An ai-based industrial edge intelligence task collaboration method and system

Abstract

The application belongs to the technical field of industrial artificial intelligence, and discloses an AI-based industrial edge intelligent task cooperation method and system, which comprises the following steps: based on the probability distribution prediction of the future resource state, a main execution DAG and multiple or DAGs are automatically generated, instant switching is realized when a preset failure mode is detected, and the system robustness is greatly improved; a deep diagnosis is performed by using a causal inference engine, and the agent is guided to perform skill specialization and skill pruning, so that an intelligent ecology capable of self-purification and evolution is constructed; and a cross-topology meta-learning framework is adopted, so that the system can master the ability of quickly adapting to new and unseen tasks with few samples. Through the three innovations of multi-path or planning, skill life cycle management driven by causality and cross-topology meta-learning, the application brings the effects of nearly zero delay of task recovery, efficient evolution of collective intelligence and rapid generalization of new tasks, and enhances the application value of the system in complex and changeable industrial scenes such as flexible manufacturing.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial artificial intelligence, and particularly relates to an AI-based industrial edge intelligent task coordination method and system. BACKGROUND

[0002] Industrial edge intelligence aims to sink AI capabilities to the production site to meet the needs of low latency and high reliability. Existing technologies often decompose tasks into directed acyclic graphs (DAGs) and use multi-agent systems (MAS) and reinforcement learning (RL) for collaborative scheduling. However, in modern industrial environments represented by flexible manufacturing, task types are diverse, production pace is fast, and working environment is dynamic (such as network jitter and device load fluctuation), which poses unprecedented challenges to intelligent coordination systems. The bottlenecks of existing technologies are increasingly prominent. In the process of implementing the present application, the inventors found that at least the following problems exist in the prior art:

[0003] 1. Existing proactive or predictive scheduling methods usually only prevent the most likely failure. When other unforeseen or minor probability failures occur, the system still needs to spend a lot of time for re-planning and scheduling, resulting in task interruption and production efficiency decline, which cannot meet the stringent requirements of business continuity in industrial scenarios.

[0004] 2. In multi-agent collaboration, although there are knowledge sharing mechanisms such as federated learning, there is a lack of discrimination of knowledge effectiveness. A skill that is effective in one scenario may produce negative transfer in another scenario, thereby reducing system performance. Existing mechanisms cannot actively identify and manage such toxic knowledge, resulting in low efficiency or even wrong direction of collective intelligence evolution.

[0005] 3. Traditional scheduling strategies are usually optimized for a specific type of task (i.e., fixed DAG topology). When the production line switches to a completely new product or process, the corresponding task DAG topology changes fundamentally, and the original scheduling strategy often fails completely. The system needs to start from scratch and go through a long learning process to reach a usable level again, which cannot meet the rapid switching needs of flexible production. SUMMARY

[0006] The present application aims to at least partially solve the above technical problems. To this end, the present application aims to provide an AI-based industrial edge intelligent task coordination method and system.

[0007] The technical solution adopted by the present application is as follows:

[0008] An AI-based industrial edge intelligent task coordination method, comprising the following steps:

[0009] S1. Receiving an industrial collaborative task and decomposing it into an initial directed acyclic graph (DAG) composed of multiple subtask nodes;

[0010] S2. Identifying multiple potential failure modes based on probabilistic prediction of edge environment states;

[0011] S3. Generating a main execution DAG and at least one corresponding contingency DAG for each potential failure mode, the contingency DAG being an optimized backup task graph for coping with a specific failure mode;

[0012] S4. Assigning and instructing execution agents on edge devices to collaboratively execute subtasks according to the main execution DAG using a collaborative scheduler;

[0013] S5. Monitoring the edge environment in real time during execution, and when an actual failure mode matching one of the potential failure modes is detected, immediately suspending execution of the main execution DAG and switching to the corresponding contingency DAG to continue task execution;

[0014] S6. After task execution, analyzing execution data using a causal inference-based diagnosis engine, and when the root cause of failure is located to be the insufficient capability of an execution agent in a specific context, triggering the agent to distill a general model into an expert skill model;

[0015] S7. When the causal inference-based diagnosis engine finds that a shared skill produces negative transfer to certain agents, triggering skill pruning of the skill library of these agents;

[0016] S8. Using a meta-learning framework to train on multiple DAG tasks with different topologies to learn a scheduling meta-strategy that can quickly adapt to new topologies with few samples.

[0017] Preferably, in S5, the switching to the corresponding contingency DAG includes starting execution from a node in the contingency DAG corresponding to the execution progress when the main execution DAG is suspended, and inheriting the results of subtasks completed by the main execution DAG.

[0018] Preferably, the step of distilling a general model into an expert skill model in S6 includes collecting context data in the specific context as supervision information, and using the data to distill the knowledge related to the specific context in the general model into a lightweight expert skill model.

[0019] Preferably, in S7, the skill pruning includes prohibiting a specific agent from loading or calling a shared skill model diagnosed to produce negative transfer.

[0020] Preferably, in S8, the meta-learning framework optimizes the scheduling meta-policy on a large number of historical DAG tasks, so that it can converge to an efficient scheduling strategy with only a small number of scheduling instances when facing a new, never-before-seen DAG topology.

[0021] Preferably, the method further comprises: when the causal inference diagnosis engine attributes the failure root cause to the inaccurate prediction in S2, the execution data leading to the failure is used as a negative sample to optimize the model relied on by the probabilistic prediction.

[0022] Preferably, in S2, the probabilistic prediction of the edge environment state is realized by a time series graph neural network.

[0023] Preferably, the industrial collaborative task is multi-process quality detection on a flexible production line or predictive maintenance of a device group.

[0024] The application also provides an AI-based industrial edge intelligent task collaboration system, comprising:

[0025] A task decomposition module for performing step S1 in the above method;

[0026] An orality planning module for performing steps S2 and S3 in the above method;

[0027] A collaborative execution and switching module for performing steps S4 and S5 in the above method;

[0028] A causally driven skill management module for performing steps S6 and S7 in the above method;

[0029] A cross-topology meta-learning module for performing step S8 in the above method.

[0030] Preferably, the causally driven skill management module is further configured to maintain a skill profile recording the skill source, applicable context, performance, and negative transfer record of each agent, to support the decision of skill pruning.

[0031] The application has the following beneficial effects:

[0032] The application pre-generates and caches orality DAGs for a plurality of possible faults, realizes zero-delay plan switching when a fault occurs, and enables the system to have unprecedented fault recovery speed and business continuity guarantee capability, so that it can maintain a very high task success rate in a highly unstable real industrial environment.

[0033] Furthermore, the system can not only learn but also proactively differentiate into experts and eliminate harmful knowledge. This adaptive intelligent ecological evolution mechanism greatly enhances the overall performance and stability of collective intelligence, especially when dealing with complex and heterogeneous problems. Attached Figure Description

[0034] Figure 1 This is a flowchart of the AI-based industrial edge intelligent task collaboration method of the present invention.

[0035] Figure 2 This is a functional block diagram of the AI-based industrial edge intelligent task collaboration system of the present invention. Detailed Implementation

[0036] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] It should be understood that, and also noted, in the embodiments, the functions / actions may appear in a different order than those shown in the figures. For example, depending on the functions / actions involved, they may actually be performed substantially concurrently, or sometimes the two figures shown consecutively may be performed in reverse order.

[0038] like Figure 1 As shown, this embodiment uses the visual inspection task of various mobile phone mid-frames on a flexible production line as an example to illustrate the implementation process of the method of the present invention, specifically including:

[0039] S1. Receiving and Decomposing Tasks:

[0040] The system receives the task "Detect the mid-frame of a Type A mobile phone", and the task decomposition module decomposes it into an initial DAG.

[0041] S2-S3. Active multi-path contingency planning:

[0042] Core Implementation: The probabilistic planning module is activated. Its built-in Temporal Graphical Neural Network (T-GNN) analyzes the historical state of each node in the edge network and outputs a probability distribution of the resource state within the next minute. The analysis shows: there is a 70% probability that everything is normal; there is a 20% probability that the GPU of Node-05 (responsible for high-precision size measurement) will be preempted, resulting in insufficient computing power; and there is a 10% probability that the network link from Node-02 to Node-05 will experience severe congestion.

[0043] Generate a multi-path DAG: Based on this prediction, the system generates:

[0044] Master execution DAG: Optimal scheduling under normal conditions at 70%.

[0045] Contingency DAG-1 (Addressing insufficient computing power): Replace the high-precision measurement task on Node-05 with an algorithm that has slightly lower precision but takes less time, or migrate it to the backup node Node-09.

[0046] Contingency DAG-2 (Addressing network congestion): Add a data compression subtask on Node-02 to reduce the amount of data transmitted.

[0047] All three DAGs are cached.

[0048] S4. Collaborative execution:

[0049] Based on the master execution DAG, the collaborative execution and switching module assigns each subtask node in the DAG to the corresponding edge agent (such as an edge server with a camera, a robotic arm controller, etc.) on the production line through its internal collaborative scheduler, and issues execution instructions. The task begins to be executed collaboratively according to the path of the master DAG.

[0050] S5. Monitoring and immediate switching:

[0051] During execution, the system's state monitoring unit continuously receives heartbeat packets and key performance indicators from each edge agent, monitoring the edge environment in real time.

[0052] When the task is halfway through, the monitoring unit detects that the GPU occupancy rate of Node-05 has soared to 99% in a short period of time, which matches the actual state of the pre-set insufficient computing power failure mode.

[0053] The collaborative execution and switching module immediately triggers the switching mechanism: it issues a stop instruction to the relevant agents executing the master DAG, and seamlessly switches to the cached contingency DAG-1. The system inherits all completed subtask results and directly continues execution from the node in contingency DAG-1 that matches the interrupted state (e.g., the backup measurement task on Node-09). The entire switching process takes milliseconds, avoiding the huge overhead of starting from scratch.

[0054] S6-S7. Self-optimizing closed loop (causally guided skill management and system self-learning):

[0055] Core implementation: After the task is completed, whether successfully or not, the causally driven skill management module will be activated for in-depth analysis of the execution data.

[0056] Skill Specialization: Suppose a task fails once, and the quality of the inspection report is not up to standard. After analysis by the causal engine, it is found that the root cause is that the agent on Node-03 performs poorly in handling a specific situation of high-gloss curved surface reflection using its general defect detection model. The system automatically triggers skill specialization: collect all failed samples (high-gloss curved surface images), force the agent on Node-03 to distill the layers responsible for feature extraction in the general model into a specialized expert skill model that specifically suppresses high-gloss and enhances edges, and label it with the context: high glare surface.

[0057] Skill Pruning: Suppose the above high-gloss suppression expert skill is shared, and the agent on Node-08 (responsible for detecting flat scratches) calls it, which actually causes its recognition rate of fine scratches to decrease. After several iterations, the causal engine finds that this skill has a significant negative causal effect on the task of Node-08. The system then triggers skill pruning, updates the skill profile of Node-08, and prohibits it from calling this skill again, thus achieving the purification of the collective knowledge base.

[0058] Predictive Model Self-Optimization: Suppose a task interruption is due to an unexpected network congestion event. After the fact, the causal engine attributes the failure to this. It further analyzes and finds that the root cause of the interruption is not execution failure, but that the T-GNN model in S2 failed to successfully predict this congestion pattern. At this time, the system will feed this case of prediction failure (including the sequence of environmental states before the event and the final congestion result) as a new negative sample back to the T-GNN model for incremental training or fine-tuning, thereby improving its future prediction accuracy and coverage, forming a complete self-optimization closed loop.

[0059] S8. Cross-topology meta-learning:

[0060] Core Implementation: The flexible production line of this factory may need to switch between producing three different models of mobile phones, A, B, and C, within a day, and each detection task corresponds to a completely different DAG topology.

[0061] Meta-strategy learning: The cross-topology meta-learning module works continuously in the background. It collects all (DAG topology, successful scheduling trajectory) data pairs of A, B, and C type tasks in the past. Using the model-agnostic meta-learning (MAML) algorithm, it does not learn a specific scheduler for A or B or C, but learns a scheduling meta-strategy. This meta-strategy contains meta-knowledge on how to adjust scheduling parameters based on topological features such as the number of nodes, edge density, and critical path length.

[0062] Few-shot fast generalization: When the second day production line receives a brand new D type mobile phone detection task, the system is faced with a never-before-seen DAG topology. At this time, the scheduler loaded with meta-strategy no longer needs to start from scratch and explore randomly. It uses the good initial parameters provided by the meta-strategy and only needs a few (e.g., less than 10) online scheduling attempts to quickly adapt to the topological characteristics of the D type task and converge to a near-optimal scheduling strategy. This is several orders of magnitude more efficient than traditional RL, which requires thousands of attempts to learn.

[0063] For a clearer understanding of the system implementation of the present application, please refer to Figure 2 . The system is a physical or logical carrier for implementing the above method, which can be deployed on a cloud server, an edge server, or a combination thereof, and communicates with the on-site edge device (executing the agent) through the network. The system mainly includes: a task decomposition module, a probabilistic planning module, a collaborative execution and switching module, a causal-driven skill management module, and a cross-topology meta-learning module.

[0064] As shown in Figure 2 , the modules of the system of the present application are described in detail:

[0065] The task decomposition module is the entrance of the system. It is responsible for receiving macro-industrial tasks from the upper MES (manufacturing execution system) or user input, such as "complete the overall quality inspection of P008 batch workpieces". The module can be a rule-based parsing engine or a trained planning large model. It parses and decomposes the macro task into a series of sub-tasks with dependencies, and outputs them in the format of an initial directed acyclic graph (DAG). For example, the overall quality inspection is decomposed into loading, surface cleanliness detection, dimension measurement, crack detection, unloading, etc. nodes, and defines the execution order between them.

[0066] The probabilistic planning module is the core of the system to achieve proactive robustness. It receives the initial DAG output by the task decomposition module. Its core function is to generate a complete and immediately executable multi-path planning scheme based on the probabilistic prediction of the future state of the environment.

[0067] The module is embedded with a time series prediction model and a DAG generator. It continuously interacts with the environment knowledge graph to obtain historical and real-time state data such as CPU / GPU load, network bandwidth, device temperature, etc. of edge nodes (such as cameras, robots, edge servers). The time series prediction model outputs multiple failure modes and their probabilities that may occur in the future (e.g., in the next 5 minutes) based on these data, such as: node A insufficient computing power probability 20%, link B congestion probability 10%.

[0068] For each high-probability failure mode, the DAG generator optimizes the initial DAG to generate a corresponding contingency DAG. Meanwhile, it also generates a main execution DAG based on the most likely scenario. Finally, the module packages this set of DAGs (one main DAG and multiple contingency DAGs) and sends them to the collaborative execution and switching module.

[0069] The collaborative execution and switching module is responsible for assigning sub-tasks in the DAG to specific edge agents for execution, monitoring the execution process in real time, and performing contingency switching when necessary. The module includes a collaborative scheduler and a state monitoring and switching unit. The scheduler efficiently assigns sub-task nodes of the main execution DAG to edge agents based on the meta-strategy provided by the cross-topology meta-learning module. The state monitoring and switching unit continuously receives real-time state reports from edge agents and matches them with the failure modes preset by the contingency planning module.

[0070] When a failure mode (such as insufficient computing power of node A) is detected, the unit immediately issues a suspension instruction to all relevant agents and seamlessly switches to the corresponding contingency DAG, instructing the agents to continue executing tasks in the new graph from the interruption point. All data during execution, including success and failure logs, are sent to the causally driven skill management module.

[0071] The causally driven skill management module performs in-depth, attributable analysis of task execution results and manages the knowledge (i.e., agent skill models) within the system throughout its lifecycle based on these results.

[0072] Internal structure and process of the causally driven skill management module:

[0073] Causal inference engine: receives execution data from the collaborative execution and switching module. It uses causal discovery algorithms (such as PC algorithm, FCI algorithm) and counterfactual inference to locate the root cause of task failure or inefficiency from complex execution logs, rather than just correlation.

[0074] Skill specialization and pruning: when the root cause is located as a skill defect of a specific agent in a particular context, the module triggers skill specialization, instructing the agent to distill an expert model using failure samples. When the root cause is located as negative transfer of a shared skill to a specific agent, skill pruning is triggered.

[0075] Skill profile database: It is a structured database for fine-grained management of each skill model in the system. A record is maintained for each skill in the database, containing: skill ID, source (generated by which agent in which task), version number, applicable scenario (such as high light reflection surface detection), historical performance (success rate, average time consumption, etc.), and a key negative transfer list. When skill pruning is triggered, the system will add the agent ID that produces negative transfer to the negative transfer list of this skill, so as to prohibit the agent from loading this skill in future task scheduling. This profile provides a solid data foundation for the accurate invocation, evaluation and purification of skills.

[0076] Self-optimizing feedback: When the causal inference engine attributes the failure root cause to prediction error, it generates an instruction to send to the probabilistic planning module, using this failure case as a new sample to optimize its internal time series prediction model.

[0077] Cross-topology meta-learning module is used to learn a "learn how to schedule" meta-ability, rather than a scheduling strategy for a specific task. It works in the background or offline state, collecting a large amount of historical task data organized by the causal-driven skill management module, which constitutes a data pair (DAG topology, efficient scheduling strategy). Using MAML and other meta-learning algorithms, the module extracts a general scheduling meta-strategy from these data across different topologies.

[0078] The trained meta-strategy is loaded into the scheduler of the collaborative execution and switching module. This enables the scheduler to start with the meta-strategy when facing an unseen DAG topology, and quickly find an efficient scheduling scheme through minimal trial and error (few-shot learning), thus achieving rapid generalization to new tasks.

[0079] In summary, the system of the present application, through the close cooperation of the above five modules, constitutes a complete closed loop from task reception, multi-path planning, robust execution to deep analysis and self-evolution.

[0080] The present application is not limited to the above optional embodiments, and anyone can derive other various forms of products under the inspiration of the present application, but regardless of any changes in shape or structure, any technical solution falling within the scope defined by the claims of the present application falls within the protection scope of the present application.

Claims

1. An AI-based industrial edge intelligence task collaboration method, characterized in that, The method comprises the following steps: S1. Receiving an industrial collaborative task and decomposing it into an initial directed acyclic graph composed of multiple subtask nodes; S2. Identifying multiple potential failure modes based on probabilistic prediction of edge environment state; S3. Generating a main execution DAG and at least one corresponding probabilistic DAG for each potential failure mode, the probabilistic DAG being an optimized backup task graph for dealing with a specific failure mode; S4. Using a collaborative scheduler to assign and instruct execution agents on edge devices to collaboratively execute subtasks according to the main execution DAG; S5. Monitoring the edge environment in real time during execution, and when an actual failure mode matching one of the potential failure modes is detected, immediately suspending execution of the main execution DAG and switching to the corresponding probabilistic DAG to continue executing the task; S6. After task execution, using a causal inference-based diagnosis engine to analyze execution data, and when it is determined that the failure root cause is that a certain execution agent lacks the ability to perform a specific task, triggering the agent to distill a general model into an expert skill model; S7. When the causal inference-based diagnosis engine finds that a certain shared skill has negative transfer to certain agents, triggering skill pruning of the skill library of these agents; S8. Using a meta-learning framework to train on multiple DAG tasks with different topologies to learn a scheduling meta-strategy that can quickly adapt to new DAG topologies with few samples.

2. The method of claim 1, wherein, In S5, the switching to the corresponding probabilistic DAG comprises: starting execution from a node in the probabilistic DAG corresponding to the execution progress when the main execution DAG is suspended, and inheriting the results of subtasks completed by the main execution DAG.

3. The method of claim 1, wherein, The step of distilling a general model into an expert skill model in S6 comprises: collecting context data under the specific situation as supervision information, and using the data to distill the knowledge related to the specific situation in the general model into a lightweight expert skill model.

4. The method of claim 1, wherein, In S7, the skill pruning comprises: prohibiting a specific agent from loading or calling a shared skill model diagnosed as having negative transfer.

5. The method of claim 1, wherein, In S8, the meta-learning framework optimizes the scheduling meta-strategy on a large number of historical DAG tasks, so that when faced with a new, unseen DAG topology, the scheduling meta-strategy can converge to an efficient scheduling strategy with only a small number of scheduling instances.

6. The method of claim 1, wherein, The method further comprises: when the causal inference-based diagnosis engine attributes the failure to inaccurate prediction in S2, using execution data leading to the failure as negative samples to optimize the model relied on by the probabilistic prediction.

7. The method of claim 1, wherein, In S2, the probabilistic prediction of edge environment state is realized by a time series graph neural network.

8. The method of claim 1, wherein, The industrial collaborative task is multi-process quality detection on a flexible production line or predictive maintenance of a device group.

9. An AI-based industrial edge intelligence task coordination system, characterized in that, It comprises: a task decomposition module for performing step S1 of claim 1; a probabilistic planning module for performing steps S2 and S3 of claim 1; a collaborative execution and switching module for performing steps S4 and S5 of claim 1; a causally-driven skill management module configured to perform steps S6 and S7 of claim 1; a cross-topology meta-learning module configured to perform step S8 of claim 1.

10. The system of claim 9, wherein, the causally-driven skill management module is further configured to maintain a skill profile recording the skill source, applicable context, performance, and negative transfer of each agent skill to support the decision of the skill pruning.

Images