A method and system for handling production process anomalies based on AI node embedding

Abstract

This invention relates to a production process anomaly handling method based on AI node embedding, belonging to the field of intelligent manufacturing. This method uses an SOP parsing engine to decompose standard operating procedures into atomic-level AI-interventionable nodes, generating a node-based process topology. At the AI ​​node embedding layer, lightweight anomaly detection models, publish-subscribe communication protocols, and self-healing mechanisms are bound to each node, forming an autonomous set of intelligent nodes. Real-time collaboration between nodes triggers anomaly suppression within nodes and propagation suppression between nodes, generating tiered response instructions. A process refactoring engine automatically skips non-critical nodes, activates backup nodes, or enables manual intervention based on production cycle time, outputting an adaptive sequence that maintains process integrity. Incremental learning of node models and continuous evolution of the rule base are achieved through knowledge distillation. This invention realizes the automation, intelligence, and self-evolution of production process anomaly handling, significantly improving anomaly response speed and production continuity, thereby increasing production efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent manufacturing, and in particular to a method and system for handling production process anomalies based on AI node embedding. Background Technology

[0002] With the deepening integration of intelligent manufacturing and industrial production, modern production processes are increasingly characterized by multi-process collaboration, high-speed operation, and flexible production. This places extremely high demands on the efficiency of anomaly detection and handling, as well as process continuity.

[0003] Traditional production process management relies primarily on standardized operating procedures (SOPs) and manual inspection experience. However, existing SOPs are mostly static text documents, making deep integration with automated execution systems difficult, leading to a disconnect between process requirements and actual control. Anomaly detection typically employs fixed threshold alarms or post-incident manual checks, resulting in high response delays and high false negative rates. This fails to meet the real-time demands of high-speed production, thus reducing the efficiency of smart manufacturing and necessitating improvements. Summary of the Invention

[0004] To improve the efficiency of intelligent manufacturing, this application provides a method and system for handling production process anomalies based on AI node embedding.

[0005] Firstly, the above-mentioned inventive objective of this application is achieved through the following technical solution: A method for handling production process anomalies based on AI node embedding, the method comprising the following steps: The production process SOP parsing engine breaks down the standard operating procedure text into multiple atomic-level AI-interventionable nodes. Each node corresponds to clear action semantics, input and output interfaces, and anomaly judgment rules, generating a node-based process execution topology diagram. The node-based topology graph is input into the AI ​​node embedding layer, and a lightweight anomaly detection model, inter-node communication protocol, and node state self-healing mechanism are bound to each node to form a set of intelligent node instances with autonomous capabilities. The set of intelligent node instances is deployed to the production execution system. The nodes exchange execution status and abnormal signals in real time based on a publish-subscribe mechanism. When a node detects an abnormality, it initiates a collaborative mechanism for intra-node abnormality suppression and inter-node abnormality propagation suppression to generate hierarchical abnormality response instructions. The hierarchical anomaly response instructions are input into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes, and outputs an adaptive execution sequence. The processing results of the adaptive execution sequence are fed back to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and regularly updates the node anomaly judgment rule base.

[0006] By adopting the above technical solutions, rigid textual procedures are transformed into a perceptible, judgmental, and self-healing intelligent node network. Compared with traditional manual line inspection, the anomaly detection response speed is improved, and the missed detection rate is effectively reduced. The inter-node publish-subscribe mechanism and the anomaly propagation suppression collaborative mechanism realize rapid isolation and hierarchical response to abnormal events, avoiding defect diffusion and chain shutdowns. Production line downtime due to anomalies is shortened. The automatic backup node call and cycle time compensation capabilities of the process dynamic reconfiguration engine ensure production continuity. The output loss caused by a single anomaly is effectively reduced. The knowledge distillation incremental learning mechanism enables the system to have self-evolution capabilities. The accuracy of identifying new anomalies continues to improve with usage time, solving the pain points of traditional systems' knowledge solidification and inability to adapt to process iterations. Overall, this solution realizes the automation, intelligence, and continuous evolution of production process anomaly handling, significantly improving the production efficiency, availability, and economy of the manufacturing system.

[0007] In a preferred example, this application can be further configured as follows: the step of decomposing the standard operating procedure text into multiple atomic-level AI-interventionable nodes through the production process SOP parsing engine, with each node corresponding to clear action semantics, input / output interfaces, and anomaly judgment rules, and generating a node-based process execution topology diagram, includes the following steps: Construct a semantic parser for production processes and use dependency parsing to extract the action subject, operation object, execution condition and judgment criteria from the SOP text; The extracted semantic elements are mapped to node attribute tuples, including node ID, node type, list of preceding dependent nodes, exception type enumeration, and response latency level. A directed acyclic graph (DAG) algorithm is used to connect the nodes, and the edge weights represent the strength of the material flow or information flow dependency between the nodes, generating an executable process execution topology graph. A virtual monitoring node is inserted into the topology graph, and the virtual monitoring node subscribes to the state change events of adjacent nodes.

[0008] In a preferred embodiment, this application can be further configured as follows: the step of inputting the nodeized topology graph into the AI ​​node embedding layer, binding a lightweight anomaly detection model, an inter-node communication protocol, and a node state self-healing mechanism to each node to form a set of intelligent node instances with autonomous capabilities includes the following steps: Instantiate a lightweight anomaly detection model for each node, process the image or time series data input to the node, and output anomaly confidence score; Define a publish-subscribe communication protocol between nodes. The protocol message includes the source node ID, event type, timestamp, anomaly level, and payload summary, supporting the rapid propagation of anomaly events across nodes. A self-healing mechanism for node states is constructed. When the confidence level of the node detection model output is lower than the threshold, the model is automatically restarted and a spare parameter set is loaded. A resource usage monitor is integrated into the node embedding layer, and a dynamic scheduler is used to adjust the inference batch size of node instances based on node CPU and memory usage.

[0009] In a preferred embodiment, this application can be further configured as follows: The step of deploying the set of intelligent node instances to the production execution system, where nodes exchange execution status and anomaly signals in real time based on a publish-subscribe mechanism, and when a node detects an anomaly, initiates a collaborative mechanism for anomaly suppression within the node and anomaly propagation suppression between nodes to generate a tiered anomaly response instruction, includes the following steps: An in-node anomaly suppression submodule is constructed. When the node detection model outputs an anomaly, the in-node retry logic is first initiated to repeatedly execute the node action and collect multiple sets of data to reduce the false judgment rate. If the error persists after retrying, the node publishes a suppression signal to the subscriber. The suppression signal contains the error type code and the suppression radius parameter. Nodes within the radius will temporarily suspend execution to prevent the error from spreading. Construct an anomaly propagation source graph to record the publication-forwarding path of anomaly events, use graph clustering algorithm to identify key hub nodes of anomaly propagation, and implement mandatory isolation of hub nodes; When generating an exception response command, a high-level response is triggered for exceptions with a false positive rate lower than the set value, and a low-level response is triggered for occasional jitter-type exceptions.

[0010] In a preferred embodiment, this application can be further configured as follows: The step of inputting the hierarchical anomaly response command into the process dynamic reconfiguration engine, which automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes based on production cycle time constraints and resource availability, and outputs an adaptive execution sequence, includes the following steps: The production cycle constraint is modeled as a temporal Petri net, with node execution time as place delay and resource occupancy as transition trigger condition, and the shortest completion time after process reconstruction is automatically calculated. Once an abnormal node is isolated, the engine searches for a functionally equivalent backup node in the knowledge graph. The backup node must have compatible input and output interfaces and resource consumption below a threshold. If no backup node is available, a manual intervention node is triggered. This node pushes AR guidance information to the production line workers, guiding them to manually complete the abnormal node task according to the standard procedure. The reconstructed execution sequence is checked for reachability to ensure that all material flow and information flow paths are connected, and an adaptive execution sequence with a resource allocation Gantt chart is output.

[0011] In a preferred example, this application can be further configured as follows: the step of feeding back the processing result of the adaptive execution sequence to the node knowledge distillation module, wherein the node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and periodically updates the node anomaly judgment rule base, includes the following steps: A teacher-student distillation framework is constructed, in which the teacher model is a high-precision large model in the cloud and the student model is a lightweight model at the node. Anomaly detection knowledge is passed through feature imitation loss. Introducing rule consistency constraints into the distillation loss ensures that the anomaly detection rules learned by the student model are logically consistent with the explicit SOP rules, thus avoiding rule drift caused by distillation. Implement an incremental learning strategy, only back-uploading and retraining difficult samples where nodes are misjudged, reducing cloud communication bandwidth and accelerating node model iteration; The updated node model parameters are packaged with the rule base version number and distributed to the corresponding nodes through a differential update mechanism. After the nodes verify the hash, they perform hot reloading updates.

[0012] Secondly, the above-mentioned inventive objective of this application is achieved through the following technical solutions: A production process anomaly handling device based on AI node embedding, the device includes: a process execution topology graph generation unit, which is used to decompose the standard operating procedure text into multiple atomic-level AI-interventionable nodes through a production process SOP parsing engine, each node corresponding to clear action semantics, input and output interfaces and anomaly judgment rules, to generate a nodeized process execution topology graph. The intelligent node instance set construction unit is used to input the nodeized topology graph into the AI ​​node embedding layer, bind a lightweight anomaly detection model, an inter-node communication protocol, and a node state self-healing mechanism to each node, and form an intelligent node instance set with autonomous capabilities. The hierarchical anomaly response instruction generation unit is used to deploy the set of intelligent node instances to the production execution system. The nodes exchange execution status and anomaly signals in real time based on a publish-subscribe mechanism. When a node detects an anomaly, it activates a collaborative mechanism for anomaly suppression within the node and anomaly propagation suppression between nodes to generate hierarchical anomaly response instructions. An adaptive execution sequence output unit is used to input the hierarchical anomaly response instruction into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes, and outputs an adaptive execution sequence. The node anomaly determination rule base update unit is used to feed back the processing results of the adaptive execution sequence to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and updates the node anomaly determination rule base regularly.

[0013] Thirdly, the above-mentioned objectives of this application are achieved through the following technical solutions: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described AI node-embedded production process anomaly handling method.

[0014] Fourthly, the above-mentioned objectives of this application are achieved through the following technical solutions: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for handling production process anomalies based on AI node embedding. Attached Figure Description

[0015] Figure 1 This is a flowchart of a production process anomaly handling method based on AI node embedding in one embodiment of this application; Figure 2 This is a schematic diagram of a production process anomaly handling device based on AI node embedding in one embodiment of this application; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of this application. Marker description

[0016] 1. Process execution topology graph generation unit; 2. Intelligent node instance set construction unit; 3. Hierarchical exception response instruction generation unit; 4. Adaptive execution sequence output unit; 5. Node exception judgment rule base update unit. Detailed Implementation

[0017] The present application will be further described in detail below with reference to the accompanying drawings.

[0018] In one embodiment, such as Figure 1 As shown, this application discloses a method for handling production process anomalies based on AI node embedding, which specifically includes the following steps: S10: The production process SOP parsing engine breaks down the standard operating procedure text into multiple atomic-level AI-interventionable nodes. Each node corresponds to clear action semantics, input and output interfaces and anomaly judgment rules, generating a node-based process execution topology diagram. In this embodiment of the application, taking the automotive OEM body welding workshop as an example, the production process SOP parsing engine reads the text of the "Body Assembly Welding Operation Instruction" and breaks it down into 6 atomic-level AI-interventionable nodes: N1 "Clamping Pneumatic Positioning" node (controls the action of the clamping cylinder, the input is the positioning signal, the output is the clamping state, and the abnormal judgment rule is "air pressure <0.5MPa or positioning deviation >2mm is abnormal"), N2 "Weld Point Visual Inspection" node (inspects the quality of weld points, the input is the weld point image, the output is the qualified / unqualified label, and the rule is "weld point diameter <4mm or spatter >3 places is abnormal"), N3 "Robot Welding" node (executes welding action), N4 "Quality Sampling Inspection" node (sampling inspection), N5 "Workpiece Flow" node (conveyor belt transport), and N6 "Manual Review" node. The engine identifies N1→N2→N3→N5 as the critical path and N4 and N6 as bypass nodes through dependency parsing, generating a directed acyclic topology graph with dependency weights. The edge weight from N3 to N4 is 0.1 (non-critical), and the edge weight from N2 to N6 is 0.9 (high dependency), providing a structural foundation for subsequent dynamic reconstruction.

[0019] S20: Input the node-based topology graph into the AI ​​node embedding layer, and bind a lightweight anomaly detection model, inter-node communication protocol and node state self-healing mechanism to each node to form a set of intelligent node instances with autonomous capabilities; Specifically, after inputting the above topology map into the AI ​​node embedding layer, each node is endowed with autonomous capabilities. The N2 "weld joint visual inspection" node is bound to a lightweight MobileNetV3 model (model size only 8MB, inference latency 12ms), and subscribes to the N1 "positioning complete" event via the MQTT protocol. Upon receiving the event, it triggers the camera to take a picture and detect. The node state self-healing mechanism is set as follows: when the N2 model outputs a confidence score <0.7 for 5 consecutive times, it automatically loads the pre-trained "low light enhancement" parameter set from the backup parameter server and restarts the model to adapt to sudden changes in workshop lighting. The N5 "workpiece flow" node priority scheduler adopts a reinforcement learning strategy based on the production cycle (60 seconds per vehicle): if the N3 welding node times out due to abnormal retry, it automatically increases the N5 conveyor belt speed from 0.5m / s to 0.8s to compensate for the time loss. The cross-node feature alignment module eliminates the difference in image brightness distribution between the N2 camera and the N6 manual inspection station through adversarial domain adaptation, ensuring that the anomaly judgment criteria are consistent.

[0020] S30: Deploy the set of intelligent node instances to the production execution system. The nodes exchange execution status and abnormal signals in real time based on the publish-subscribe mechanism. When a node detects an abnormality, it starts the collaborative mechanism of intra-node abnormality suppression and inter-node abnormality propagation suppression to generate hierarchical abnormality response instructions. Specifically, after the intelligent node instance is deployed to the production line, node N2 detects a weld point diameter of only 3.5mm when inspecting the 1000th car body, triggering a "weld point too small" anomaly (confidence level 0.95). The node's internal suppression mechanism is activated: the flow of this workpiece is immediately paused, and robot N3 is notified to suspend subsequent welding. Simultaneously, suppression signals are sent to N1, N3, and N5 (suppression radius = 2 nodes). N1 suspends the positioning of the next car body, N3 suspends welding, and N5 stops the conveyor belt to prevent defective workpieces from flowing downstream. The anomaly propagation traceability diagram records the propagation path of the anomaly from N2 → N6 (manual review) → N3 (parameter adjustment), identifying N2 as the source of the anomaly. A tiered response command is generated: due to a false positive rate of <5%, a high-level response (red alert) is triggered. The system determines that it is an equipment problem rather than accidental fluctuations and notifies the maintenance station to intervene.

[0021] S40: Input the hierarchical anomaly response command into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes and outputs an adaptive execution sequence. Specifically, after receiving the N2 red alert, the process dynamic refactoring engine, based on production cycle constraints, analyzed that waiting for N2 repair to resume would take 30 minutes, resulting in a backlog of 60 car bodies. The engine searched for backup nodes in the knowledge graph and found that the N6 manual review node had the "magnifying glass + plug gauge" manual inspection capability and was currently idle. N6 was automatically activated as a backup node, and the 1000th car body was transferred to the N6 station. After manual inspection confirmed the defect, the N3 robot was triggered to repair the weld point. At the same time, the non-critical node N4 quality inspection was skipped (because N6 had already performed equivalent inspection), keeping the overall cycle time at a loss of only 2 minutes. The refactored execution sequence shows: N1→N2 (abnormal)→N6 (backup)→N3 (repair welding)→N5. The path is connected, and the resource allocation Gantt chart shows that subsequent car bodies can catch up with the cycle time, thus maintaining the integrity of the process.

[0022] S50: The processing result of the adaptive execution sequence is fed back to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and regularly updates the node anomaly judgment rule base. Specifically, after the adaptive execution sequence processing is completed, the weld defect of vehicle body number 1000 is confirmed by manual review by N6 as being caused by electrode tip wear. The event feature vector (weld image, current waveform, number of electrode tip uses) is fed back to the node knowledge distillation module. After analyzing the sample, the cloud-based teacher model (ResNet50) discovers that "blurred weld edges caused by electrode wear" is a new anomaly pattern, and distills its knowledge into the student model (MobileNetV3) of N2. After the student model's weights are updated, its sensitivity to weld edge features increases by 40%. Simultaneously, a new rule is added to the anomaly judgment rule base: "When the weld edge gradient < 0.3 and the number of electrode uses > 5000 times, it is judged as an electrode wear anomaly." One week later, N2 detects a similar feature again, with the confidence level increasing from 0.75 to 0.92 and the false positive rate decreasing by 60%. The continuous evolution of node capabilities continuously enhances the system's ability to identify rare anomalies. It should be noted that the data in this embodiment may differ from those in other embodiments, and may contain different data values.

[0023] This application transforms rigid textual procedures into a perceptible, analytical, and self-healing intelligent node network through SOP node decomposition and AI embedding. Compared to traditional manual line inspection, this improves anomaly detection response speed and effectively reduces the missed detection rate. The inter-node publish-subscribe mechanism and anomaly propagation suppression collaborative mechanism enable rapid isolation and tiered response to abnormal events, preventing defect spread and cascading downtime, reducing production line downtime by 70%. The automatic backup node invocation and cycle time compensation capabilities of the process dynamic reconfiguration engine ensure production continuity, reducing output loss caused by a single anomaly by 85%. The knowledge distillation incremental learning mechanism enables the system to self-evolve, with the accuracy of identifying new anomalies continuously improving over time. This addresses the pain points of traditional systems' knowledge stagnation and inability to adapt to process iterations. Overall, this solution achieves automation, intelligence, and continuous evolution in production process anomaly handling, significantly improving the robustness, availability, and economy of the manufacturing system.

[0024] In S10: The standard operating procedure (SOP) text is broken down into multiple atomic-level AI-interventionable nodes through the production process SOP parsing engine. Each node corresponds to a clear action semantic, input / output interface, and anomaly judgment rules, generating a node-based process execution topology diagram. This step includes the following steps: S11: Construct a semantic parser for production processes and use dependency parsing to extract the action subject, operation object, execution condition and judgment criteria from the SOP text; S12: Map the extracted semantic elements to node attribute tuples, including node ID, node type, list of preceding dependent nodes, exception type enumeration, and response latency level; S13: A directed acyclic graph (DAG) algorithm is used to connect the nodes. The edge weights represent the strength of the material flow or information flow dependency between the nodes, generating an executable process execution topology graph. S14: Insert a virtual monitoring node into the topology graph. The virtual monitoring node subscribes to the state change events of adjacent nodes.

[0025] In step S20: Inputting the nodeized topology graph into the AI ​​node embedding layer, binding a lightweight anomaly detection model, inter-node communication protocol, and node state self-healing mechanism to each node to form a set of intelligent node instances with autonomous capabilities, the steps include the following: S21: Instantiate a lightweight anomaly detection model for each node, process the image or time series data input to the node, and output the anomaly confidence score; S22: Defines a publish-subscribe communication protocol between nodes. The protocol message includes the source node ID, event type, timestamp, anomaly level, and payload summary, supporting the rapid propagation of anomaly events across nodes. S23: Construct a node state self-healing mechanism. When the confidence level of the node detection model output is lower than the threshold, the model will be automatically restarted and a spare parameter set will be loaded. S24: Integrate a resource usage monitor into the node embedding layer. The dynamic scheduler adjusts the inference batch size of node instances based on node CPU and memory usage.

[0026] In a car body welding assembly line at a certain automobile manufacturer, the AI ​​node embedding layer endows each workstation with intelligent autonomous capabilities: Specifically, in step S21, this embodiment of the application instantiates a lightweight anomaly detection model for five key nodes. The N2 "weld joint visual inspection" node adopts a MobileNetV3-Tiny architecture (2.3M parameters), deployed on an edge computing box at the workstation, processes 640×480 weld joint images, and outputs three types of anomaly confidence scores: "weld joint too small," "weld burn-through," and "splash." The single-frame inference latency is only 12ms. The N3 "robot welding" node is bound to a vibration timing detection model, receives accelerometer data from the robot's end effector (sampling rate 1kHz), and outputs anomaly scores for "electrode wear" and "trajectory deviation" through a lightweight TCN network. The model, after TensorRT quantization and compression, has a size of <5MB, suitable for node-side deployment, solving the communication latency problem of traditional cloud-based centralized detection and achieving real-time response for anomaly judgment. When N2 detects a weld joint with a diameter of 3.5mm, it immediately outputs an anomaly confidence score of 0.92, triggering local interception without waiting for cloud feedback.

[0027] In step S22, a publish-subscribe communication protocol between nodes is defined. When N2 detects an anomaly, the node publishes a message to the "welding quality" topic: {src_id:"N2", event:"weld_defect", ts:"2024-01-15T10:23:45.123Z", level:"high", payload_digest:"sha256(image feature vector)"}.

[0028] The protocol supports rapid cross-node propagation: after the N1 "fixture positioning" node subscribes to the N2 event, it immediately pauses the next vehicle body positioning to avoid defect accumulation; after receiving the event, the N5 "workpiece transfer" node stops the conveyor belt and routes the defective vehicle body to the rework station. The protocol's payload digest design compresses the message size from complete image data (approximately 300KB) to only 200 bytes, improving communication efficiency by 1500 times and ensuring millisecond-level event broadcasting.

[0029] In step S23, a node self-healing mechanism is constructed. During night shift production, node N2 experiences a sudden change in workshop lighting, causing its model confidence to consistently fall below the 0.7 threshold (normally it should be >0.85). The node self-healing mechanism is automatically triggered: first, a warm restart is performed, unloading the current model and clearing the GPU cache; then, a "low-light enhancement" pre-trained parameter set (fine-tuned on historical night shift data) is loaded from the backup parameter server; after restarting, the model confidence recovers to 0.88. The entire process is completed within 3 seconds without manual intervention. This mechanism achieves node-level fault isolation, preventing the failure of a single model from paralyzing the entire production line. Node availability is increased from 95% to 99.8%, significantly enhancing system robustness.

[0030] In step S24, a resource usage monitor is integrated into the node embedding layer. When the N3 robot welding node experiences CPU utilization of 85% and memory usage of 3.2GB (out of a total of 4GB) due to processing complex welds, the dynamic scheduler automatically reduces its inference batch size from 8 images to 2 and lowers the feature map resolution from 640×480 to 320×240, reducing resource usage to 62% and 2.1GB respectively. Simultaneously, the scheduler detects that the N1 fixture node is only at 30% load and dynamically allocates some of N1's idle computing resources (1 CPU core + 0.5GB memory) to N3, achieving cluster load balancing. This elastic scheduling ensures stable operation of all nodes under resource constraints, preventing node overload and crashes, improving overall system resource utilization by 40%, and ensuring that high-priority nodes (N2 quality inspection) always receive sufficient resources.

[0031] Steps S21-S24 construct a fully autonomous intelligent node cluster through lightweight model instantiation, efficient communication protocols, node self-healing mechanisms, and dynamic resource scheduling. Compared to traditional centralized monitoring systems, node-level real-time detection effectively reduces anomaly response latency, significantly improving production safety. The efficiency of the publish-subscribe communication protocol increases cross-node collaboration efficiency by several orders of magnitude, avoiding batch losses caused by anomaly propagation. The node self-healing mechanism brings system availability close to telecom-grade standards, significantly reducing manual maintenance costs. Elastic resource scheduling optimizes the return on hardware investment, enabling production lines of the same scale to reduce investment in edge computing equipment. Overall, this AI node embedding layer enables decentralized, autonomous, and collaborative handling of production process anomalies, providing a solid technical foundation for flexible manufacturing and zero-downtime production goals.

[0032] In step S30: Deploying the set of intelligent node instances to the production execution system, with nodes exchanging execution status and anomaly signals in real time based on a publish-subscribe mechanism, and when a node detects an anomaly, activating a collaborative mechanism for anomaly suppression within the node and anomaly propagation suppression between nodes to generate a tiered anomaly response instruction, the following steps are included: S31: Construct an intra-node anomaly suppression submodule. When the node detection model outputs an anomaly, first start the intra-node retry logic, repeatedly execute the node action and collect multiple sets of data to reduce the false judgment rate. S32: If the exception persists after retrying, the node publishes a suppression signal to the subscriber. The suppression signal contains the exception type code and the suppression radius parameter. Nodes within the radius will temporarily suspend execution to prevent the exception from spreading. S33: Construct an anomaly propagation source tracing graph, record the publication-forwarding path of anomaly events, use graph clustering algorithm to identify key hub nodes of anomaly propagation, and implement mandatory isolation of hub nodes; S34: When generating an abnormal response command, a high-level response is triggered for abnormalities with a false positive rate lower than the set value, and a low-level response is triggered for occasional jitter-type abnormalities.

[0033] In this embodiment, at the "weld point visual inspection" node (N2) of the automotive body welding assembly line, when the lightweight model detects an "oversized weld point" anomaly for the first time, the anomaly suppression submodule within the node automatically triggers retry logic: controlling the robotic arm to reposition the vehicle body to the inspection position, triggering the camera to take pictures again, and collecting multiple sets of weld point image data. By repeatedly executing the inspection action and fusing the features collected multiple times, the system effectively distinguishes between real defects and occasional misjudgments caused by light fluctuations or dust obstruction, significantly reducing the false alarm rate caused by environmental interference and avoiding unnecessary line stoppages.

[0034] When the N2 node's anomaly persists after retrying, confirming the weld defect as a genuine problem, the node immediately sends a suppression signal to subscribers. This signal includes the "weld defect" anomaly type code and a suppression radius parameter (covering adjacent workstations before and after). Upon receiving the suppression signal, the N1 "fixture positioning" node suspends the positioning action of the next car body; the N3 "robot welding" node pauses the welding program to prevent material waste caused by continued welding on the defective car body; and the N5 "workpiece flow" node stops the conveyor belt, isolating the defective car body at the current workstation. Through the coordinated suspending of nodes within the radius, the defective car body is effectively prevented from flowing into downstream workstations, blocking the spread of the anomaly in the process chain and ensuring the quality of mass production.

[0035] The system continuously records the release-forward path of abnormal events, constructing an anomaly propagation source map. In a series of "solder joint defects" events, the source map shows that the anomaly propagated frequently through the path N2→N3→N5, with node N3, as the welding execution unit, located at the center of the graph connecting multiple upstream detection nodes and downstream transfer nodes. After analysis using a graph clustering algorithm, N3 was identified as the critical hub node for anomaly propagation. Even if N2's detection is normal, wear on N3's electrode tip can still cause batch solder joint defects. The system forcibly isolates N3, suspending its welding function and triggering a maintenance work order, cutting off the anomaly propagation path at its source and significantly improving the overall stability of the production line.

[0036] The safety control decision module manages anomalies in a tiered manner: For serious anomalies detected by N2, such as "solder joint size being below standard for three consecutive times," which have an extremely low false positive rate and exhibit a systematic pattern, a high-level response is triggered, including immediate line stoppage, notification of quality engineers, and initiation of full batch traceability. Conversely, for occasional jitter anomalies such as "blurred solder joint image" caused by occasional light flicker, which have a higher false positive rate, a low-level response is triggered, only logging is performed and camera exposure parameters are automatically adjusted without interrupting production. This tiered management strategy prioritizes the handling of high-risk anomalies with limited production resources, avoiding frequent line stoppages due to over-response, and effectively improving overall production line efficiency while ensuring safety.

[0037] The intra-node anomaly suppression submodule significantly reduces the false positive rate caused by environmental interference through retry logic, improving the accuracy and robustness of anomaly detection. The inter-node suppression signal broadcasting mechanism enables rapid isolation and propagation blocking of anomalies, preventing the cascading propagation of defects in the process chain and significantly reducing batch quality incidents. Anomaly propagation tracing and hub node identification functions enable the system to locate systemic risk sources hidden deep within the process (such as equipment wear), achieving a leap from superficial handling to root cause eradication. The hierarchical anomaly management strategy optimizes the allocation of response resources, enabling rapid handling of severe anomalies and flexible processing of occasional disturbances, significantly improving the continuous operation capability and overall production efficiency of the production line while ensuring safety. Overall, this collaborative mechanism enables precise, proactive, and intelligent handling of production process anomalies, significantly reducing downtime losses and quality costs caused by anomalies.

[0038] In step S40: Inputting the hierarchical anomaly response command into the process dynamic reconfiguration engine, which automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes based on production cycle constraints and resource availability, and outputs an adaptive execution sequence, includes the following steps: S41: Model the production cycle constraint as a temporal Petri net, with node execution time as place delay and resource occupation as transition trigger condition, and automatically calculate the shortest completion time after process refactoring; S42: After an abnormal node is isolated, the engine searches for a functionally equivalent backup node in the knowledge graph. The backup node must meet the requirements of compatible input and output interfaces and resource consumption below the threshold. S43: If no backup node is available, a manual intervention node is triggered. This node pushes AR guidance information to the production line workers, guiding them to manually complete the abnormal node task according to the standard procedure. S44: Perform reachability verification on the reconstructed execution sequence to ensure that all material flow and information flow paths are connected, and output an adaptive execution sequence with a resource allocation Gantt chart.

[0039] In step S50: the processing result of the adaptive execution sequence is fed back to the node knowledge distillation module, which has a teacher-student distillation framework to achieve incremental learning of the node model and periodically updates the node anomaly judgment rule base. This step includes the following steps: S51: Construct a teacher-student distillation framework. The teacher model is a high-precision large model in the cloud, and the student model is a lightweight model at the node. Anomaly discrimination knowledge is passed through feature imitation loss. S52: Introduce rule consistency constraints in the distillation loss to ensure that the anomaly judgment rules learned by the student model are logically consistent with the explicit SOP rules, and avoid rule drift caused by distillation. S53: Implement an incremental learning strategy, only back-uploading and retraining difficult samples where nodes are misjudged, reducing cloud communication bandwidth and accelerating node model iteration; S54: Package the updated node model parameters and rule base version number, and distribute them to the corresponding nodes through a differential update mechanism. After the node verifies the hash, it performs a hot reload update.

[0040] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0041] In one embodiment, a production process anomaly handling device based on AI node embedding is provided, which corresponds one-to-one with the production process anomaly handling method based on AI node embedding in the above embodiments. For example... Figure 2 As shown, the production process anomaly handling device based on AI node embedding includes: a process execution topology graph generation unit 1, which is used to decompose the standard operating procedure text into multiple atomic-level AI-interventionable nodes through the production process SOP parsing engine. Each node corresponds to a clear action semantic, input and output interface and anomaly judgment rule, generating a node-based process execution topology graph. The intelligent node instance set construction unit 2 is used to input the nodeized topology graph into the AI ​​node embedding layer, bind a lightweight anomaly detection model, an inter-node communication protocol and a node state self-healing mechanism to each node, and form an intelligent node instance set with autonomous capabilities. The hierarchical anomaly response instruction generation unit 3 is used to deploy the set of intelligent node instances to the production execution system. The nodes exchange execution status and anomaly signals in real time based on a publish-subscribe mechanism. When a node detects an anomaly, it activates a collaborative mechanism for anomaly suppression within the node and anomaly propagation suppression between nodes to generate hierarchical anomaly response instructions. The adaptive execution sequence output unit 4 is used to input the hierarchical abnormal response instruction into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes and outputs an adaptive execution sequence. The node anomaly determination rule base update unit 5 is used to feed back the processing results of the adaptive execution sequence to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and updates the node anomaly determination rule base regularly.

[0042] Specific limitations regarding the AI ​​node-embedded production process anomaly handling device can be found in the limitations of the AI ​​node-embedded production process anomaly handling method described above, and will not be repeated here. Each module in the aforementioned AI node-embedded production process anomaly handling device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the electronic device, or stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of each module.

[0043] In one embodiment, an electronic device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores the database. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a production process anomaly handling method based on AI node embedding.

[0044] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: The production process SOP parsing engine breaks down the standard operating procedure text into multiple atomic-level AI-interventionable nodes. Each node corresponds to clear action semantics, input and output interfaces, and anomaly judgment rules, generating a node-based process execution topology diagram. The node-based topology graph is input into the AI ​​node embedding layer, and a lightweight anomaly detection model, inter-node communication protocol, and node state self-healing mechanism are bound to each node to form a set of intelligent node instances with autonomous capabilities. The set of intelligent node instances is deployed to the production execution system. The nodes exchange execution status and abnormal signals in real time based on a publish-subscribe mechanism. When a node detects an abnormality, it initiates a collaborative mechanism for intra-node abnormality suppression and inter-node abnormality propagation suppression to generate hierarchical abnormality response instructions. The hierarchical anomaly response instructions are input into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes, and outputs an adaptive execution sequence. The processing results of the adaptive execution sequence are fed back to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and regularly updates the node anomaly judgment rule base.

[0045] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The production process SOP parsing engine breaks down the standard operating procedure text into multiple atomic-level AI-interventionable nodes. Each node corresponds to clear action semantics, input and output interfaces, and anomaly judgment rules, generating a node-based process execution topology diagram. The node-based topology graph is input into the AI ​​node embedding layer, and a lightweight anomaly detection model, inter-node communication protocol, and node state self-healing mechanism are bound to each node to form a set of intelligent node instances with autonomous capabilities. The set of intelligent node instances is deployed to the production execution system. The nodes exchange execution status and abnormal signals in real time based on a publish-subscribe mechanism. When a node detects an abnormality, it initiates a collaborative mechanism for intra-node abnormality suppression and inter-node abnormality propagation suppression to generate hierarchical abnormality response instructions. The hierarchical anomaly response instructions are input into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes, and outputs an adaptive execution sequence. The processing results of the adaptive execution sequence are fed back to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and regularly updates the node anomaly judgment rule base.

[0046] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0047] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0048] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for handling production process anomalies based on AI node embedding, characterized in that, The method includes the following steps: using a production process SOP parsing engine to decompose the standard operating procedure text into multiple atomic-level AI-interventionable nodes, each node corresponding to clear action semantics, input and output interfaces and anomaly judgment rules, generating a node-based process execution topology diagram; The node-based topology graph is input into the AI ​​node embedding layer, and a lightweight anomaly detection model, inter-node communication protocol, and node state self-healing mechanism are bound to each node to form a set of intelligent node instances with autonomous capabilities. The set of intelligent node instances is deployed to the production execution system. The nodes exchange execution status and abnormal signals in real time based on a publish-subscribe mechanism. When a node detects an abnormality, it initiates a collaborative mechanism for intra-node abnormality suppression and inter-node abnormality propagation suppression to generate hierarchical abnormality response instructions. The hierarchical anomaly response instructions are input into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes, and outputs an adaptive execution sequence. The processing results of the adaptive execution sequence are fed back to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and regularly updates the node anomaly judgment rule base.

2. The method for handling production process anomalies based on AI node embedding according to claim 1, characterized in that, The step of breaking down the standard operating procedure (SOP) text into multiple atomic-level AI-interventionable nodes using a production process SOP parsing engine, with each node corresponding to clear action semantics, input / output interfaces, and anomaly judgment rules, to generate a node-based process execution topology diagram, includes the following steps: Construct a semantic parser for production processes and use dependency parsing to extract the action subject, operation object, execution condition and judgment criteria from the SOP text; The extracted semantic elements are mapped to node attribute tuples, including node ID, node type, list of preceding dependent nodes, exception type enumeration, and response latency level. A directed acyclic graph (DAG) algorithm is used to connect the nodes, and the edge weights represent the strength of the material flow or information flow dependency between the nodes, generating an executable process execution topology graph. A virtual monitoring node is inserted into the topology graph, and the virtual monitoring node subscribes to the state change events of adjacent nodes.

3. The method for handling production process anomalies based on AI node embedding according to claim 2, characterized in that, The step of inputting the nodeized topology graph into the AI ​​node embedding layer, binding a lightweight anomaly detection model, inter-node communication protocol, and node state self-healing mechanism to each node to form a set of intelligent node instances with autonomous capabilities includes the following steps: Instantiate a lightweight anomaly detection model for each node, process the image or time series data input to the node, and output anomaly confidence score; Define a publish-subscribe communication protocol between nodes. The protocol message includes the source node ID, event type, timestamp, anomaly level, and payload summary, supporting the rapid propagation of anomaly events across nodes. A self-healing mechanism for node states is constructed. When the confidence level of the node detection model output is lower than the threshold, the model is automatically restarted and a spare parameter set is loaded. A resource usage monitor is integrated into the node embedding layer, and a dynamic scheduler is used to adjust the inference batch size of node instances based on node CPU and memory usage.

4. The method for handling production process anomalies based on AI node embedding according to claim 3, characterized in that, The step of deploying the set of intelligent node instances to the production execution system, exchanging execution status and abnormal signals in real time between nodes based on a publish-subscribe mechanism, and initiating a collaborative mechanism for intra-node abnormality suppression and inter-node abnormality propagation suppression when a node detects an abnormality to generate a tiered abnormality response instruction includes the following steps: An in-node anomaly suppression submodule is constructed. When the node detection model outputs an anomaly, the in-node retry logic is first initiated to repeatedly execute the node action and collect multiple sets of data to reduce the false judgment rate. If the error persists after retrying, the node publishes a suppression signal to the subscriber. The suppression signal contains the error type code and the suppression radius parameter. Nodes within the radius will temporarily suspend execution to prevent the error from spreading. Construct an anomaly propagation source graph to record the publication-forwarding path of anomaly events, use graph clustering algorithm to identify key hub nodes of anomaly propagation, and implement mandatory isolation of hub nodes; When generating an exception response command, a high-level response is triggered for exceptions with a false positive rate lower than the set value, and a low-level response is triggered for occasional jitter-type exceptions.

5. The method for handling production process anomalies based on AI node embedding according to claim 4, characterized in that, The step of inputting the hierarchical anomaly response command into the process dynamic reconfiguration engine, which automatically skips non-critical nodes, starts backup nodes, or enables manually intervened nodes based on production cycle time constraints and resource availability, and outputs an adaptive execution sequence, includes the following steps: The production cycle constraint is modeled as a temporal Petri net, with node execution time as place delay and resource occupancy as transition trigger condition, and the shortest completion time after process reconstruction is automatically calculated. Once an abnormal node is isolated, the engine searches for a functionally equivalent backup node in the knowledge graph. The backup node must have compatible input and output interfaces and resource consumption below a threshold. If no backup node is available, a manual intervention node is triggered. This node pushes AR guidance information to the production line workers, guiding them to manually complete the abnormal node task according to the standard procedure. The reconstructed execution sequence is checked for reachability to ensure that all material flow and information flow paths are connected, and an adaptive execution sequence with a resource allocation Gantt chart is output.

6. The method for handling production process anomalies based on AI node embedding according to claim 5, characterized in that, The step of feeding back the processing results of the adaptive execution sequence to the node knowledge distillation module, which has a teacher-student distillation framework to achieve incremental learning of the node model and periodically updates the node anomaly judgment rule base, includes the following steps: A teacher-student distillation framework is constructed, in which the teacher model is a high-precision large model in the cloud and the student model is a lightweight model at the node. Anomaly detection knowledge is passed through feature imitation loss. Introducing rule consistency constraints into the distillation loss ensures that the anomaly detection rules learned by the student model are logically consistent with the explicit SOP rules, thus avoiding rule drift caused by distillation. Implement an incremental learning strategy, only back-uploading and retraining difficult samples where nodes are misjudged, reducing cloud communication bandwidth and accelerating node model iteration; The updated node model parameters are packaged with the rule base version number and distributed to the corresponding nodes through a differential update mechanism. After the nodes verify the hash, they perform hot reloading updates.

7. A production process anomaly handling device based on AI node embedding, applied to the production process anomaly handling method based on AI node embedding as described in any one of claims 1 to 6, characterized in that, The device includes: a process execution topology generation unit (1), which is used to decompose the standard operating procedure text into multiple atomic-level AI-interventionable nodes through the production process SOP parsing engine, each node corresponding to a clear action semantics, input and output interfaces and anomaly judgment rules, and generate a node-based process execution topology. The intelligent node instance set construction unit (2) is used to input the nodeized topology graph into the AI ​​node embedding layer, bind a lightweight anomaly detection model, inter-node communication protocol and node state self-healing mechanism to each node, and form an intelligent node instance set with autonomous capabilities. The hierarchical anomaly response instruction generation unit (3) is used to deploy the set of intelligent node instances to the production execution system. The nodes exchange execution status and anomaly signals in real time based on the publish-subscribe mechanism. When a node detects an anomaly, it starts the collaborative mechanism of anomaly suppression within the node and anomaly propagation suppression between nodes to generate hierarchical anomaly response instructions. The adaptive execution sequence output unit (4) is used to input the hierarchical abnormal response instruction into the process dynamic reconstruction engine. Based on production cycle constraints and resource availability, the engine automatically skips non-critical nodes, starts backup nodes, or enables manual intervention nodes and outputs an adaptive execution sequence. The node anomaly determination rule base update unit (5) is used to feed back the processing results of the adaptive execution sequence to the node knowledge distillation module. The node knowledge distillation module has a teacher-student distillation framework to realize incremental learning of the node model and regularly updates the node anomaly determination rule base.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the production process exception handling method based on AI node embedding as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the production process exception handling method based on AI node embedding as described in any one of claims 1 to 6.

Images