is used for preprocessing. Taking the position data as an example, the filtering formula is:

[0054]

[0055] wherein, is the filtered position data, is the size of the sliding window (i.e. take the last 5 collected data), is the collection period, which can effectively reduce the interference of noise on the subsequent association calculation.

[0056] The triggering of event data needs to meet the preset conditions. Taking the "obstacle detection event" as an example, when the obstacle distance data collected by the laser radar , the event is triggered, wherein is the obstacle distance threshold (set according to the safety requirements of the robot movement); the triggering condition of the "low power warning event" is (set according to the endurance characteristics of the robot). The event data after triggering needs to record the event type type , the triggering time and the key parameters (such as obstacle distance , remaining power ), forming an event triple , which provides the original data of the event dimension for the construction of the state-event association matrix in the subsequent association modeling step; the preprocessed running state data is used as the original data of the state dimension, which jointly supports the construction of the state-event association matrix and the event-state influence matrix in the subsequent association modeling step.

[0057] S2: association modeling step, constructing state-event association matrix and event-state influence matrix.

[0058] The state-event association matrix is constructed by counting the number of simultaneous occurrences of a specific state and a specific event, the total number of occurrences of the state alone, and the total number of occurrences of the event alone, and the co-occurrence association strength of the state and the event is quantified by using the intersection ratio correction method. The event-state influence matrix is constructed by calculating the change amplitude of a specific state after a specific event is triggered, and normalizing the change amplitude with the maximum change amplitude of all events on the state, to quantify the influence amplitude of the event on the state.

[0059] In the construction of the state-event association matrix (denoted as matrix, the dimension is , wherein is the number of state types, For the number of event types, first, the running state data and event data after preprocessing in S1 step need to be counted: let be the number of times that the state (such as "low power" ) and the event (such as "low power warning event") appear simultaneously in the statistical window (the statistical window length matches the data collection period, set to 10 seconds, corresponding to 100 collection steps), count be the total number of times that the state appears alone, be the total number of times that the event appears alone, and all of them are obtained by traversing the state sequence and the event sequence output by S1. The correlation strength calculation formula using intersection ratio correction is:

[0060]

[0061] In the formula, the denominator count is the total number of times that the state or the event appears, to avoid misjudgment of the correlation degree due to the high frequency of the state or event appearing alone, and the larger the value is, the stronger the co-occurrence correlation between the state and the event is.

[0062] When constructing the event-state influence matrix (denoted as matrix, the dimension is ), the dynamic influence of the event on the state needs to be calculated: let be the change amplitude of the state in the unit time (consistent with the collection period, 0.1 seconds) after the event is triggered, such as when the event is "obstacle detection event", the state is "robot moving speed", that is, the difference between the speed after triggering and the speed before triggering, which is obtained by comparing the state data before and after S1 collection; be the maximum change amplitude of the state caused by all events , which is obtained by counting the historical event-triggering records (such as event-state change data in the past 1 hour). The normalized influence amplitude calculation formula is:

[0063]

[0064] wherein, the larger the value, the stronger the influence of the event on the state . The matrix and the matrix form a complement, respectively quantifying the state-event relationship from the two dimensions of "co-occurrence association" and "dynamic influence".

[0065] The method of this step further comprises fusing the state-event association matrix and the event-state influence matrix into a state-event-time three-dimensional tensor, wherein the value of any "state-event-time" element in the three-dimensional tensor is calculated by the geometric mean of the association degree of the state and the event and the influence degree of the event on the state, to uniformly quantify the dynamic association relationship among the state, the event and the time; the three-dimensional tensor here is denoted as , with dimensions of , wherein is the number of time steps (such as , corresponding to a time sequence of 10 seconds), is the time of the th time step ( ). The calculation formula of the tensor element (corresponding to the state , the event and the time ) is:

[0066]

[0067] wherein, is the element value of the matrix at the time (the element value is updated in real time by a sliding window with a window step of 1 time step), is the element value of the matrix at the time , the geometric mean operation can balance the weights of the association degree and the influence degree, avoid the information of one dimension being covered by the excessively large value of the other dimension, and realize the dynamic unified representation of the state-event relationship in the time dimension. It further comprises decomposing the three-dimensional tensor into the product of a state core tensor, an association core tensor and a time core tensor by using a low-rank decomposition algorithm, wherein the state core tensor represents the low-dimensional characteristics of the state dimension, the association core tensor represents the low-dimensional characteristics of the state and event interaction, and the time core tensor represents the low-dimensional characteristics of the time dimension, so as to reduce the calculation complexity of the three-dimensional tensor from the three-order complexity of the number of states, the number of events and the number of time steps to linear complexity; here, a tensor train (TT) decomposition algorithm is used, and the decomposition formula is:

[0068]

[0069] ​​​

[0070] In the formula, For the state core tensor, Its rank parameter (initially set to 10) is used, and each row corresponds to a low-dimensional feature vector of a state; To associate the core tensor, Its rank parameter (initially set to 10) characterizes the interaction features between state and event; This is a time-core tensor, with each column corresponding to an eigenvector of a time step. The computational complexity of decomposing the first three-dimensional tensor is O(n log n). ,like (Number of state types) (Number of event types) When the time steps are , the complexity is O(n). The complexity is reduced to [value] after decomposition. This significantly reduces the pressure of high-dimensional data processing and meets real-time requirements.

[0071] It also includes dynamically adjusting the rank parameter of the low-rank decomposition based on task priority. For high-priority tasks, the rank parameter is increased to increase the feature dimension of the core tensor, preserving detailed information about state-event associations; for low-priority tasks, the rank parameter is decreased to reduce the feature dimension of the core tensor, improving computational efficiency within acceptable information loss limits. Task priority ( The value is calculated through subsequent steps S4 (for the current task), and its range is [value range missing]. ,in Defined as a high-priority task (such as a precision assembly task). Defined as a low-priority task (such as a routine inspection task). The rank parameter adjustment rule is as follows:

[0072] like And currently ( (where the maximum rank threshold is), then ;

[0073] like And currently ( (where the minimum rank threshold is used), then ;

[0074] like or If so, the rank parameter remains unchanged.

[0075] To avoid re-decomposing the tensor completely after rank adjustment, an incremental update strategy is adopted: when the rank parameter changes from... Upgraded to At that time, only the newly added factor vectors need to be optimized (such as...) The objective function to be optimized is:

[0076]

[0077] wherein, is the rank of the reconstructed tensor, is the tensor outer product operation, i.e. (vector), (matrix), (vector) are superimposed after being extended to the same dimension tensor through outer product, ensuring consistency with the X dimension, is the Frobenius norm, this strategy can reduce the calculation time of rank adjustment above, while ensuring the reconstruction accuracy (reconstruction error , meeting the task requirements).

[0078] S3: Threshold adjustment step, dynamically generating clustering threshold and filtering threshold based on the state-event association matrix and event-state influence matrix.

[0079] The clustering threshold is determined by superimposing the basic threshold on the maximum association weight of the event and all states, and the filtering threshold is determined by superimposing the minimum threshold on the sum of the influence degree of all events on the state.

[0080] The clustering threshold here is used to aggregate the operating state data collected in S1 (such as grouping "low power " and "low power " into the same low power state), and the filtering threshold is used to filter the event data generated in S1 (such as eliminating redundant events that have a small impact on the state). Let be the clustering basic threshold (set according to the state dimension, the value range is , such as the of the location state), be the clustering threshold adjustment coefficient (to be optimized variable, initial value , for the first class event , its corresponding clustering threshold The calculation formula is:

[0081] wherein, is the maximum association weight of the event and all states (from the matrix constructed in S2), the larger the association weight, the higher the clustering threshold, which can avoid excessive aggregation leading to loss of state details.

[0082] Let be the filtering minimum threshold (set according to the event type, the value range is , such as the of the environmental event), The filter threshold adjustment coefficient (the variable to be optimized, initial value) ), for the first Class State The corresponding filtering threshold The calculation formula is:

[0083] In the formula, For all events ( ) for state Total influence (from S2 construction) The larger the total influence of the matrix, the higher the filtering threshold, and events that have a significant impact on the state can be retained first.

[0084] The method in this step specifically includes: further, it includes autonomously optimizing the threshold coefficient using an intelligent optimization algorithm. This intelligent optimization algorithm constructs a weighted summation objective function with the goals of task completion rate, resource utilization rate, and information loss rate. By iteratively adjusting the threshold coefficient, the objective function value is maximized, achieving a multi-objective balance between task efficiency, resource utilization, and information retention. The threshold coefficient here is the one described above. (Clustering threshold adjustment coefficient) and (Filtering threshold adjustment coefficient) is optimized using the Multi-Objective Deep Deterministic Policy Gradient (MO-DDPG) algorithm. First, the objective function is defined. :

[0085] In the formula, Weighting coefficient (dynamically adjusted based on task priority; higher priority tasks can have their weighting increased). TaskRate represents the task completion rate. (Real-time statistics from S6 process execution steps); TaskRate represents the task completion rate. (Real-time statistics from S6 process execution steps); Redundancy is the rate of redundant events. (Valid events refer to events filtered by the threshold); LossInfo is the loss rate of aggregated state information. ,in The entropy of the state before aggregation. The entropy of the aggregated state ( For state The probability of its occurrence is derived from S1 data statistics.

[0086] The algorithm adjusts iteratively. and make Maximize, in each iteration, the current state (e.g.) Matrix mean Matrix mean Resource surplus rate Input policy network, output and The adjustment amount, and then based on the objective function. The feedback updates the network parameters until Converging to a stable value (fluctuation range) ).

[0087] It also includes constructing a multi-robot interaction graph, with each robot as a graph node. The graph node features include the robot's current threshold coefficient, task list, and fault risk. A graph attention mechanism is used to calculate the similarity of the node features between two robots, and this similarity is normalized to obtain the task dependency between the two robots. Based on the task dependency, a threshold conflict correction strategy is dynamically adjusted, using a smaller conflict correction factor for robots with high task dependency. Assume the scenario includes... The robot, the first The node characteristics of the robot are ,in For robots The current threshold coefficient, Its task list is represented by a 4-dimensional embedded encoding (type / priority / time / resource) (e.g., task type and priority, derived from the S4 task priority calculation steps). Its failure probability (failure prediction from the S8 pre-planning step). Dimension set to ( Each is 1 dimension. Encoded in 4 dimensions, 1D (with 1D expansion reserved).

[0088] First, through the linear mapping matrix ( (reducing dimensionality to a lower-dimensional space) and Perform feature mapping to obtain and Then through attention vectors Calculate similarity:

[0089]

[0090] In the formula, This is a vector concatenation operation, with LeakyReLU as the activation function (slope set to 0.2). For robots and The original similarity is then normalized using Softmax to obtain the attention weights (i.e., task dependency).

[0091]

[0092]

[0093] In the formula, is the neighbor node set of the robot (the robot that has task dependency, such as the robot that has the material taking task dependency robot that has the material carrying task), The greater the value, the higher the dependency.

[0094] When multiple robots have threshold conflicts due to resource competition (such as simultaneously applying for a charging station), first calculate the global conflict degree , wherein is the number of robots applying for resources at time t, is the total number of resources (such as the number of charging stations); then calculate the basic correction factor according to the conflict degree:

[0095]

[0096] In the formula, Conf is the conflict threshold, Finally, adjust the conflict correction factor of the robot in combination with the task dependency:

[0097]

[0098] In the formula, is the maximum dependency of the robot and the neighbor node, the higher the dependency, the smaller, the correction is more moderate, avoiding the task dependency from being broken due to the dramatic adjustment of the threshold (such as the material taking task of the robot being affected by the large adjustment of the material carrying threshold of the robot ). The adjusted threshold coefficient is , ensuring the threshold collaborative optimization of multiple robots.

[0099] S4: Task priority calculation step, combining the time sensitivity coefficient corresponding to the task deadline, the task resource demand proportion, the task urgency coefficient, and the product sum of the association degree of the task corresponding state and the effective event and the influence degree of the effective event on the state, the task priority is calculated in a weighted manner.

[0100] In the method of this step, the task priority is calculated according to the above-mentioned multi-factor weighting method, which provides priority basis for subsequent resource allocation. First, the specific definition, value source and mathematical calculation logic of each calculation factor are defined to ensure that all parameters can be obtained from the output of the previous step or the system preset configuration, and the objectivity and traceability of the priority calculation are guaranteed:

[0101] For the time-sensitive coefficient corresponding to the task deadline ( representing the th task to be calculated), the core is to quantify the urgency of the task due to the passage of time. Let the current system time be (from the system real-time clock module, unit: seconds), the preset deadline of task is (from the parameter configuration in the task initialization instruction, such as the assembly task deadline seconds), and the remaining time of the task is calculated first . In order to avoid the distortion of the coefficient caused by negative remaining time (task timeout) or too large remaining time (long-term task), the exponential normalization formula is used to calculate:

[0102] In the formula, (reference time threshold, preset to 1 hour, which can be adjusted according to the scene, such as 1800 seconds in the peak season of logistics), ; the shorter the remaining time, the closer the coefficient is to 1.0, and the coefficient of the overtime task is fixed at 1.0, ensuring that time-sensitive tasks are prioritized.

[0103] For the task resource demand ratio , the demand intensity of the task for multiple types of core resources of the system needs to be considered. The system core resources include types (such as power , computing power , device occupation time ), and the total available amount of the th type of resource in the system is (from the system resource configuration table, such as total power , total computing power ), and the demand of task for the th type of resource is (from the resource application parameters in the task instruction, such as the assembly task power , computing power ). First, calculate the single-resource demand ratio , then take the arithmetic mean of the resource ratios to get:

[0104]

[0105] In the formula, The higher the resource demand ratio, the larger the coefficient. Tasks that significantly consume system resources will be prioritized to avoid excessive resource concentration that would cause other tasks to have "no resources available".

[0106] For the task urgency coefficient This coefficient is an inherent attribute of the task and is preset by the system or user according to the task type: high-urgency tasks (such as equipment failure repair, emergency material delivery) are set to [value missing]. Routine tasks (such as daily sorting and routine assembly) are set as Low-urgency tasks (such as data backup and equipment inspection) are set to For example, "robot motor fault repair task" This is directly used as one of the weighting factors to reflect the differences in the urgency and priority of tasks.

[0107] The sum of the products of the correlation between the task's corresponding state and the effective event, and the influence of the effective event on that state. The calculation needs to be combined with the output results of the previous steps: First, determine the task. Core state set (For example, the core state of the "material grabbing task" is the "robotic arm position state") "Grasping force status" ), valid event set For events filtered by the S3 threshold (i.e., those that meet the criteria) The incident, Event-state influence matrix from S2 (Filter threshold calculation from S3). For each core state Calculate its relationship with all valid events. "relevance" The product of influence, then... The average of the product of the core states is taken:

[0108] In the formula, For state With the event Co-occurrence correlation (from the state-event correlation matrix of S2). For the event State The magnitude of the impact (from the event-state impact matrix of S2). The larger the sum of the products, the closer the task is to the effective events in the current system. Prioritizing the execution of these tasks can better utilize the current state-event interaction information and reduce uncertainty in task execution.

[0109] Finally, the final priority of the task is calculated by weighted summation , with the weights of the four factors being (time sensitivity), (resource requirement), (emergency level), and (associated impact), satisfying (the weights can be dynamically adjusted according to the scene, such as the production workshop can increase to 0.3 and decrease to 0.2), and the calculation formula is:

[0110]

[0111] wherein , the larger the priority value, the more the task needs to be executed in priority; after the calculation is completed, all the tasks to be executed are sorted in descending order according to , and a task priority sequence is generated, providing a clear scheduling basis for the resource allocation step of S5.

[0112] S5: Resource allocation step, resource allocation is performed based on the task priority and resource constraints, and the resource constraint is that the total resource consumption of each task in the aggregated state does not exceed the total system resource.

[0113] First, the core objects and parameter definitions of resource allocation are defined to ensure that all variables are associated with the previous steps or system configuration: the resource types cover the core resources required for robot operation, including power resources (unit: mAh), computing power resources (unit: GFLOPS), and device occupation time resources (unit: s), which are denoted as resource categories (1 = power, computing power, device time). Among them, the total system resource is a preset configuration parameter (such as total power resource , total computing power resource , and total device time resource ) stored in the system resource configuration module; the resource consumption of the task in the aggregated state is calculated based on the aggregated state data of S2 association modeling layer—by extracting the resource demand characteristics of the core state corresponding to task i from the state core tensor of S2, and further correcting by combining the resource demand proportion of the task in the task priority calculation step of S4, wherein is the task ​Basic resource consumption (from task initialization instructions, such as the basic power consumption of assembly tasks) ), 0.2 is the resource consumption correction coefficient (adapting to the fluctuation of resource demand under the aggregation state).

[0114] Resource allocation must satisfy the constraint that "the total resource consumption of all tasks in the aggregated state does not exceed the total system resources". Resource allocation decision variables ( Represented as a task Allocate resources, (Indicates no allocation), then for each type of resource The mathematical expression for the constraint is:

[0115]

[0116] In the formula, This represents the total number of tasks awaiting resource allocation (derived from the list of tasks to be executed after priority sorting by S4); simultaneously, to prioritize high-priority tasks, resource allocation must be combined with the task priorities calculated by S4. The "priority sorting-greedy allocation" strategy is adopted: first, the tasks to be executed are sorted according to... Sort the tasks from largest to smallest, and allocate resources accordingly, up to a specific resource category. Remaining available amount ( Stop allocating resources for this type of task (for tasks currently pending assignment) to ensure that high-priority tasks receive resource support first.

[0117] The method in this step specifically includes: it also includes aggregating the local resource parameters of multiple robots using a federated learning framework, wherein during the aggregation process, parameters exceeding the local parameter mean are first removed. The extreme parameters are calculated by multiplying the standard deviation by a factor of 1, and then the remaining parameters are weighted by the robot's task completion rate. The local resource parameter here is the predicted value of each robot for the "task-resource matching degree". ( (where N is the total number of robots). Generated by the robot's local computing unit (S1's operational status data + S4's task resource requirements), it is used to characterize the robot. Resource utilization efficiency when performing a certain type of task (e.g.) The larger the robot The less resources are wasted in executing the task.

[0118] First, extreme parameter elimination is performed: Let... The set of local parameters for each robot is Calculate the mean of the set. with standard deviation :

[0119] wherein, is the local parameter mean, is the standard deviation; extreme parameters satisfying are removed (such parameters are mostly caused by robot sensor abnormalities or data noise, such as resource efficiency misjudgment caused by motor current sensor failure in S1), and the remaining parameter set is obtained.

[0120] Subsequently, weighted average aggregation is performed, and the weight is the task completion rate of each robot (real-time statistics from the execution steps of S6 flow, , and after weight normalization:

[0121]

[0122] wherein, ; the final global resource matching degree parameter after aggregation is:

[0123]

[0124] This parameter is used to correct the resource allocation matching relationship between tasks and robots, such as Higher robots are preferentially allocated high resource demand tasks, improving overall resource utilization efficiency.

[0125] Before uploading the local parameter, Laplace noise is added to the parameter, and the noise scale is determined by the maximum variation range of the parameter, the privacy protection level, and the number of robots. Let the local parameter be the maximum variation range (calculated by traversing all local parameters of robots), the privacy protection level (industrial scenarios commonly use , meeting the differential privacy safety standard), and the number of robots , then the scale parameter of the Laplace noise is:

[0126]

[0127] wherein, determines the noise intensity, is smaller (the higher the privacy protection requirement), is smaller (the fewer the number of robots), is larger, and the noise is stronger. After adding noise to the local parameter , the privacy protection parameter is obtained:

[0128]

[0129] wherein, This indicates a mean of 0 and a scale parameter of . Laplace distributed random variable; parameters after adding noise Uploading to the federated learning server avoids the leakage of raw local parameters (including sensitive information such as robot resource consumption details and task execution preferences) while ensuring the availability of the aggregated global parameters, complying with industrial data security standards. Local parameters are defined as those whose differences from the global parameters exceed a threshold, and the difference threshold is set to the mean of the global parameters. That is, the absolute deviation between the local parameter and the global parameter mean. When this happens, the upload of this local parameter is triggered.

[0130] Through the above steps, the resource allocation process satisfies system resource constraints, achieves robust aggregation of heterogeneous resource parameters among multiple robots through federated learning, and also protects privacy. The final output is a task-robot resource allocation scheme, such as "Task 1 (priority 0.92) is allocated to robot A (remaining battery power)." Ranked #1), it consumes 2800mAh of power, has a computing power of 12GFLOPS, and a device duration of 180s, providing clear resource scheduling instructions for the S6 process execution steps.

[0131] S6: Process execution steps, execute the task process according to the task priority and resource allocation results.

[0132] In this step, based on the task priorities and resource allocation results determined in the previous steps, the corresponding task flow is executed to complete the programming and implementation of complex robot tasks. First, the task priority sequence in S4 and the resource allocation scheme in S5 need to be converted into specific instructions executable by the robot, forming a four-dimensional execution scheme of "task-robot-resource-timing": Let the task sequence after S4 be... ( (Highest priority), S5 for each task The assigned robots are (like Resource quotas allocated to robot A are (corresponding to power consumption, computing power, and device duration respectively), then it needs to be... Generate execution instructions containing action sequences and resource usage timings—action sequences (e.g., "move to material area") Grab materials Move to assembly area Assembly) is based on task type presets, and resource usage timing is controlled by "resource consumption rate". The estimated execution time is (From task initialization instructions, such as) Then the power consumption rate (unit: ), ensuring even distribution of resource consumption and avoiding resource depletion within a short period of time.

[0133] During task execution, real-time monitoring of execution progress and resource consumption status is required to avoid deviation from the expected plan. Execution progress is quantified by "sub-step completion rate". Assuming that a task contains sub-steps (e.g., an assembly task contains 4 sub-steps: positioning, grasping, aligning, and fastening), the number of completed sub-steps is detected in real time through S1's action feedback sensors (e.g., joint angle sensors, grasping force sensors) , and the execution progress calculation formula is:

[0134]

[0135] In the formula, , when , it indicates that the task is completed; at the same time, S1's resource sensors collect the actual resource consumption of the robot at the current execution time , and calculate the resource consumption progress:

[0136]

[0137] In the formula, is the resource category, ; if (the preset resource overspending threshold ), a resource warning is triggered, the task execution is suspended and feedback is given to S5 resource allocation step, and the resource quota is adjusted again (e.g., redundant resources are adjusted from other low-priority tasks).

[0138] To cope with execution deviations (e.g., robot positioning deviation, material grasping failure), a deviation detection and correction mechanism needs to be established. Taking positioning deviation as an example, assuming that the target position of a certain sub-step of a task is (from the task instruction, such as ), and the real-time position of the robot is (from S1's GPS / SLAM sensor), then the positioning deviation is:

[0139]

[0140] In the formula, is the planar positioning deviation (unit: m); if (the preset deviation threshold, such as ), deviation correction is triggered: based on S2's state-event association matrix ​​Query the correlation between the "positioning deviation" event and the "position calibration" status. ,like (For high correlation) the robot's built-in calibration algorithm (such as point cloud matching calibration based on LiDAR) is invoked to correct the position until... If the correlation is low or the calibration fails, the alternative process generated by the S8 pre-planning step is triggered (such as replacing the robot with one that has higher positioning accuracy to perform the sub-step).

[0141] For tasks requiring multi-robot collaborative execution (such as...) Robot A needs to pick up materials, and robot B needs to transport materials. Robot task dependency needs to be based on S3. (Dependency between robots A and B) Achieving time synchronization: Assume the estimated time for robot A to complete the grasping sub-step is... Then, a command to "delay the start of the transportation sub-step" is issued to robot B, with the start time set to [time value missing]. ( (For synchronization buffer time), ensure that robot B can promptly connect to the transportation after robot A completes the grasping, and avoid task disconnection.

[0142] After the task is completed ( The execution results data need to be collected and fed back to the relevant steps: First, the task completion status (success / failure) and actual execution time need to be recorded. Actual resource consumption Feedback is sent to the S4 task priority calculation step to adjust the time sensitivity coefficient of subsequent tasks. (If the actual duration exceeds the estimated duration, then the duration of similar tasks will be increased.) (weighting); secondly, to factor in resource consumption deviations. Feedback is sent to the S5 resource allocation step to optimize resource quota calculation (e.g.) If the value remains positive, it will increase the number of similar tasks. Basic resource quota Third, the execution progress curve and deviation correction records are fed back to the S10 visualization feedback step to dynamically update the BPMN flowchart and realize the transparency of the execution status of the entire chain.

[0143] S7: Cross-scene adaptation steps.

[0144] The cross-scene adaptation step includes: extracting visual image features, laser point cloud features and task text features of a new scene; taking the task text features as an anchor point, calculating the similarity of the visual image features, the laser point cloud features and the task text features, and normalizing the similarity to obtain attention weights of each modality feature, and aligning the multi-modal features based on the attention weights to generate a scene implicit representation; calculating the similarity of the scene implicit representation and a historical scene implicit representation, and migrating associated model parameters of the historical scene based on the similarity.

[0145] In the method of the present step, the cross-scene adaptation operation is performed according to the above steps to realize the migration application of the associated model in different scenes. First, the extraction method and parameter definition of each modality feature are determined to ensure that the feature source is traceable and meets the scene adaptation requirements: the visual image features The images are collected by deploying a high-definition industrial camera in the new scene, extracted by a pre-trained ResNet50 model (weights from ImageNet pre-training library, fine-tuned for industrial scene), and the output dimension is , that is , each element corresponds to the local texture, target contour and other visual features of the image; the laser point cloud features The point cloud data of the new scene is collected by a 3D laser scanner (the number of point clouds is set to 1024 per frame), processed by a PointNet model (pre-trained for robot scene obstacle and terrain recognition), and the output dimension is , that is , representing the spatial structure, distance distribution and other features of the point cloud; the task text features The task instruction text of the new scene (such as "outdoor logistics sorting" and "indoor precision assembly") is generated, and a BERT-base pre-training model (fine-tuned for industrial task semantics) is used, and the output dimension is , that is , capturing the semantic connotation of the text (such as the task type difference between "sorting" and "assembly").

[0146] When calculating the similarity of each modality feature with the task text features as an anchor point, the cosine similarity is used to quantify the semantic correlation between features to avoid similarity deviation caused by feature dimension difference. The similarity calculation formula of the visual image features and the text features is:

[0147]

[0148] In the formula, is the text feature The linear mapping layer (weight matrix is pre-trained on a cross-modal dataset, Pre-trained on public industrial robot cross-scene multi-modal dataset (e.g. RobotScene-100) which contains 50+ industrial scene visual, laser, text modal data, ensuring that the mapping layer adapts to the industrial scene semantics) dimensionality reduction results, ensuring consistency with dimensionality; representing vector dot product, representing L2 norm. Similarly, the similarity between laser point cloud features and text features is:

[0149]

[0150] where is the dimensionality reduction result of the text feature through another linear mapping layer (Ltext) , and the similarity of the text feature itself is set to (reference value). The above similarity is normalized by Softmax to obtain the attention weight of each modal feature, ensuring that the sum of the weights is 1, and the formula is:

[0151]

[0152]

[0153]

[0154] where are the attention weights of the visual, laser point cloud, and text modal, respectively, satisfying ; the stronger the association between a certain modal and the text semantics (e.g. "indoor assembly" text and high sim value of indoor image), the greater the corresponding weight, ensuring that the aligned features focus on consistent information.

[0155] When aligning multi-modal features based on attention weights to generate scene implicit representation, first map each modal feature to the same dimension (avoid dimension difference affecting fusion effect) through a unified linear layer, obtaining , and then generate implicit representation through weighted summation, the formula is:

[0156]

[0157] This representation condenses the multi-modal implicit association of the new scene (e.g. in the "outdoor sorting" scene, the "logistics box" feature of the image, the "open terrain" feature of the point cloud, and the "sorting" semantic feature of the text are fused), providing a core basis for cross-scene transfer.

[0158] When calculating the similarity between the scene implicit representation and the historical scene implicit representation, the historical scene implicit representation set ​The number of historical scenes comes from the historical scene feature library stored in the system, each... For the first The implicit representations generated by the above steps for each historical scene. The new scene and the... The similarity between the historical scenes is calculated using cosine similarity:

[0159]

[0160] In the formula, The larger the value, the more similar the new scene is to the first one. The stronger the implicit connection between historical scenarios (such as "outdoor logistics sorting" and "outdoor warehousing and handling"), the better. Close to 1). Normalization is performed to obtain the transfer weights:

[0161]

[0162] In the formula, For the first A historical scene Matrix (from historical modeling results of S2); similarly, TT core tensor , This migration method eliminates the need to build a correlation model from scratch for new scenarios; it only requires a small amount of new scenario data. By fine-tuning the migration parameters (such as dynamic rank adjustment of S2) for each sample, the system can be quickly adapted, significantly shortening the deployment cycle.

[0163] S8: Pre-planning steps.

[0164] The pre-planning steps include: extracting historical information gain sequences, state-event correlation trends, and task complexity variation characteristics; constructing a time-series prediction model; separating the long-term trend characteristics and periodic characteristics of information gain; focusing on key historical time steps through an attention mechanism to predict the information gain of future time steps; collecting robot motor current, temperature, and sensor signal strength characteristics; assigning high weights to key fault features through an attention mechanism; using an encoder-decoder model to calculate the error between input features and reconstructed features, and the deviation between encoded latent features and normal sample latent features; and weighting the error and deviation to map them to the robot fault probability; when the predicted information gain is lower than a set threshold or the fault probability is higher than a set threshold, generating an alternative process containing a transition task, wherein the similarity of the state features of the transition task with the currently executed task and the target task is not lower than a set threshold.

[0165] In this step, the pre-planning operation is performed according to the above steps to predict risks in advance and generate alternative processes, ensuring the stability of task execution.

[0166] When extracting historical information gain sequences, information gain (IG) is defined as the increment of information in the state-event correlation matrix after a new event is triggered, based on the S2 state-event correlation matrix. Calculation: Let Let be the matrix at time t. If the event is triggered at time t, then ,in For matrix entropy, (Normalized), representing the uncertainty of the matrix; the historical information gain sequence is Historical time steps (taking data from the past 24 hours, with one time step per hour, i.e.) The trend of changes in the correlation between state and event is observed through... Calculation of the mean difference of a matrix in a sliding window: The window size; the task complexity variation characteristics are calculated based on the task priority sequence of S4: Let be the number of tasks at time t. For the task Priority.

[0167] When constructing the time series prediction model, an LSTM-attention hybrid model is adopted: firstly, the long-term trend and periodic characteristics of information gain are separated through linear decomposition, and the long-term trend... To fit the parameters, the least squares method is used. Solve for periodic characteristics The main period is obtained by Fourier transform. (like (Hours, corresponding to work shift cycles). The trend and cycle characteristics are concatenated into an input sequence. Input to LSTM layer to extract latent features Then, focus on key time steps through the attention mechanism: attention weights Depend on With query vector The similarity is determined after (random initialization and training optimization). The key step is Take the front (Time step of weights). Future step( The predicted information gain value (i.e., for the next 4 hours) is:

[0168]

[0169] When collecting robot motor current, temperature, and sensor signal strength characteristics, the feature vector is: ,in Real-time motor current (unit: , current sensor from , sampling frequency 10 Hz), , motor temperature (unit: , temperature sensor from , , sensor signal strength (normalized to , communication module from S1). Assign high weight to key failure features through attention mechanism: calculate mutual information between each feature and historical failure label , Fault), , ( , failure label, 1 represents failure), normalize to get attention weight Similarly, the higher the weight, the more critical the feature to failure prediction (e.g. Motor current

[0170] Calculate error and bias using encoding-decoding model (autoencoder): encoder maps weighted features to hidden features (dimension 20), decoder reconstructs features ; error between input and reconstructed features is . The set of hidden features of normal samples is , from historical non-failure samples), the mean is , the bias of the encoded hidden features from the normal mean is . The weighted error and bias are mapped to the failure probability: , where is the weight, sigmoid function to ensure .

[0171] When the predicted information gain is lower than the set threshold or the failure probability is higher than the set threshold, trigger the alternative process generation: information gain threshold (take of the historical maximum value, avoid insufficient decision basis due to too low information gain), failure probability threshold (based on system reliability requirements, ensure that the false alarm rate is less than 5% with 95% confidence). Transition tasks are selected from the historical task library, with state features ( , core states such as power, position, load), the state features of the current executing task are , the state features of the target task are , and the similarity is calculated using cosine similarity: , Similarly, set the threshold to 0.8, and only keep two transition task (e.g., when the current task is "material grabbing" and the target task is "material assembly", the transition task can be "material carrying", which has a state feature highly similar to both). The generated alternative process is "current task transition task target task", and stores the resource requirements, execution duration, etc. of the transition task (from historical task records) to ensure that the original process can be quickly replaced when triggered.

[0172] S9: visual feedback step.

[0173] Map the task nodes, event nodes, and resource nodes into a dynamic flowchart, where the task nodes are labeled with the calculated task priority, the proportion of task resource requirements to the total system resources, and the associated impact of the task's corresponding state and effective events; the event nodes are labeled with the impact of events on the state; the proportion of task delay time to the total task duration is used as the increment of the dynamic rank adjustment error threshold in the association modeling step, and the proportion of task delay time to the deadline is used as the deduction item of the reward function of the intelligent optimization algorithm in the threshold adjustment step, to realize the feedback optimization of the execution results on the model parameters.

[0174] In the method of this step, the visual feedback operation is performed according to the above steps to realize the visual monitoring of the task process and the optimization feedback of the execution results on the model parameters. First, the mapping rules of each node and the calculation logic of the labeled parameters are determined to ensure that the visual information is directly related to the output of the previous steps: the dynamic flowchart is constructed according to the BPMN (Business Process Model and Notation) standard, the task nodes are represented by rectangular boxes, the event nodes are represented by circular boxes, and the resource nodes are represented by diamond boxes. The order of the process is represented by arrows connecting the nodes (the arrow direction represents the task execution flow).

[0175] Among the parameters labeled in the task nodes, the task priority directly references the results calculated in S4 , to intuitively distinguish the priority levels, the node fill color changes with red, yellow, green); the proportion of task resource requirements to the total system resources is calculated for each type of resource (electricity, computing power, device duration), and the formula is , where is the resource consumption quota of the task in S5, ​For the total resources of the system (preset configuration), the label is presented in the format of "power ratio: XX%, computing power ratio: XX%, time length ratio: XX%"; the total sum of the associated influence of the task corresponding state and the effective event is multiplied, and the calculation logic in S4 is cited , and the formula is , wherein (S2 state-event association matrix), (S2 event-state influence matrix), (S3 filtered effective event) are all from the previous steps, and the value is labeled below the task node, which is used to reflect the close degree of association between the task and the current system event.

[0176] The "event influence on state" labeled in the event node is the maximum influence value of the event on all states in the S2 event-state influence matrix , and the formula is , wherein is the influence amplitude of the event on the state , and the influence degree is distinguished by color during labeling ( orange, blue, gray), which facilitates quick identification of high-impact events (such as usually ).

[0177] The resource node labels the current resource remaining rate, and the formula is , wherein is the total resource allocated to the task in S5, and the remaining rate is lower than , the node border flashes as a warning, prompting the need for timely scheduling due to resource shortage.

[0178] In the execution result feedback optimization part, first define the task delay time related parameters: the actual execution time of the task comes from the execution progress statistics of S6 (the time recorded when ), the expected execution time comes from the task initialization instruction (such as the assembly task ), and the task delay time (delay time is non-negative, and is 0 when there is no delay); the total task duration is the actual execution time , and the task deadline comes from the time sensitivity coefficient calculation in S4 (preset by the task initialization instruction).

[0179] The proportion of task delay time to total task duration is taken as the increment of the dynamic rank adjustment error threshold in the association modeling step: let the basic error threshold of dynamic rank adjustment in S2 be (Pre-set to 0.05, i.e. reconstruction error is acceptable), delay ratio , then the error threshold increment , where is the increment coefficient (determined based on system debugging, to ensure that the impact of delay on the error threshold is controllable); the adjusted error threshold is This value is fed back to the dynamic rank adjustment link of S2 - if the task delay is caused by the low rank parameter leading to the loss of associated information (such as low rank leading to state judgment deviation), then increasing the error threshold can allow higher reconstruction error, indirectly promoting the rank parameter to be raised (such as originally when the rank is 8, the rank can be raised to 9 after adjustment ), reducing the delay caused by insufficient information.

[0180] The ratio of task delay time to deadline is taken as the deduction item of the reward function of the intelligent optimization algorithm in the threshold adjustment step: the original reward function of the MO-DDPG algorithm in S3 is , and the delay ratio is defined as the start time of the task, from the execution record of S6), then the reward function deduction item , where is the deduction coefficient (to balance the impact of delay on the reward, to avoid excessive punishment); the adjusted reward function is This function is fed back to the intelligent optimization algorithm of S3 - if the task delay is caused by unreasonable threshold coefficient (such as too high filtering threshold to exclude key events), then the deduction of reward will promote the algorithm to adjust (clustering threshold adjustment coefficient), (filtering threshold adjustment coefficient), optimize the threshold generation logic (such as reducing the filtering threshold to retain more key events), reduce the subsequent task delay.

[0181] After each task execution is completed, the above feedback optimization process is automatically triggered, and the adjusted and are updated to the parameter library of S2 and S3, respectively, for the associated modeling and threshold adjustment of the next task, forming a closed loop of "execution-feedback-optimization", continuously improving the adaptability of the system to task delay, and reducing the repeated delay rate of similar tasks (tests have shown that this feedback mechanism can reduce the task delay rate from the initial to below).

[0182] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all described as a combination of a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0183] The present scheme constructs a complete logical chain from data input to closed-loop optimization through "full-link technical means collaborative linkage". The design characteristics of each link technical means can objectively deduce the corresponding technical effect, and the specific deduction process is as follows:

[0184] Firstly, in the data acquisition link, the technical means of "multi-dimensional sensor (robot body sensor + environmental perception equipment + multi-modal acquisition equipment) + sliding average filtering pretreatment" is adopted: the former ensures the full-dimensional data input covering the robot running state (position, power, temperature), environmental events (obstacles, abnormalities) and scene multi-modal characteristics (vision, laser, text), avoiding modeling deviation caused by missing data dimensions; the latter smooths the noise data (such as position sensor Gaussian noise, power sensor fluctuation) through sliding window, ensuring the accuracy of the original data - this means directly deduces the effect of "providing high-quality, full-dimensional data foundation for subsequent correlation modeling", eliminating the interference of noise and data missing on modeling accuracy.

[0185] Secondly, the correlation modeling link adopts the progressive technical means of "double matrix quantization (state-event correlation matrix + event-state influence matrix) → three-dimensional tensor fusion → TT low-rank decomposition + task perception dynamic rank adjustment": the double matrix quantizes the state-event relationship from "co-occurrence correlation" and "dynamic influence", solving the problem that traditional single matrix cannot fully characterize the interaction characteristics; tensor fusion integrates the double matrix with time dimension, preserving the spatiotemporal dynamic correlation and avoiding the consistency conflict of independent matrix; TT decomposition reduces the high-dimensional computational complexity from three orders to linear, solving the problem of computational delay in large-scale scenarios; dynamic rank adjustment adapts the rank parameter according to task priority (high-priority task increases rank to preserve accuracy, low-priority task reduces rank to improve efficiency), avoiding the precision waste or insufficient efficiency caused by "one-size-fits-all" rank setting - this series of means cooperatively deduces the effect of "considering state-event dynamic correlation characterization, computational real-time and task adaptability", meeting the modeling needs of different priority tasks in complex scenarios.

[0186] Furthermore, the threshold adjustment link adopts the technical means of "MO-DDPG multi-objective optimization threshold coefficient + graph attention multi-machine collaborative conflict correction": MO-DDPG constructs a weighted function with task completion rate, resource utilization rate, and information loss rate as the target to iteratively optimize the clustering / filtering threshold, replacing the experience dependence of traditional manual threshold setting and avoiding local optimization caused by single-objective optimization; the graph attention mechanism calculates the multi-robot task dependency through node feature similarity, dynamically adjusts the conflict correction factor (the higher the dependency, the gentler the correction), and avoids task coordination rupture caused by multi-machine threshold conflict - this means derives the effect of "threshold coefficient autonomous optimization and multi-machine collaborative conflict-free", balancing multi-objective demand while ensuring the threshold collaboration of heterogeneous robot clusters.

[0187] The task priority calculation link adopts a multi-factor weighting technical means of "time sensitivity coefficient + resource demand proportion + emergency level coefficient + state-event association impact product": the time sensitivity coefficient quantifies the deadline urgency, the resource demand proportion reflects the task's dependence on system resources, the emergency level coefficient reflects the inherent priority of the task, and the association impact product reflects the close association between the task and the current event - multi-factor comprehensive weighting avoids scheduling bias caused by a single dimension (such as only looking at time), and derives the effect of "accurate matching of task scheduling priority with system state and task characteristics", ensuring that high-value tasks (time-sensitive, resource-critical, urgent, and closely associated) have priority in obtaining scheduling resources.

[0188] The resource allocation link adopts the technical means of "federated learning robust aggregation (extreme parameter elimination + task completion rate weighting) + Laplace noise differential privacy": robust aggregation eliminates abnormal local parameters (such as extreme resource parameters caused by sensor failure), and assigns aggregation weights according to the robot task completion rate, avoiding aggregation bias caused by heterogeneous data (such as parameter differences between robots from different manufacturers); differential privacy protects sensitive parameters (such as process path and resource consumption details) by adding Laplace noise, in line with data security specifications - this means derives the effect of "accurate aggregation of heterogeneous robot resource parameters and privacy security of sensitive data", balancing resource allocation rationality and data security needs in industrial scenarios.

[0189] The process execution link adopts the technical means of "real-time progress / resource monitoring + deviation correction (positioning deviation calibration) + multi-machine timing synchronization (dependency-driven start time adaptation)": real-time monitoring ensures timely detection of execution deviations (such as resource overruns and positioning shifts), deviation correction calls calibration algorithms (such as laser point cloud matching) to correct deviations based on associated modeling results, and multi-machine synchronization adjusts start times based on task dependency (such as buffering time for transportation tasks lagging behind grasping tasks) - this means derives the effect of "timely correction of task execution process deviations and multi-machine collaboration without disconnection", avoiding task failure or coordination rupture caused by deviation accumulation.

[0190] The cross-scene adaptation link uses the technical means of "multimodal feature extraction (vision / laser / text) + text anchor attention alignment + historical parameter weighted transfer": multimodal features comprehensively depict the characteristics of the scene, text anchor alignment ensures that multimodal features are consistent around task semantics (such as "sorting" text aligned with visual logistics box features), and weighted transfer reuses historical associated model parameters based on scene similarity - this means derives the effect of "new scene without modeling from zero, quickly realizing associated model adaptation", solving the problems of long modeling cycle and high data dependence in traditional cross-scene modeling.

[0191] The pre-planning link uses the technical means of "temporal prediction information gain (trend + cycle separation + attention focus key time step) + fault probability calculation (key feature weighting + autoencoder error mapping) + alternative process triggering": information gain prediction identifies the risk of insufficient decision-making basis in advance, fault probability calculation focuses on key fault features such as motor current and temperature for accurate early warning, and alternative processes (including transition tasks) seamlessly connect when risks are triggered - this means derives the effect of "identifying insufficient information and fault risks in advance, and avoiding task interruption with alternative processes", improving the robustness of task execution.

[0192] The visual feedback link uses the technical means of "dynamic node labeling (task priority / resource occupancy / correlation + event impact + resource remaining rate) + execution result feedback optimization (delay ratio driven modeling / threshold parameter adjustment)": dynamic labeling enables intuitive monitoring of task, event, and resource status, and feedback optimization converts execution delays into associated modeling error threshold increments and threshold adjustment reward function deduction items - this means derives the effect of "transparent human-machine monitoring and closed-loop iterative optimization of system parameters", continuously improving the accuracy of modeling and adjustment in subsequent links.

[0193] In summary, the whole-link technical means achieve layer-by-layer connection and collaborative optimization through "data - modeling - adjustment - scheduling - execution - adaptation - pre-control - feedback", ultimately deriving the overall technical effect of "robot complex task arrangement with dynamic adaptability (adaptation to multiple scenes / multiple robots), scheduling rationality (multiple factor priority + precise resource allocation), execution stability (risk pre-control + deviation correction), privacy security (differential privacy), and sustainable optimization (closed-loop feedback)", fully solving the core problems of traditional static arrangement in complex dynamic scenarios, such as poor adaptation, weak collaboration, and uncontrollable risks.

[0194] The above description is merely exemplary of the application and of the application of the principles thereof and the application is not limited to the disclosed technical features or combinations thereof. It is intended to be apparent to one skilled in the art that the scope of the disclosure is not limited to the technical solutions formed by the specific combinations of the technical features disclosed above, and also includes other technical solutions formed by the combinations of the technical features disclosed above or their equivalent features without departing from the above disclosed concept. For example, technical solutions formed by the mutual replacement of the above features and technical features with similar functions disclosed in the application (but not limited to) are also included.

Claims

1. A method for dynamic orchestration of complex tasks by robots, characterized in that, The method comprises the following steps: a data collection step for collecting running state data and event data of the robot; an association modeling step for constructing a state-event association matrix and an event-state influence matrix, wherein the state-event association matrix quantifies the co-occurrence association strength of a state and an event by counting the number of simultaneous occurrences of a specific state and a specific event, the total number of occurrences of the state alone, and the total number of occurrences of the event alone, and using an intersection ratio correction method, and the event-state influence matrix quantifies the influence amplitude of an event on a state by calculating the change amplitude of a specific state after a specific event is triggered, and normalizing the change amplitude with the maximum change amplitude of all events on the state; a threshold adjustment step for dynamically generating a clustering threshold and a filtering threshold based on the state-event association matrix and the event-state influence matrix, wherein the clustering threshold is determined by superimposing a basic threshold on the maximum association weight of the event and all states, and the filtering threshold is determined by superimposing a minimum threshold on the sum of the influence degrees of all events on the state; a task priority calculation step for calculating the task priority in a weighted manner by combining the time sensitivity coefficient corresponding to the task deadline, the task resource demand proportion, the task urgency coefficient, and the product sum of the association degree of the state corresponding to the task and the effective event and the influence degree of the effective event on the state; a resource allocation step for allocating resources based on the task priority and resource constraints, wherein the resource constraints are that the total resource consumption of all tasks in the aggregated state does not exceed the total system resources; a process execution step for executing the task process according to the task priority and resource allocation result.

2. The method of claim 1, wherein, The association modeling step further comprises fusing the state-event association matrix and the event-state influence matrix into a state-event-time three-dimensional tensor, wherein any "state-event-time" element value in the three-dimensional tensor is calculated by taking the geometric mean of the association degree of the state and the event and the influence degree of the event on the state, to uniformly quantify the dynamic association relationship among the state, the event, and the time.

3. The method of claim 2, wherein, The association modeling step further comprises decomposing the three-dimensional tensor into the product of a state core tensor, an association core tensor, and a time core tensor using a low-rank decomposition algorithm, wherein the state core tensor represents the low-dimensional characteristics of the state dimension, the association core tensor represents the low-dimensional characteristics of the state and event interaction, and the time core tensor represents the low-dimensional characteristics of the time dimension, to reduce the calculation complexity of the three-dimensional tensor from the three-order complexity of the number of states, the number of events, and the number of time steps to linear complexity.

4. The method of claim 3, wherein, The association modeling step further comprises dynamically adjusting the rank parameter of the low-rank decomposition based on the task priority, increasing the rank parameter to increase the feature dimension of the core tensor for high-priority tasks to retain the detailed information of the state-event association, and reducing the rank parameter to reduce the feature dimension of the core tensor for low-priority tasks to improve the calculation efficiency within an acceptable information loss range.

5. The method of claim 1, wherein, The threshold adjusting step further comprises autonomously optimizing the threshold coefficient by an intelligent optimization algorithm, the intelligent optimization algorithm constructs a weighted summation objective function with task completion rate, resource utilization rate and information loss rate as targets, maximizes the objective function value by iteratively adjusting the threshold coefficient, and realizes multi-objective balance of task efficiency, resource utilization and information retention.

6. The method of claim 5, wherein, The threshold adjusting step further comprises constructing a multi-robot interaction relationship graph, taking each robot as a graph node, the graph node features include the current threshold coefficient of the robot, the task list and the fault risk; the similarity of the features of two robot nodes is calculated through a graph attention mechanism, and the task dependency of the two robots is obtained by normalizing the similarity; the threshold conflict correction strategy is dynamically adjusted based on the task dependency, and a smaller conflict correction factor is used for robots with high task dependency.

7. The method of claim 1, wherein, It further comprises a cross-scene adaptation step, which comprises: extracting visual image features, laser point cloud features and task text features of the new scene; taking the task text features as anchor points, calculating the similarity of the visual image features, laser point cloud features and task text features, and normalizing the similarity to obtain the attention weight of each modality feature, aligning the multi-modal features based on the attention weight to generate scene implicit representation; calculate the similarity of the scene implicit representation and the historical scene implicit representation, and transfer the associated model parameters of the historical scene based on the similarity.

8. The method of claim 1, wherein, It further comprises a pre-planning step, which comprises: extracting historical information gain sequence, state-event correlation degree change trend and task complexity change feature, constructing a time series prediction model, separating the long-term trend feature and periodic feature of information gain, and focusing on key historical time steps through attention mechanism to predict information gain at future time steps; collecting robot motor current, temperature and sensor signal intensity features, giving high weight to key fault features through attention mechanism, calculating the error between input features and reconstructed features, the deviation between encoding implicit features and normal sample implicit features using an encoding-decoding model, and mapping the weighted error and deviation to robot failure probability; when the predicted information gain is lower than the set threshold or the failure probability is higher than the set threshold, generate an alternative process containing transition tasks, the state features of the transition tasks, the current execution tasks and the target tasks have a similarity not lower than a set threshold.

9. The method of claim 1, wherein, The resource allocation step further comprises aggregating local resource parameters of multiple robots using a federated learning framework, first removing extreme parameters beyond ±2 standard deviations of the local parameter mean in the aggregation process, then performing weighted averaging on the remaining parameters with the robot task completion rate as the weight; before uploading the local parameters, add Laplacian noise to the parameters, the noise scale is determined by the maximum change range of the parameters, the privacy protection level and the number of robots.

10. The method of claim 1, wherein, Further comprising a visual feedback step, mapping the task nodes, event nodes and resource nodes into a dynamic flow chart, the task nodes are labeled with the task priority calculated based on claim 1, the proportion of task resource demand in total system resource, and the associated impact product of the corresponding state of the task and the effective event, the event nodes are labeled with the impact degree of the event on the state; the proportion of the task delay time in the total task time length is taken as the increment of the dynamic rank adjustment error threshold in the associated modeling step, and the proportion of the task delay time in the deadline time is taken as the deduction item of the intelligent optimization algorithm reward function in the threshold adjustment step, to realize the feedback optimization of the execution result on the model parameters.

Images