Monitoring and closed-loop control method and system based on large model and multi-agent cooperation

Abstract

The application relates to the technical field of equipment monitoring and intelligent control, and discloses a monitoring and closed-loop control method and system based on large models and multi-agent cooperation, which comprises the following steps: obtaining a control instruction stream and a response data stream, performing dynamic time warping and feature decoupling, and generating a structured perception event; determining a target physical root cause in combination with an operation context and a mechanism knowledge graph; determining a target adjustment strategy according to material attribute boundaries and equipment motion boundaries; generating a bottom-layer control instruction and executing the same after completing a rehearsal check, while back-feeding a newly-added response data stream to form a closed loop. The application performs state perception by combining a control instruction stream and a response data stream, realizes root cause positioning in combination with a mechanism knowledge graph, determines an adjustment strategy under the constraint of physical boundaries, forms a closed loop through rehearsal checking and feedback back-feeding, and thus improves the accuracy of abnormality identification and the executability of an adjustment strategy, and improves the stability of a manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of equipment monitoring and intelligent control technology, and in particular to a monitoring and closed-loop control method and system based on a large model and multi-agent collaboration. Background Technology

[0002] As industrial manufacturing moves towards higher precision and intelligence, the ability to monitor and control manufacturing equipment under complex conditions has become a key factor affecting processing quality and operational stability. In processing involving the coupling of heat transfer, material flow, and mechanical motion, subtle fluctuations in local parameters can easily accumulate and amplify in subsequent processing, leading to a decline in forming quality or operational abnormalities. Therefore, how to effectively sense and stably regulate the processing process has become a key focus of existing technologies.

[0003] In existing technologies, process monitoring largely relies on external observation information, such as visual inspection, temperature detection, or single electrical parameter detection. While these methods can reflect local state changes, they lack the ability to perceive coupled changes within the processing process. In cases of complex processing structures or obstructions, internal anomalies such as thermal hysteresis, execution deviations, and vibration disturbances are difficult to detect in a timely manner, leading to discrepancies between monitoring results and actual operating conditions, and affecting the accuracy of anomaly identification.

[0004] In anomaly identification, existing technologies mostly rely on threshold comparisons or empirical rules for judgment. When monitored data exceeds preset ranges, they can only indicate an abnormal state, but it is difficult to further analyze the specific cause of the anomaly. Because there are coupling relationships between various process parameters, the same abnormal phenomenon may correspond to different causes, and monitoring a single parameter cannot reveal their inherent connections, resulting in diagnostic results remaining superficial. Subsequent adjustments rely on trial and error based on experience, lacking specificity.

[0005] While some technologies can adjust parameters based on monitoring results during the control and regulation process, the adjustment process often lacks the coordinated constraints of material properties and equipment motion. Under complex operating conditions, it is easy to generate adjustment results that exceed the actual physical boundaries, affecting processing stability. At the same time, the selection criteria for candidate regulation schemes are relatively singular, making it difficult to achieve a comprehensive balance between quality, efficiency, and energy consumption.

[0006] Furthermore, existing technologies suffer from weak integration between monitoring, diagnosis, and control processes, making it difficult to promptly relay feedback after control execution back to subsequent processing, thus hindering the formation of a continuous closed loop. This incomplete information transmission makes it difficult for the system to continuously correct its operational state, reducing overall control effectiveness. Therefore, existing technologies still have shortcomings in multi-source state perception, anomaly root cause analysis, constrained adjustment, and closed-loop feedback, making it difficult to meet the application requirements of complex manufacturing scenarios. Summary of the Invention

[0007] The main objective of this invention is to provide a monitoring and closed-loop control method and system based on a large model and multi-agent collaboration, aiming to solve the technical problem that existing technologies are unable to accurately perceive the internal state of the processing process and locate the root cause of anomalies under multi-source coupled conditions, and generate an executable and continuous closed-loop regulation and control strategy under physical constraints.

[0008] To achieve the above objectives, this invention provides a monitoring and closed-loop control method based on a large model and multi-agent collaboration, comprising:

[0009] Acquire control command streams and response data streams, and extract the runtime context based on the control command streams;

[0010] The control command stream and the response data stream are subjected to dynamic time warping to generate an aligned residual sequence;

[0011] The aligned residual sequence is subjected to feature decoupling processing to separate state feature components, and structured sensing events are generated based on the deviation of the state feature components from the historical normal process distribution.

[0012] The structured perception event and the runtime context are mapped to entity nodes in the mechanism knowledge graph, and inverse probability inference is performed along the causal edges in the mechanism knowledge graph to determine the target physical root cause;

[0013] Physical safety boundary conditions are constructed based on material property boundaries and equipment motion boundaries. A set of candidate strategies is generated based on the target physical root cause. The set of candidate strategies is screened using the physical safety boundary conditions. Finally, a target adjustment strategy is determined from the screened set of candidate strategies using a value evaluation function.

[0014] Perform physical feasibility simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility simulation and thought chain verification into low-level control instructions and write them into the device control terminal.

[0015] The newly generated response data stream is collected and used as a response data stream for subsequent acquisition of control command stream and response data stream.

[0016] Furthermore, to achieve the above objectives, the present invention provides a monitoring and closed-loop control system based on a large model and multi-agent collaboration, comprising:

[0017] The dual-stream acquisition module is used to acquire control command streams and response data streams, and extract the runtime context based on the control command streams;

[0018] The timing alignment module is used to perform dynamic time warping on the control command stream and the response data stream to generate an aligned residual sequence.

[0019] The decoupling sensing module is used to perform feature decoupling processing on the aligned residual sequence, separate the state feature components, and generate structured sensing events based on the deviation of the state feature components from the historical normal process distribution.

[0020] The root cause inference module is used to map the structured perception event and the running context to entity nodes in the mechanism knowledge graph, and perform inverse probability inference along the causal edges in the mechanism knowledge graph to determine the target physical root cause;

[0021] The safety decision module is used to construct physical safety boundary conditions based on material property boundaries and equipment motion boundaries, generate a set of candidate strategies based on the target physical root causes, filter the set of candidate strategies using the physical safety boundary conditions, and determine the target adjustment strategy from the filtered set of candidate strategies through a value evaluation function.

[0022] The pre-simulation delivery module is used to perform physical feasibility pre-simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility pre-simulation and thought chain verification into low-level control instructions and write them into the device control terminal.

[0023] The feedback feedback module is used to collect the new response data stream generated by executing the underlying control command, and to use the new response data stream as a response data stream for subsequent acquisition of control command stream and response data stream.

[0024] Beneficial Effects: This invention relates to the field of equipment monitoring and intelligent control technology, and discloses a monitoring and closed-loop control method and system based on a large model and multi-agent collaboration. The method includes: acquiring control command streams and response data streams and extracting the operating context; performing dynamic time warping on the control command streams and response data streams to generate aligned residual sequences; performing feature decoupling on the aligned residual sequences to separate state feature components, and generating structured sensing events based on the state feature components; mapping the structured sensing events and operating context to a mechanistic knowledge graph to perform inverse probabilistic deduction to determine the target physical root cause; screening a set of candidate strategies based on material property boundaries and equipment motion boundaries and determining the target adjustment strategy; performing physical feasibility pre-playing and thought chain verification on the target adjustment strategy; converting the target adjustment strategy that meets the conditions into a low-level control command and writing it into the equipment control terminal; collecting the newly generated response data streams from the execution of the low-level control commands and feeding them back into the subsequent processing to form closed-loop control. This invention extracts state features by fusing control command flow and response data flow, locates the root cause of anomalies by combining mechanistic knowledge graphs, and selects adjustment strategies under the constraints of material property boundaries and equipment motion boundaries. At the same time, it forms a closed loop through physical feasibility simulation and feedback reinjection, thereby improving the accuracy of anomaly identification and the executability of adjustment strategies, and enhancing the stability of the manufacturing process. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0026] Figure 1 This is a schematic diagram of an application environment for a monitoring and closed-loop control method based on large model and multi-agent collaboration in one embodiment of the present invention.

[0027] Figure 2 This is a flowchart illustrating an embodiment of the monitoring and closed-loop control method based on large model and multi-agent collaboration of the present invention.

[0028] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the monitoring and closed-loop control system based on large model and multi-agent collaboration of the present invention.

[0029] Figure 4 This is a system architecture diagram of a monitoring and closed-loop control method based on large model and multi-agent collaboration in one embodiment of this application;

[0030] Figure 5 This is a flowchart of the PMA agent workflow in a monitoring and closed-loop control method based on large model and multi-agent collaboration in one embodiment of this application.

[0031] Figure 6 This is a flowchart of EDA agent logic reasoning in a monitoring and closed-loop control method based on large model and multi-agent collaboration in one embodiment of this application;

[0032] Figure 7 This is a DSA agent decision control logic diagram in a monitoring and closed-loop control method based on large model and multi-agent collaboration in one embodiment of this application;

[0033] Figure 8 This is a diagram of the virtual-real migration deployment architecture in a monitoring and closed-loop control method based on large model and multi-agent collaboration in one embodiment of this application. Detailed Implementation

[0034] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0035] The monitoring and closed-loop control method based on large model and multi-agent collaboration provided in this invention can be applied to applications such as... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain control command streams and response data streams from the client and extract the runtime context. It then performs dynamic time warping on the control command streams and response data streams to generate aligned residual sequences. These sequences are then decoupled to separate state feature components, and structured sensing events are generated based on these components. The structured sensing events and runtime context are mapped to a mechanistic knowledge graph, and inverse probabilistic inference is performed to determine the target physical root cause. A candidate strategy set is selected based on material property boundaries and equipment motion boundaries, and the target adjustment strategy is determined. Physical feasibility pre-running and thought chain verification are performed on the target adjustment strategy. Target adjustment strategies that meet the conditions are converted into underlying control commands and written to the equipment control terminal. New response data streams generated by executing the underlying control commands are collected and fed back into subsequent processing, forming a closed-loop control. This invention extracts state features by fusing control command flow and response data flow, combines mechanistic knowledge graphs to locate the root cause of anomalies, and filters adjustment strategies under the constraints of material property boundaries and equipment motion boundaries. Simultaneously, it forms a closed loop through physical feasibility simulation and feedback loop, thereby improving the accuracy of anomaly identification and the executability of adjustment strategies, and enhancing the stability of the manufacturing process. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0036] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the monitoring and closed-loop control method based on large model and multi-agent collaboration provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0037] like Figure 2 As shown, the monitoring and closed-loop control method based on large model and multi-agent collaboration proposed in this invention includes the following steps:

[0038] S10, acquire the control command stream and response data stream, and extract the running context based on the control command stream;

[0039] In this embodiment, acquiring the control command stream corresponds to the time-sequential acquisition of the command sequence continuously output by the equipment control terminal during processing. This command sequence essentially reflects the scheduling intent and action arrangement of the equipment's execution behavior. The source of the control command stream can be control codes generated during CNC program execution, motion command sequences issued by the upper-level control system, or execution sequences generated by the internal scheduling unit of the embedded controller. In the implementation process, a communication interface is set up to listen to the command output of the equipment control terminal, and the continuously output control commands are sequentially stored through a buffer queue. A unified time base is attached when writing to the buffer, thereby forming a control command stream with temporal continuity. To ensure the stability of subsequent processing, abnormal interruptions, duplicate writes, or missing commands are corrected during the caching process. For example, missing commands are filled through time interval detection or duplicate commands are eliminated through deduplication strategies, thereby obtaining a control command stream with continuous structure and complete timing.

[0040] The acquisition of the response data stream corresponds to the unified organization of data collected from multiple sensors during equipment operation. The sources of the response data stream include thermal response data collected by temperature sensors, mechanical response data collected by vibration sensors, and electric drive response data collected by current sensors. These data reflect the actual physical response of the equipment after executing control commands. In implementation, data acquisition channels with each sensor are established through an edge computing terminal. Data from different sampling frequencies are resampled according to a unified time base and aggregated according to the sampling time order to form a response data stream with a unified time axis. To improve data consistency, time offsets between different data sources are aligned and corrected during the aggregation process. For example, timestamp mapping or interpolation is used to align data from different sampling frequencies to the same time scale, ensuring that the response data stream accurately reflects the continuous changes in the equipment's operating state.

[0041] Extracting the runtime context corresponds to extracting a set of information from the control command stream that characterizes the current processing stage and behavioral semantics. The runtime context originates from the control field information implicit in the control command stream; for example, motion path information reflects the current processing area of ​​the equipment, energy control information reflects heating or driving intensity, and level switching information reflects changes in processing stage. In implementation, the control command stream is semantically parsed, each control command is broken down into multiple field structures, and these fields are categorized and mapped. Motion-related fields are mapped to processing segment identifiers, action-related fields are mapped to action type identifiers, and state switching fields are mapped to operating condition switching identifiers. These identifiers are then combined in chronological order to form a runtime context sequence, ensuring that each time slice corresponds to a clear processing semantic expression. To enhance expressive power, historical command fragments or context information from adjacent time slices can be further integrated into the runtime context, forming a continuous context description through a sliding window approach, enabling subsequent processing to make judgments based on more complete behavioral information.

[0042] In one implementation, control command streams are acquired via a serial communication interface, using a circular buffer to cache the commands. During caching, the commands are parsed in real time, and the parsed results are stored synchronously with the original commands for rapid subsequent retrieval. Response data streams are acquired through multi-channel data acquisition modules, with different sampling frequencies set for different sensors, and a time synchronization module unifies the data from all channels to the same time base. In another implementation, control command and response data streams can be acquired via network communication. A message queue-based communication mechanism is established between the device control terminal and the edge terminal, enabling command and data transmission in an event-driven manner, thereby reducing data congestion. In implementations suitable for high real-time scenarios, an interrupt-triggered mechanism can be used to immediately trigger acquisition and writing when control commands are output or sensor data is updated, thereby reducing time latency.

[0043] This embodiment unifies the acquisition of control command streams and response data streams and extracts the operating context, enabling the device behavior input, physical response output, and processing semantic expression to form a continuous and interconnected data expression structure. This provides a consistent data foundation for subsequent anomaly identification and causal analysis, improving the completeness and accuracy of state perception.

[0044] S20, perform dynamic time warping on the control command stream and the response data stream to generate an aligned residual sequence;

[0045] In this embodiment, dynamic time warping corresponds to correcting the time offset between the control command stream and the response data stream, so that the control behavior and the actual response correspond on a unified time scale. The control command stream is usually output according to the control cycle, and the response data stream is usually acquired according to the sensor sampling cycle. The two differ in sampling frequency, time interval, and change rhythm, so direct comparison can easily cause time misalignment, resulting in the control action and the actual response not being accurately matched.

[0046] In implementation, the control command stream is divided into multiple control command time segments according to a unified time window. Each control command time segment contains a set of control behaviors within a certain time range. The response data stream is divided into multiple response data time segments according to the same time window. Each response data time segment contains response information such as temperature changes, vibration changes, and current changes within the corresponding time range. The unified time window can be set based on the equipment control cycle, sensor sampling frequency, or processing cycle time.

[0047] To match control command time segments with response data time segments, feature extraction is required for both types of time segments. Features in the control command time segment include action switching frequency, setpoint variation amplitude, control command duration, and control command density, reflecting changes in control behavior. Features in the response data time segment include temperature fluctuation amplitude, vibration peak value variation, current variation trend, and response stability, reflecting changes in the physical state of the equipment.

[0048] After feature extraction, distance calculations are performed on the control command feature sequences and response data feature sequences to generate a distance matrix. Each element in the distance matrix represents the degree of difference between a control command time segment and a response data time segment. Distance calculations can use Euclidean distance, absolute distance, or cosine distance, etc., to reflect the proximity of the two types of time segments in the feature space.

[0049] Based on the distance matrix, the matching relationships between each time segment are cumulatively calculated to generate a cumulative distance matrix. Each position in the cumulative distance matrix represents the minimum cumulative difference from the starting time segment to the current position. Subsequently, the mapping path corresponding to the minimum cumulative distance is determined according to the cumulative distance matrix, so that an optimal correspondence is formed between the control command time segment and the response data time segment.

[0050] After determining the mapping path, the control command stream and response data stream are time-aligned according to the mapping path to obtain a time-consistent sequence of control command segments and response data segments. Then, the difference between the time-aligned control command segment sequence and response data segment sequence is calculated to obtain an aligned residual sequence reflecting the deviation between the control behavior and the physical response. Each residual value in the aligned residual sequence corresponds to the degree of deviation between the control behavior and the actual response under the same temporal semantics, and can be expressed as amplitude difference, rate of change difference, or trend deviation.

[0051] In one implementation, the unified time window is divided into fixed-length segments, such as segmenting the control command stream and response data stream synchronously at millisecond intervals, so that different time segments have a uniform time scale. In another implementation, the unified time window is divided into dynamic-length segments, automatically adjusting the time segment length when the control command changes significantly or when the response data shows fluctuation peaks, in order to improve the time matching accuracy.

[0052] During feature extraction, statistical methods can be used to extract the average, maximum, minimum, and rate of change, while frequency domain analysis can be used to extract frequency components and fluctuation features. In distance calculation, different weights can be assigned to different types of features; for example, increasing the weights of temperature change features and action switching features can enhance time matching capabilities in specific scenarios.

[0053] During time alignment, a segment-by-segment matching method can be used to map time segments, or a global matching method can be used to achieve unified alignment of the entire sequence. When the device has a long running time and a large amount of data, a sliding window update method can be used, calculating only the newly added time segments, thereby reducing the consumption of computing resources.

[0054] This embodiment eliminates time misalignment caused by different sampling mechanisms and timing rhythms by performing dynamic time warping on the control command stream and response data stream, enabling an accurate correspondence between control behavior and physical response, and generating an aligned residual sequence that reflects the degree of deviation, thereby improving the consistency and comparability of data in subsequent analysis.

[0055] S30, perform feature decoupling processing on the aligned residual sequence, separate the state feature components, and generate structured sensing events based on the deviation of the state feature components from the historical normal process distribution.

[0056] In this embodiment, the aligned residual sequence reflects the deviation between control behavior and physical response. However, the sources of deviation in the aligned residual sequence often have mixed characteristics, and different types of state changes may overlap in the same residual sequence. Therefore, feature decoupling processing is required to separate state changes from different sources. Feature decoupling processing corresponds to splitting the mixed state changes into multiple state feature components with independent physical meanings, thereby enabling the subsequent identification process to distinguish different sources of anomalies.

[0057] In the implementation process, the aligned residual sequence is input into the feature mapping unit to perform multidimensional representation transformation on the residual data, so that the time dimension, amplitude dimension, and trend dimension can be preserved simultaneously. Subsequently, feature separation is performed on the multidimensional representation results, so that different change patterns form relatively independent feature expressions in different dimensions. For example, a continuously increasing but slowly changing residual can reflect the state of heat accumulation, a residual that fluctuates frequently in a short period of time can reflect the state of mechanical vibration, and a residual with strong randomness and no obvious periodicity can reflect the state of environmental disturbance.

[0058] State feature components correspond to different state representations obtained from the aligned residual sequence, and each state feature component describes a specific change in the operating state of the equipment. State feature components may include thermal hysteresis state feature components, mechanical vibration state feature components, environmental disturbance state feature components, and may also include pressure fluctuation state feature components, current drift state feature components, or motion instability state feature components. Different state feature components are independent of each other and can reflect different types of abnormal symptoms.

[0059] After obtaining the state characteristic components, it is necessary to determine whether the current state deviates from the historical normal process distribution. The historical normal process distribution corresponds to the statistical results of the state characteristics formed by the equipment during the stable operation phase. Its source can be long-term collected normal operation data, or it can be a collection of normal state samples formed under different materials and process parameters. The historical normal process distribution can include the mean range, fluctuation range, change trend range, and distribution density range, which are used to describe the state change law of the equipment under normal operating conditions.

[0060] Deviation results correspond to the degree of difference between the state characteristic components and the historical normal process distribution. When the state characteristic components exceed the range of the historical normal process distribution, it can be determined that the current state has a deviation. Deviation results can manifest as abnormal thermal hysteresis, abnormal vibration amplitude, abnormal fluctuation frequency, or abnormal trend. To improve the accuracy of judgment, different deviation judgment conditions can be set for different state characteristic components. For example, the duration and rise rate can be jointly judged for thermal hysteresis state characteristic components, and the peak amplitude and fluctuation frequency can be jointly judged for mechanical vibration state characteristic components.

[0061] Structured sensing events correspond to encapsulated state descriptions of deviations, transforming previously scattered state information into a unified representation. Structured sensing events can include event type, occurrence time, state feature component categories, deviation degree, and impact range. For example, when thermal hysteresis state feature components consistently exceed the historical normal process distribution range, a structured sensing event containing information such as thermal hysteresis anomaly, duration, and deviation magnitude can be generated. Through structured sensing events, complex state changes can be converted into a unified, directly processable representation.

[0062] In one implementation, the aligned residual sequence is decoupled using a dimensionality reduction mapping method, allowing different types of state changes to be projected onto different feature spaces. In another implementation, the aligned residual sequence is processed using frequency decomposition, using the low-frequency variation to represent the thermal hysteresis state and the high-frequency variation to represent the mechanical vibration state.

[0063] In establishing historical normal process distributions, statistical models can be formed based on long-term normal operation data, or corresponding normal state distributions can be established based on different material types, equipment models, and process parameter conditions. In the deviation determination process, a fixed threshold method can be used to judge the degree of deviation, or a dynamic threshold method can be used to automatically adjust the judgment criteria according to the fluctuation range of historical normal process distributions.

[0064] In the generation of structured events, either tabular fields or vectorized representation can be used. When collaboration with subsequent data processing modules is required, event level and event duration fields can be added to structured events to differentiate the severity of different abnormal states.

[0065] This embodiment can separate state changes from different sources into independent state feature components by performing feature decoupling processing on the aligned residual sequence, and generate structured sensing events based on the degree of deviation between the state feature components and the historical normal process distribution, thereby improving the accuracy of state recognition and the consistency of abnormal expression.

[0066] S40, map the structured perception event and the running context to entity nodes in the mechanism knowledge graph, and perform inverse probability inference along the causal edges in the mechanism knowledge graph to determine the target physical root cause;

[0067] In this embodiment, structured sensing events are used to characterize the abnormal status information that has been identified during equipment operation, while the operating context is used to characterize semantic information such as control behavior, processing area, and operating condition switching. Structured sensing events only describe the abnormal phenomena that have been observed, and the operating context only describes the current operating conditions. The two correspond to abnormal performance information and operating condition background information, respectively. When used alone, it is difficult to accurately locate the source of the abnormality. Therefore, it is necessary to jointly map structured sensing events and operating context to entity nodes in the mechanism knowledge graph, so that abnormal phenomena and operating conditions can be associated in a unified knowledge representation space.

[0068] Entity nodes in the mechanism knowledge graph are used to carry information about different types of objects during equipment operation. The sources of entity nodes include process parameter information, equipment component information, physical process information, abnormal behavior information, and quality defect information. Process parameter information can include heating temperature, feed rate, motion acceleration, and material conveying status; equipment component information can include nozzle areas, drive mechanisms, thermal management components, and actuators; physical process information can include heat accumulation, material flow instability, mechanical vibration transmission, and local overload; and abnormal behavior information can include current fluctuations, temperature drift, vibration enhancement, and response delay. Mapping structured sensing events to entity nodes in the mechanism knowledge graph essentially involves corresponding the anomaly categories, anomaly times, anomaly states, and anomaly degrees contained in the structured sensing events to the anomaly behavior entity nodes in the mechanism knowledge graph. Mapping the runtime context to entity nodes in the mechanism knowledge graph essentially involves corresponding the current processing segment, action type, and operating condition switching state contained in the runtime context to the operating condition entity nodes, process parameter entity nodes, or equipment status entity nodes in the mechanism knowledge graph.

[0069] The mapping process requires addressing the semantic alignment issue between event information and graph nodes. In implementation, structured events can first be vectorized, allowing anomaly categories, intensity, duration, and associated states to be expressed in a unified feature form. Then, the operational context is vectorized, forming contextual feature representations of condition information, operational information, and environmental information. Subsequently, based on vector similarity relationships, node attribute matching relationships, and graph relationship constraints, the structured events and operational contexts are associated with corresponding entity nodes in the mechanistic knowledge graph. After mapping, anomaly manifestations are no longer merely isolated observations but have entered into a knowledge representation structure organized by causal relationships, providing a basis for subsequent root cause identification.

[0070] In a mechanistic knowledge graph, causal edges represent the causal transmission relationships between different entity nodes. The direction of the causal edge reflects the direction of the interaction between cause and effect. For example, abnormal heat transfer can cause temperature shifts, which can further cause changes in material flow state, and changes in material flow state can further cause forming anomalies. Causal edges can also represent the relationship between changes in control parameters and changes in equipment response, the relationship between changes in component state and abnormal behavior, and the relationship between operating conditions and fault evolution. The existence of causal edges enables the graph to not only express the connections between nodes but also the order of influence and transmission direction during the abnormal evolution process.

[0071] Reverse probabilistic deduction corresponds to starting from the already mapped abnormal behavior entity nodes and operating condition entity nodes, and performing a backward analysis of potential causes along the causal edges in the mechanistic knowledge graph. The goal of this process is not to continue describing the abnormal behavior, but to deduce the set of potential causes from the abnormal result, and then identify the most probable cause from this set. Reverse probabilistic deduction needs to consider both the abnormal state represented by the structured perceived event and the operating conditions represented by the operational context, because the same abnormal behavior may correspond to different causes under different operating conditions. For example, temperature deviation may have different causes under low-speed and high-speed operating conditions, and increased vibration may also correspond to different causes under turning processing sections and straight processing sections.

[0072] In the implementation process, the mapped abnormal behavior entity nodes can be used as the observation starting point, and the upstream entity nodes with causal relationships to the observation starting point can be used as candidate cause entity nodes. Then, based on the propagation relationship of causal edges, the prior probability of the candidate cause entity nodes, and the constraints imposed on them by the current operating context, the posterior probability of each candidate cause entity node is calculated. The posterior probability reflects the likelihood that each candidate cause entity node will become the true cause given the abnormal behavior and operating conditions. The target physical root cause corresponds to the candidate cause entity node with the highest posterior probability. The target physical root cause can manifest as specific physical causes such as thermal hysteresis, unstable transport, amplified local vibration, abnormal material state, and unbalanced drive response. Through this approach, the anomaly identification results are elevated from the superficial level to the causal level, allowing subsequent processing to directly revolve around the target physical root cause.

[0073] In one implementation, structured sensing events are mapped to anomalous behavior entity nodes in the mechanism knowledge graph using vector encoding, and the runtime context is mapped to operating condition entity nodes in the mechanism knowledge graph using context encoding. Candidate causal entity nodes are then selected based on the common adjacency relationships between the anomalous behavior entity nodes and the operating condition entity nodes. In another implementation, structured sensing events and runtime context are first jointly encoded, and then matching weights with multiple entity nodes are calculated using graph attention. The entity node with the higher matching weight is used as the mapping result. For causal edges in the mechanism knowledge graph, a fixed set of causal relationships can be used, or the strength of the causal edges can be updated based on historical equipment operating data. For inverse probabilistic deduction, a deduction method based on prior probability and conditional probability can be used, or a probability update method based on graph structure propagation can be used. In scenarios with frequent changes in operating conditions, the weight of the runtime context in the mapping process can be increased, making the selection process of candidate causal entity nodes more closely aligned with the current operating state. In scenarios where abnormal behavior persists for a long period, structured perception events within multiple consecutive time windows can be merged, and then mapping and inverse probability deduction can be performed to improve the stability of target physical root cause identification.

[0074] This embodiment maps structured perception events and runtime contexts to entity nodes in a mechanistic knowledge graph, and performs inverse probability deduction along the causal edges in the mechanistic knowledge graph. This unifies abnormal phenomena and runtime conditions into a causal expression space, thereby improving the targeting and accuracy of abnormal cause identification and providing clear causal basis for determining the target physical root cause.

[0075] S50, construct physical safety boundary conditions based on material property boundaries and equipment motion boundaries, generate a candidate strategy set based on the target physical root cause, filter the candidate strategy set using the physical safety boundary conditions, and determine the target adjustment strategy from the filtered candidate strategy set through a value evaluation function;

[0076] In this embodiment, material property boundaries characterize the physical range that a material can withstand during processing, while equipment motion boundaries characterize the range of motion that the equipment actuator can achieve during operation. The sources of material property boundaries include the material's thermal decomposition threshold, melting range, thermal conductivity, viscosity variation range, and cooling shrinkage characteristics. The sources of equipment motion boundaries include upper limits on movement speed, acceleration, stroke range, drive power limits, and mechanical response capabilities. The material property boundaries and equipment motion boundaries together determine the safe range that is permissible during equipment operation, preventing subsequent adjustments from exceeding the equipment's capacity or the material's processing capabilities.

[0077] Physical safety boundary conditions correspond to a set of constraints formed based on material property boundaries and equipment motion boundaries. Physical safety boundary conditions include not only the upper and lower limits of individual parameters, but also the combined constraints between multiple parameters. For example, even if the heating temperature does not exceed the thermal decomposition threshold, simultaneous operation under high speed and high extrusion volume conditions may still lead to localized overheating or unstable material flow. Therefore, physical safety boundary conditions also need to reflect the synergistic constraints between different parameters. Physical safety boundary conditions can be expressed as parameter ranges, parameter combination intervals, rate of change limits, and continuous operation limits, used to define the range for generating subsequent candidate strategy sets.

[0078] Target physical causes correspond to identified sources of anomalies and essentially reflect the most pressing need for equipment adjustment. Different target physical causes correspond to different types of adjustment directions. For example, when the target physical cause is thermal hysteresis, the adjustment direction can vary around heating intensity, cooling rate, and movement rhythm; when the target physical cause is increased material flow resistance, the adjustment direction can vary around extrusion speed, nozzle temperature, and movement path; when the target physical cause is increased local vibration, the adjustment direction can vary around acceleration, path turning amplitude, and movement smoothness.

[0079] The candidate strategy set corresponds to multiple alternative adjustment combinations generated around the target physical root cause. Each candidate strategy in the set contains multiple adjustable parameters, and the differences between different candidate strategies mainly lie in the parameter values ​​and the way they are combined. Adjustable parameters may include nozzle temperature, platform temperature, extrusion speed, motion speed, acceleration, cooling airflow, and path spacing, among others. The generation process of the candidate strategy set involves combining multiple sets of values ​​for the adjustable parameters in the adjustment direction corresponding to the target physical root cause, thereby forming multiple differential adjustment results.

[0080] After the candidate strategy set is generated, it needs to be screened using physical safety boundary conditions. The screening process involves boundary verification of each candidate strategy in the set to determine if it exceeds material property boundaries or equipment motion boundaries. For example, a candidate strategy that reduces thermal hysteresis might cause the nozzle temperature to exceed the material's thermal decomposition threshold or cause the motion speed to exceed the equipment's allowable range; in this case, the candidate strategy cannot be retained. Through this screening process, candidate strategies with physical or equipment risks can be eliminated, retaining only those that meet the physical safety boundary conditions.

[0081] The value evaluation function is used to comprehensively evaluate the shortlisted candidate strategies. The input to the value evaluation function comes from multiple evaluation indicators corresponding to the candidate strategies, including changes in geometric accuracy, processing time, energy consumption, stability, and risk of exceeding limits. Different evaluation indicators reflect the degree of influence of the candidate strategies in different dimensions. For example, one candidate strategy may have high processing accuracy but increase running time and energy consumption; another candidate strategy may have low energy consumption but reduce forming stability. The value evaluation function obtains the comprehensive evaluation result for each candidate strategy by weighting and combining different evaluation indicators.

[0082] The target adjustment strategy corresponds to the candidate strategy that yields the best comprehensive evaluation results. This strategy must not only meet physical safety boundary conditions but also achieve a balance between accuracy, efficiency, stability, and energy consumption. Determining the target adjustment strategy provides specific parameter combinations for subsequent execution processes, enabling the equipment's operating state to be adjusted towards greater stability and safety.

[0083] In one implementation, material property boundaries are determined based on a material database, with different materials corresponding to different thermal decomposition thresholds, melting ranges, and flow characteristics. Equipment motion boundaries are determined based on equipment specifications, with different equipment corresponding to different speed limits, acceleration limits, and motion ranges. In another implementation, material property boundaries and equipment motion boundaries can be dynamically adjusted based on historical operating data, making the physical safety boundary conditions more closely match actual operating conditions.

[0084] During the generation of the candidate strategy set, a fixed-step approach can be used to combine the parameters to be adjusted, or an adaptive search approach can be adopted to apply more intensive value changes to key parameters based on the target physical root causes. In the value assessment process, the weight of geometric accuracy evaluation indicators can be increased to adapt to high-precision machining scenarios; the weight of machining time and energy consumption evaluation indicators can also be increased to adapt to batch production scenarios. In scenarios where equipment operating conditions change rapidly, the size of the candidate strategy set can be reduced to decrease computation time; in scenarios where equipment operating conditions are relatively stable, the size of the candidate strategy set can be expanded to increase the selection space of the target adjustment strategy.

[0085] This embodiment constructs physical safety boundary conditions based on material property boundaries and equipment motion boundaries, generates a set of candidate strategies around the target physical root causes, and then uses physical safety boundary conditions for screening and value evaluation functions for comprehensive evaluation. This can avoid the adjustment results from exceeding the allowable range of equipment and materials, while improving the overall rationality of the target adjustment strategy in terms of accuracy, efficiency and stability.

[0086] S60, Perform physical feasibility pre-simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility pre-simulation and thought chain verification into low-level control instructions and write them into the device control terminal;

[0087] In this embodiment, the target adjustment strategy corresponds to the parameter combination result after screening and comprehensive evaluation. However, the target adjustment strategy is not directly equivalent to the executable result, because the target adjustment strategy still belongs to the high-level parameter expression oriented towards adjustment logic, and its actual executability during equipment operation needs to be further verified. The physical feasibility pre-simulation corresponds to predicting the equipment state changes that the target adjustment strategy may cause before it is actually written into the equipment control terminal, in order to determine whether the target adjustment strategy will lead to problems such as local overheating, motion instability, path conflict, response lag, or processing interruption.

[0088] Physical feasibility simulation requires prediction based on the current operating status of the equipment, the state of the materials, and the status of the control parameters. Changes in the parameters of the target adjustment strategy will affect the thermal, mechanical, and material responses of the equipment. Therefore, during the simulation, it is necessary to simulate changes in nozzle temperature, movement speed, extrusion speed, cooling intensity, and movement trajectory according to the target adjustment strategy. The simulation results can be represented as an equipment evolution response trajectory, which describes the process of equipment operating status changing over time under the action of the target adjustment strategy. The equipment evolution response trajectory can include trends in temperature change, vibration change, current change, displacement change, and material deposition change.

[0089] The thought chain verification corresponds to checking the logical consistency and causal rationality of the target adjustment strategy during execution. Multiple parameters in the target adjustment strategy are interdependent, therefore it is necessary to determine whether parameter changes conform to the equipment's operating logic and processing logic. For example, when the target adjustment strategy includes reducing the nozzle temperature, it is necessary to determine whether the extrusion speed decreases synchronously; when the target adjustment strategy includes increasing the movement speed, it is necessary to determine whether the acceleration and path turning degree are adjusted synchronously. If the parameter changes lack reasonable coordination, even if the target adjustment strategy itself does not exceed the physical safety boundary, new anomalies may occur during execution.

[0090] During the thought chain verification process, it is necessary to analyze the parameter change sequence, parameter dependencies, and the rationality of parameter combinations in the target adjustment strategy. The parameter change sequence is used to determine whether each adjustment action is arranged according to the correct execution order; parameter dependencies are used to determine whether a change in one parameter requires other parameters to change simultaneously; and the rationality of parameter combinations is used to determine whether simultaneous changes in multiple parameters will cause new risks. For example, when the nozzle temperature decreases while the movement speed remains constant, it may lead to insufficient material melting; when the movement speed increases while the cooling airflow remains constant, it may lead to increased local heat accumulation. The result of the thought chain verification can be expressed as logical pass or logical fail. Execution is only permitted if the target adjustment strategy simultaneously meets the requirements of physical feasibility simulation and thought chain verification.

[0091] Low-level control commands correspond to control codes that the equipment control terminal can directly recognize and execute. Parameter changes in the target adjustment strategy are typically expressed using high-level semantics, such as increasing nozzle temperature, decreasing movement speed, or adjusting cooling intensity. However, the equipment control terminal can only recognize specific control fields and values; therefore, the target adjustment strategy needs to be converted into low-level control commands. This conversion process involves mapping parameter changes in the target adjustment strategy to specific temperature setpoints, speed setpoints, movement path commands, and execution action commands, based on the equipment control protocol and communication format.

[0092] Writing to the device control terminal involves transmitting the converted low-level control commands to the device's internal control unit, enabling the device to continue operating according to the new parameter configuration. During the writing process, it's crucial to ensure consistency between the low-level control commands and the current device state to prevent new commands from being written prematurely before old commands have been completed. To guarantee writing stability, the low-level control commands can be buffered and sorted based on the current control cycle, current motion state, and current processing segment, ensuring the device control terminal executes the updated control content in the correct order.

[0093] In one implementation, physical feasibility pre-simulation is conducted in a virtual simulation environment to predict temperature, motion, and current changes after the target adjustment strategy is executed, and the prediction results are used to determine whether there is a risk of exceeding the allowable range. In another implementation, physical feasibility pre-simulation compares historical operating data, matching the target adjustment strategy with the operating results of similar historical scenarios, and determining its feasibility based on the historical results.

[0094] During the thought chain verification process, rule matching can be used to check parameter dependencies, while logical deduction can be used to analyze the rationality of parameter combinations. During device control protocol adaptation, different underlying control field mapping rules can be configured for different device types, enabling the target adjustment strategy to be recognized by different device control terminals. During the writing process to the device control terminal, a synchronous writing method can be used to immediately update the underlying control commands, or a buffered writing method can be used to execute command switching after the current control cycle ends.

[0095] This embodiment can identify potential risks and logical conflicts during the execution of the target adjustment strategy by performing physical feasibility simulation and thought chain verification. The verified target adjustment strategy is then converted into low-level control instructions and written into the device control terminal, thereby improving the executability of the target adjustment strategy and the stability of device operation.

[0096] S70, acquire the new response data stream generated by executing the underlying control command, and use the new response data stream as a response data stream for subsequent acquisition of control command stream and response data stream.

[0097] In this embodiment, the newly added response data stream corresponds to a continuous set of data generated by changes in the device's operating state after the underlying control command is executed. This data set reflects the actual operating results of the device under the new parameter configuration, including temperature change information, motion state information, execution state information, and material change information. The sources of the newly added response data stream include data output from the device's internal sensing units and control execution feedback data, such as real-time temperature sequences output by temperature sensors, displacement sequences output by position encoders, current sequences output by motor drive units, and execution state flag change information. These data are continuously collected in time series form and organized according to timestamps, enabling data from different sources to correspond in the time dimension.

[0098] The data acquisition process involves continuously reading feedback information from various sensing and execution units within the device during the execution of low-level control commands. Acquisition can be achieved through the data acquisition interface built into the device's control terminal, or by reading feedback data from the device's communication bus using an external acquisition module. During acquisition, different types of data need to be processed to unify the time base, aligning temperature, displacement, and current data on the same time scale to form a complete new response data stream. To avoid data loss or conflicts during acquisition, data can be cached and its sequence verified to ensure the continuity and consistency of the new response data stream.

[0099] The newly added response data stream is used for subsequent processing. This corresponds to directly replacing or updating the input data for the next round of processing with the data generated in the current execution phase. The update method for the response data stream can be either to completely overwrite the original response data stream with the new one, or to append the new response data stream to the end of the original response data stream to form an extended data sequence. In the overwrite method, subsequent processing is based solely on the data from the current execution phase; in the append method, subsequent processing can utilize both historical and current data for comprehensive analysis. The choice between the different update methods depends on the degree to which subsequent processing relies on historical information.

[0100] By using newly added response data streams as the main response data stream, a continuously updated data input structure can be formed, enabling subsequent acquisition of control command streams and response data streams to be processed based on the latest device status. This update process allows for continuous perception of the device's operating status and enables the feedback of actual operating results to the next round of processing after each execution, thereby achieving data-driven dynamic adjustment.

[0101] In one implementation, new response data streams are periodically acquired through the device control terminal. The acquisition period is set according to the device's operating frequency to ensure that the data sampling frequency is consistent with the device control frequency. In another implementation, data acquisition can be triggered by events. Data acquisition is initiated when the execution of a low-level control command is detected or when key parameter changes reach set conditions, thereby reducing data redundancy.

[0102] During the update process of the response data stream, a sliding window approach can be used to retain the data sequence over a recent period, enabling subsequent processing to capture both the current state and short-term trends. In scenarios with high time alignment requirements, a unified timestamp generation mechanism can be used to synchronize data from different sources, improving the accuracy of data correspondence. In scenarios with large datasets, newly added response data streams can be compressed or have features extracted to reduce the computational burden on subsequent processing.

[0103] This embodiment collects the new response data stream after the execution of the underlying control command and uses it to update the response data stream, so that subsequent processing can be based on the latest equipment operating status, thereby improving data timeliness and enhancing the adaptability of the subsequent adjustment process to the actual operating conditions.

[0104] In one embodiment, step S10 includes:

[0105] S101, establish an instruction communication connection between the device control terminal and the edge computing terminal, and establish a data communication connection between the sensor acquisition terminal and the edge computing terminal;

[0106] S102, continuously read control commands from the device control terminal and buffer them according to the output order to form a control command stream;

[0107] S103, the control command stream is split and parsed to obtain each control command containing a motion control field, an energy control field, and a layer switching control field, and a corresponding time identifier is written for each control command;

[0108] S104 reads temperature sampling values, vibration sampling values, and current sampling values ​​from the sensor acquisition terminal, and aggregates them according to the sampling time sequence to form a response data stream;

[0109] S105, extract the current processing segment identifier, action type identifier, and working condition switching identifier from the motion control field, energy control field, and layer switching control field in each control instruction with the corresponding time identifier, and combine the current processing segment identifier, the action type identifier, and the working condition switching identifier to form a running context.

[0110] In this embodiment, acquiring the control command stream corresponds to the time-sequential reception and organization of control information continuously output by the device control terminal during operation. The device control terminal undertakes the output tasks of control content such as motion execution, energy regulation, and level switching. The output format can be a continuous coded sequence or discrete control messages. To ensure stable reception of control information, an instruction communication connection needs to be established between the device control terminal and the edge computing terminal. The instruction communication connection is used to transmit control information, maintain communication status, and ensure data sequence consistency. In implementation, a serial interface, bus interface, or network communication interface corresponding to the device control terminal can be configured in the edge computing terminal, and the interface parameters can be uniformly configured so that the edge computing terminal can continuously receive control information according to the output rhythm of the device control terminal. After establishing the communication connection, the edge computing terminal buffers the continuously arriving control information. The buffering operation corresponds to temporarily storing and queuing the control information according to the output order, avoiding instruction omissions, overlaps, or out-of-order processing due to the inconsistency between the output speed of the control terminal and the parsing speed of the processing terminal. The buffered data forms a control command stream according to the time sequence. The control command flow is not a simple accumulation of data, but a control sequence with sequential, temporal, and execution attributes. The position of each control information in the sequence is related to the current motion and processing state of the equipment.

[0111] The response data stream corresponds to the feedback sequence formed after synchronously observing the equipment's operating results. The response data does not originate from the control terminal itself, but rather from the sensor acquisition terminal. The sensor acquisition terminal is responsible for outputting observational data related to the real-time status of the equipment. Temperature sampling values ​​reflect changes in the thermal state of heated areas, deposition areas, or component surfaces; vibration sampling values ​​reflect mechanical disturbances in actuators, transmission structures, or localized processing areas; and current sampling values ​​reflect load changes in drive components during execution. To enable these feedback results to enter a unified processing flow, a data communication connection needs to be established between the sensor acquisition terminal and the edge computing terminal. This data communication connection handles sensor data uploading, sampling synchronization, and feedback transmission. After establishing the data communication connection, the edge computing terminal receives and aggregates the temperature, vibration, and current sampling values ​​according to their sampling time sequence. The aggregation process corresponds to organizing data from different sensors into the same time series according to a unified time reference, allowing different physical quantities to form corresponding feedback expressions at the same time scale. The aggregated feedback results form the response data stream. The response data stream reflects the actual operating state of the equipment under control commands, and subsequent processing relies on this data stream to determine whether there is a deviation between the control results and the physical response.

[0112] The splitting and parsing of the control command stream corresponds to the semantic decomposition of the continuous control sequence. Continuous control information in its raw state is typically represented as a combination of multiple fields; without splitting, it is difficult to extract the effective information related to the operating state. The splitting operation corresponds to dividing the control command stream into individual control commands according to control command boundaries. Each control command contains motion control fields, energy control fields, and layer-switching control fields. The motion control field characterizes actions such as changes in device position, path, speed, or acceleration. The energy control field characterizes control content related to energy input or output, such as nozzle temperature, platform temperature, drive power, and cooling intensity. The layer-switching control field characterizes stage-specific changes such as layer switching, processing stage switching, or operating mode switching. The parsing operation corresponds to identifying, reading, and classifying the above fields, enabling the edge computing terminal to obtain control information related to motion behavior, energy behavior, and layer behavior. To establish a correspondence between subsequent control information and response information, a corresponding time identifier needs to be written for each control command. The corresponding time stamp is used to characterize the time when a control command is output or takes effect, enabling the control behavior to be correlated with the response result in the time dimension. The corresponding time stamp can be generated from the original output time of the device control terminal, the receiving time of the edge computing terminal, or a calibrated unified time base. After the time stamp is written, each control command retains not only the semantic field content, but also the time attributes that can be used for timing correlation.

[0113] The extraction of operational context corresponds to extracting descriptive information reflecting the current processing state and control semantics from control commands. The current processing segment identifier characterizes the processing area, path interval, or geometric processing interval the equipment is currently in. The action type identifier characterizes whether the equipment is currently performing extrusion, movement, pause, heating adjustment, speed adjustment, or other control actions. The operating condition switching identifier characterizes whether the equipment's operating state has undergone a phase change, such as switching from continuous deposition to a transitional state, or from low-power operation to high-load operation. The extraction process is not a simple copying of field content, but rather a summarization and transformation based on the motion control, energy control, and layer-switching control fields in each control command with corresponding time identifiers, converting discrete control information into a state description with operational semantics. The resulting current processing segment identifier, action type identifier, and operating condition switching identifier represent the processing position, action nature, and state change, respectively. The combination process corresponds to associating these multiple identifiers under the same temporal semantics to form the operational context. The operational context is not a single identifier, but a combination of information reflecting position, action, and state simultaneously. This combined result can describe the region the equipment is in at a certain moment or within a short time range, what actions it is performing, and whether a state transition has occurred. The runtime context plays a role in constraint and interpretation in subsequent processing, enabling operational information that cannot be expressed by control sequences alone to enter the analysis process in a structured form.

[0114] This embodiment establishes command communication connections between the device control terminal and the edge computing terminal, as well as data communication connections between the sensor acquisition terminal and the edge computing terminal, enabling the integration of control and feedback information into a unified processing environment. By caching, splitting, and parsing continuous control information, a control command stream with sequential and temporal attributes can be formed, resulting in individual control commands containing motion control, energy control, and layer-switching control fields. By aggregating temperature, vibration, and current sampling values ​​in a temporal sequence, a response data stream reflecting the actual operating status of the device can be generated. By extracting the current processing segment identifier, action type identifier, and operating condition switching identifier from each control command with corresponding time identifiers and combining them to form an operating context, the control information can acquire operating condition semantics in addition to its temporal attributes. Therefore, control input, execution feedback, and operating semantics can be expressed simultaneously on a unified temporal basis, improving the completeness of subsequent processing in identifying the device's operating status and the accuracy of data correspondence.

[0115] In one embodiment, prior to step S10, the method further includes:

[0116] S1001, constructing a thermal-fluid coupled digital twin environment;

[0117] S1002 introduces friction coefficient, thermal conductivity, density and ambient temperature disturbances into the digital twin environment to form multiple sets of unsteady virtual working conditions;

[0118] S1003, a cloud-based high-dimensional time series model trained based on multiple sets of unsteady virtual working conditions;

[0119] S1004, the trained cloud-based high-dimensional time series model is identified as the teacher model.

[0120] In this embodiment, a digital simulation training stage is set up before acquiring the control command stream and response data stream. The purpose is to first develop a high-dimensional temporal cognitive ability capable of covering complex operating conditions, and then apply this ability to subsequent edge-side processing. Acquiring the control command stream and response data stream involves the actual operation of the equipment. In real-world processes, abnormal samples are sparsely distributed, operating conditions change continuously, and parameters are significantly coupled. Relying solely on real-time field data for model learning can easily lead to problems such as insufficient sample coverage, excessively high training costs at the edge, and insufficient model recognition capabilities for unsteady states. Placing the training stage before acquiring the control command stream and response data stream allows the model, upon which subsequent processing depends, to complete parameter learning through virtual operating conditions before entering the field operation phase. This reduces the computational burden on the edge and improves the adaptability of the subsequent recognition process to complex operating conditions.

[0121] A thermal-fluid coupling digital twin environment is used to characterize the coupling relationship between heat transfer, material flow, and structural motion during equipment operation. In thermal-fluid coupling, "heat" corresponds to changes in heating source input, temperature conduction, convective heat dissipation, and heat accumulation; "flow" corresponds to the flow state of materials in nozzles, channels, or processing areas, as well as the flow regime changes caused by viscosity variations after heating. The digital twin environment is not a static geometric model, but a computable virtual environment built around the equipment structure, material properties, control parameters, and environmental conditions. During construction, it is necessary to import the equipment geometric model, heat source location distribution, material thermal properties, boundary heat dissipation conditions, and motion control parameters, enabling the virtual environment to simultaneously reflect temperature field changes and material flow changes over time. The geometric model can be derived from equipment design parameters, actuator dimensions, and processing area boundary information. Material thermal properties can be derived from a material database, including thermal conductivity, specific heat capacity, density, melting temperature range, and viscosity variation relationships. Environmental conditions can be derived from workshop ambient temperature, airflow conditions, and heat dissipation methods. Through these inputs, the digital simulation environment can calculate and express the thermal and flow responses under different operating conditions before the equipment enters actual operation.

[0122] Introducing perturbations of friction coefficient, thermal conductivity, density, and ambient temperature into a digital twin environment corresponds to applying random or range-based variations to existing virtual environment parameters, preventing the virtual operating conditions from remaining in a single steady-state state. The friction coefficient reflects the resistance variation between contact interfaces, originating from the contact characteristics between materials and components, the degree of wear of moving parts, or differences in material surface conditions. Thermal conductivity reflects the ability of materials and structures to diffuse heat, originating from differences in the material's own thermal properties, changes in thermal conductivity caused by local temperature variations, or changes in the heat conduction path caused by changes in deposition state. Density reflects the mass distribution per unit volume of the material, originating from batch differences in materials, changes in gas content, thermal expansion and contraction effects, or changes in doping ratios. Ambient temperature reflects the external thermal boundary conditions of the equipment, originating from fluctuations in workshop temperature, changes in cooling air conditions, and changes in the heat dissipation state around the equipment. The introduction of perturbations is not a single parameter replacement, but rather applying variations of different magnitudes and combinations to these parameters across multiple simulation rounds, causing the heat transfer and flow processes in the virtual environment to exhibit non-repetitive and non-constant variations. The operating conditions formed in this way can cover the deviation of motion resistance caused by frictional changes, the temperature response difference caused by changes in thermal conductivity, the flow instability caused by density changes, and the changes in heat dissipation conditions caused by fluctuations in ambient temperature.

[0123] Multiple sets of unsteady-state virtual operating conditions correspond to a series of processing states that are continuous in time but exhibit non-constant state changes, generated under different combinations of perturbations. Unsteadiness is manifested in the continuous change of temperature distribution over time, the alteration of material flow parameters with both control input and environmental conditions, and the phased fluctuations in equipment operating conditions under the influence of multiple parameters. Multiple sets do not represent repeated sampling of a single sample, but rather the formation of multiple differentiated state sequences through multiple combinations of perturbation parameters. During the formation process, different values ​​or different change trajectories can be set for the friction coefficient, thermal conductivity, density, and ambient temperature in each simulation, combined with the control input to form complete time-series operating condition data. The resulting data includes various time-series information such as temperature field distribution sequences, flow state sequences, velocity change sequences, heat accumulation change sequences, and load change sequences. The significance of multiple sets of unsteady-state virtual operating conditions lies in expanding the coverage of training samples, ensuring that the training data is not limited to ideal states or a small number of typical operating conditions, but can cover operational changes under the combined effects of material differences, environmental differences, and equipment state differences.

[0124] A cloud-based high-dimensional time-series model corresponds to a high-dimensional representation model formed after learning from multiple sets of unsteady virtual operating conditions in an environment with sufficient computing resources. "Cloud" refers to the deployment location in a high-computing environment, used to undertake large-scale time-series sample training tasks. "High-dimensional" means that the model input and intermediate representation simultaneously contain multiple physical quantities, multiple time steps, and multiple parameter correlation information, not limited to a single signal or a single window. A time-series model corresponds to a model structure that takes continuous time changes as input and learns about state evolution, parameter coupling, and abnormal trends. The training process uses multiple sets of unsteady virtual operating conditions as input, enabling the model to learn the temporal relationships between heat transfer changes, flow changes, and control response changes under different combinations of parameter perturbations. During training, the model can receive temperature field time series, flow state time series, control input sequences, and environmental perturbation sequences, extracting high-dimensional correlation patterns through parameter updates. The resulting cloud-based high-dimensional time-series model can not only represent local physical quantity changes but also the coupled evolution characteristics of multiple physical quantities in the time dimension. After training, the model has the ability to represent, predict, or extract features from complex operating conditions.

[0125] The trained high-dimensional temporal model in the cloud is designated as the teacher model, corresponding to its guiding role in subsequent transfer learning and knowledge distillation. The teacher model does not directly undertake the final edge operation task, but serves as a high-capacity, high-representational-capability reference model, providing target features and learning constraints for subsequent lightweight models. The teacher model is formed only after sufficient training under multiple sets of non-steady-state virtual operating conditions; therefore, it retains temporal representation capabilities under multiple physical quantities, multiple time scales, and multiple operating condition perturbations. Designating the trained high-dimensional temporal model in the cloud as the teacher model ensures that subsequent lightweight model training does not start from scratch, but rather uses the high-dimensional representation provided by the teacher model for compressed transfer, reducing the problem of insufficient expressive power in edge models under insufficient sample conditions. Positioning the teacher model before acquiring control command and response data streams ensures that subsequent edge processing modules have a high-level representational reference before entering real device operation, thereby improving model adaptation efficiency during the field deployment phase.

[0126] For example, the formula for the physical parameter perturbation vector:

[0127]

[0128] in, This represents the physical parameter perturbation vector. This represents the coefficient of friction. This indicates thermal conductivity. Indicates density. This represents the ambient temperature. The superscript T indicates transpose, used to organize multiple disturbance parameters into a column vector.

[0129] The formula for random sampling of unsteady virtual operating conditions is:

[0130]

[0131] in, This represents the set of random physical parameters sampled in the t-th evolution iteration. This represents the probability distribution of physical parameters. A set of parameters representing a probability distribution. (Symbol) It indicates that it follows or is sampled from a certain probability distribution.

[0132] This embodiment constructs a heat-fluid coupled digital twin environment and introduces friction coefficient, thermal conductivity, density, and ambient temperature disturbances into the digital environment to form non-steady-state virtual operating conditions covering various operational differences. By training a cloud-based high-dimensional time series model based on multiple sets of non-steady-state virtual operating conditions and determining the trained cloud-based high-dimensional time series model as the teacher model, the model on which subsequent processing depends can obtain a more comprehensive ability to represent complex operating conditions before entering the real equipment operation stage. This improves the adaptability of subsequent edge-side processing to changes in non-steady-state conditions and reduces the pressure of on-site training.

[0133] In one embodiment, prior to step S10, the method further includes:

[0134] S1005 uses a teacher model to receive all physical field data and generate teacher characteristics;

[0135] S1006: Receives sensor time-series data using a student model and generates student characteristics;

[0136] S1007, based on teacher and student characteristics, performs knowledge distillation training to obtain a lightweight edge model;

[0137] S1008 performs quantization compression on the lightweight edge model and deploys the quantized and compressed lightweight edge model to the edge terminal.

[0138] In this embodiment, the teacher model receives the full physics field data and generates teacher features. The full physics field data corresponds to multidimensional continuous data that can fully characterize heat transfer, material flow, mechanical motion, and load changes during equipment operation. The sources of the full physics field data can include temperature field distribution, velocity field distribution, pressure field distribution, displacement field distribution, and energy input distribution output from a digital simulation environment, or state evolution results obtained after multiphysics solution. The teacher model receives the full physics field data to extract the implicit global correlation information in the high-dimensional state evolution process, enabling the teacher features to simultaneously express the coupling changes between multiple physical quantities, the continuity in the time dimension, and the state transfer relationships between different regions. The teacher features are not a simple compression of the original data, but a high-dimensional representation result formed after deep encoding of the full physics field data. This representation result can reflect the synchronous changes, lag changes, and mutual constraints between different physical quantities under complex operating conditions, providing a high-information-density target representation for subsequent lightweight model learning.

[0139] The student model receives sensor time-series data and generates student features. This sensor time-series data corresponds to the actual operational data that the edge terminal can directly collect. The data includes time series outputs from temperature, vibration, and current sensors, as well as other state sampling sequences related to equipment operation. Compared to full physical field data, sensor time-series data has fewer observation dimensions, more discrete sampling point distribution, and limited information coverage. The student model receives this sensor time-series data with the aim of extracting student features that characterize the current operational state from limited sensing conditions. These student features need to approximate the high-level state relationships expressed by the teacher features as closely as possible, given the limited information dimensions. Therefore, the student model must not only maintain local sensing capabilities for temperature, vibration, and current changes but also recover the implicit correlations between multiple observations in the feature space. The construction results of the student features directly determine the edge terminal's ability to represent complex operating conditions in subsequent processing.

[0140] Knowledge distillation training involves using the correspondence between teacher and student features to supervise the updating of the student model, enabling it to achieve representational capabilities close to the teacher model when receiving time-series sensor data. Knowledge distillation training is not a simple replication of the teacher model's parameters, but rather an optimization objective for the student model by comparing the distribution, temporal, and structural differences between teacher and student features in the feature space. Teacher features provide a high-dimensional state representation under full physics data, while student features provide a local state representation under sensor time-series data. The difference between the two reflects the deviation between the student model's current expressive capability and the target expressive capability. Knowledge distillation training adjusts the student model parameters to gradually approach the teacher features while maintaining the edge-side input format. The result is a lightweight edge model. A lightweight edge model corresponds to a model form where parameter size, computational complexity, and storage overhead are all controlled, while still retaining sufficient state representation capability. The formation of a lightweight edge model means that the temporal representation capability trained in a high-computing environment is compressed into a model structure suitable for edge terminal operation.

[0141] Quantization compression involves further reducing the parameter representation and numerical storage methods of lightweight edge models to decrease the computational and storage resource requirements of edge terminals during model deployment. During quantization compression, high-precision parameters can be mapped to low-precision parameters, and the weight matrix, activation value range, and intermediate calculation results can be discretized. The goal of quantization compression is to reduce model storage size, lower computational bandwidth requirements, and shorten inference latency while preserving the representational capabilities of lightweight edge models as much as possible. Quantized and compressed lightweight edge models are more suitable for the operating conditions of embedded processors, edge computing chips, or resource-constrained control terminals.

[0142] Deploying to edge terminals involves writing the quantized and compressed lightweight edge model into the edge computing platform, enabling the edge terminal to directly call this model for real-time processing during subsequent operation. The edge terminal is the computing carrier on the field side, used to receive data output from the device control terminal and data uploaded from the sensor acquisition terminal, and to perform processing tasks such as status analysis, feature extraction, or anomaly identification locally. Deploying the quantized and compressed lightweight edge model to the edge terminal requires not only that the model file be correctly loaded on the edge terminal, but also that the model execution flow be consistent with the edge terminal's data acquisition cycle time, cache management mechanism, and real-time task scheduling mechanism. After deployment, the edge terminal has the ability to run the compressed model under limited computing resources, thus freeing subsequent data processing from continuous dependence on high-computing-power environments.

[0143] There are clear dependencies between the teacher model, teacher features, student model, student features, knowledge distillation training, lightweight edge model, quantization compression, and edge terminal. The teacher model corresponds to a high-capacity representation source, and teacher features are the output form of high-dimensional state information. The student model corresponds to the edge-side fitting object, and student features are the state expression results under edge input conditions. Knowledge distillation training, through the correspondence constraints between teacher features and student features, enables the student model to acquire high-level representation capabilities. The lightweight edge model is the model result after knowledge distillation training, and quantization compression further reduces the resource consumption of the lightweight edge model. The edge terminal serves as the final execution location, receiving the compressed model for execution. The entire process corresponds to transferring high-dimensional temporal knowledge from the full physical field data to an edge model driven by sensor temporal data, enabling the edge side to maintain strong state expression capabilities even under limited resource conditions.

[0144] For example, the formula for the total loss of knowledge distillation between teachers and students is:

[0145]

[0146] in, This represents the total loss during knowledge distillation training. This represents the weight of the distributed distillation loss. This represents the Kullback-Leibler divergence between the output distribution of the teacher model and the output distribution of the student model. This represents the output distribution of the teacher model. This represents the distribution of the student model output. This represents the feature alignment loss weights. Indicates the characteristics of a teacher. Indicates student characteristics. This represents the result after performing dimension adaptation mapping on student features. This represents the squared difference between teacher characteristics and mapped student characteristics.

[0147] This indicates a loss of mission objectives.

[0148] This embodiment utilizes a teacher model to receive full physical field data to generate teacher features and a student model to receive sensor time-series data to generate student features, thereby establishing a correspondence between high-dimensional global state representation and edge-side local observation representation. By performing knowledge distillation training based on teacher and student features, the student model can obtain state representation capabilities close to those of a high-dimensional model under limited input conditions, thus forming a lightweight edge model. By performing quantization compression on the lightweight edge model and deploying it to the edge terminal, the computational and storage resources required for model operation can be reduced, enabling the edge terminal to have strong real-time processing capabilities without relying on continuous participation from a high-computing environment.

[0149] In one embodiment, step S20 above includes:

[0150] S201, the control command stream and the response data stream are segmented according to a unified time window to obtain a set of control command time segments and a set of response data time segments;

[0151] S202, extract action switching features and setpoint change features from the control command time segment set to form a control command feature sequence, and extract response fluctuation features and response peak features from the response data time segment set to form a response data feature sequence;

[0152] S203, calculate a distance matrix based on the control command feature sequence and the response data feature sequence, and calculate a cumulative distance matrix based on the distance matrix;

[0153] S204, determine the mapping path corresponding to the minimum cumulative distance based on the cumulative distance matrix;

[0154] S205, according to the mapping path, the control instruction time segment set and the response data time segment set are time-aligned to obtain an aligned control instruction segment sequence and an aligned response data segment sequence;

[0155] S206, calculate the timing difference based on the alignment control instruction fragment sequence and the alignment response data fragment sequence, and generate the alignment residual sequence.

[0156] In this embodiment, the control command stream and response data stream are segmented according to a unified time window. The purpose is to convert the original continuous time-series data into a set of time segments that can be locally compared and matched. The control command stream records the control behavior output by the device control terminal, including information such as changes in action, changes in setpoints, and changes in execution rhythm. The response data stream records the physical responses returned by the sensor acquisition terminal, including temperature changes, vibration changes, current changes, and other feedback information related to the execution state. Since the sampling methods, update times, and change rhythms of the control command stream and response data stream are inconsistent, direct comparison at the original sequence level can easily lead to time offset and semantic misalignment. Therefore, segmentation processing needs to be completed under a unified time scale first. The unified time window corresponds to using the same time length as the segmentation benchmark for both types of time-series data, so that the local control behavior in the control command stream and the local feedback changes in the response data stream are expressed correspondingly within the same time range. The length of the unified time window can be set according to the control cycle, sampling cycle, task cycle time, and device response delay. The control command time segment set corresponds to multiple control command segment sequences obtained by dividing the data within a unified time window, and the response data time segment set corresponds to multiple response data segment sequences obtained by dividing the data within the same unified time window. The formation process of the time segment set allows subsequent processing to establish a correlation based on changes within local time periods, rather than performing a rough match directly across the entire long time series data.

[0157] Extracting action switching features and setpoint change features from the set of control command time segments forms a control command feature sequence, corresponding to converting the control semantics and control intensity in the control command time segments into computable feature expressions. Action switching features reflect the transitions in control behavior, such as switching between movement and deposition commands, linear and turning movements, and heating / regulation and maintenance states. The sources of action switching features are various action codes, control mode identifiers, and execution state identifiers in the control commands. Setpoint change features reflect the trend and magnitude of changes in control parameters within a time segment, such as increases or decreases in speed setpoints, adjustments to temperature setpoints, and changes in power setpoints. The sources of setpoint change features are parameter fields, power fields, and stage control fields in the control commands. After extracting the action switching features and setpoint change features, the control behavior in a time segment can be condensed into finite-dimensional feature results, which are then arranged chronologically to form a control command feature sequence. The control command feature sequence preserves the sequential relationship between time segments while retaining the structural differences in control behavior within each time segment.

[0158] Response fluctuation features and response peak features are extracted from the set of response data time segments to form a response data feature sequence. This corresponds to converting feedback changes in the response data time segments into a computable feature expression. Response fluctuation features reflect the degree of fluctuation, frequency of change, and trend shift of the feedback data in the time segment, such as temperature fluctuation amplitude, vibration intensity change, and current change slope. Response peak features reflect extreme value information and sudden change information of the feedback data in the time segment, such as temperature peaks, vibration peaks, and current spikes. The source of response fluctuation features and response peak features is continuous sampling values ​​in the response data stream, obtained through statistical analysis, extreme value detection, and local trend analysis in each response data time segment. The response data feature sequence is arranged in chronological order, allowing feedback changes in different time segments to form a continuous expression. The resulting response data feature sequence is no longer a point-by-point accumulation of the original sensor data, but a time-series feature set after local induction, facilitating matching and comparison with the control command feature sequence.

[0159] A distance matrix is ​​calculated based on the control command feature sequence and the response data feature sequence, and a cumulative distance matrix is ​​calculated based on the distance matrix, corresponding to the establishment of a global difference distribution between control behavior and response changes. Each element in the distance matrix reflects the degree of difference between a control command time segment and a response data time segment. Sources of difference include inconsistencies between action switching changes and response fluctuation changes, inconsistencies between setpoint adjustments and response peak changes, and inconsistencies in the changing trends of the two types of time segments. After the distance matrix is ​​established, the matching degree between the control command feature sequence and the response data feature sequence at each time position can be visually observed. The cumulative distance matrix is ​​a recursive result formed based on the distance matrix, used to record the minimum cumulative difference from the initial time position to the current time position. The cumulative distance matrix does not simply sum all differences, but selects the cumulative combination with the smallest difference under time order constraints, so that subsequent path selection can reflect the overall matching degree while maintaining the time order. The existence of the cumulative distance matrix allows local time offsets, local response delays, and local fluctuation enhancements to be considered uniformly in the overall matching process.

[0160] The mapping path corresponding to the minimum cumulative distance is determined based on the cumulative distance matrix, which corresponds to selecting the correspondence with the minimum overall difference between the control command feature sequence and the response data feature sequence. The mapping path describes the optimal time pairing result between the set of control command time segments and the set of response data time segments. The mapping path is not a one-to-one fixed alignment, but allows one control command time segment to correspond to multiple response data time segments, and also allows multiple control command time segments to correspond to one response data time segment, thereby absorbing the response lag, response compression, and response stretching phenomena that exist in the actual operation of the equipment. The determination of the mapping path makes the time misalignment between control behavior and response change no longer dependent on fixed time offset correction, but obtains an adaptive correspondence based on the principle of minimum feature difference. The mapping path obtained in this way can both preserve the order of control behavior and feedback change and adapt to the non-constant delay of the equipment under different operating conditions.

[0161] Based on the mapping path, the control command time segment set and the response data time segment set are time-aligned to obtain an aligned control command segment sequence and an aligned response data segment sequence. This corresponds to reorganizing the two types of time segments that originally had time discrepancies into the same temporal semantics. Time alignment is not a simple truncation or shifting of the original sequence, but rather a rearrangement, repetition, compression, or expansion of the control command time segment and response data time segment according to the mapping path, so that each aligned control command segment corresponds to a semantically closest aligned response data segment. The aligned control command segment sequence retains the structural expression of control behavior over time, and the aligned response data segment sequence retains the state expression of feedback changes over time. After time alignment, the comparison of control behavior and response changes at the same time position has clear meaning, and subsequent difference calculations are no longer affected by time misalignment.

[0162] The timing difference is calculated based on the aligned control command fragment sequence and the aligned response data fragment sequence to generate an aligned residual sequence, which quantitatively expresses the degree of deviation between control behavior and feedback changes. The timing difference originates from the differences between the aligned control characteristics and response characteristics in dimensions such as amplitude, trend, rate of change, and local extrema. Each residual result in the aligned residual sequence reflects the magnitude of the difference between the control behavior and the actual response under the current time semantics. The aligned residual sequence can be represented as a single difference sequence or as a set of multi-dimensional difference results. This sequence is used to support state decoupling and anomaly identification in subsequent processing because the residual information obtained after time alignment has eliminated most of the time offset directly caused by sampling asynchrony and response delay, making the remaining deviations closer to the state anomalies and execution anomalies in real operating conditions.

[0163] For example, the formula for the dynamic time warping cost of control command characteristics and response characteristics is:

[0164]

[0165] in, This represents a sequence of control commands. This represents the characteristic sequence of the response data. This represents the set of mapping paths. Indicates the first in the control command feature sequence In the feature sequence of the feature and response data, the first feature is... The number of matching points for each feature. K represents the total number of matching points in the mapping path. This represents the distance cost corresponding to the k-th matching point. This represents the Euclidean distance between the control command characteristics and the response data characteristics. This represents the minimum normalization cost between the control command feature sequence and the response data feature sequence.

[0166] This embodiment segments the control command stream and response data stream according to a unified time window and extracts the control command feature sequence and response data feature sequence, which can convert the original continuous data into a comparable expression with temporal local semantics. By constructing a distance matrix and a cumulative distance matrix and determining the mapping path corresponding to the minimum cumulative distance, the optimal correspondence between control behavior and feedback changes can be established under the conditions of response delay and sampling rhythm differences. By aligning the control command time segment set and the response data time segment set in time and calculating the time difference, an aligned residual sequence with clear temporal semantics can be generated, thereby reducing the interference of time misalignment on deviation analysis and improving the data consistency of subsequent state identification and anomaly judgment.

[0167] In one embodiment, step S30 above includes:

[0168] S301, The aligned residual sequence is input into the manifold learning module to perform feature decoupling processing, and the aligned residual sequence is dimensionality reduced and mapped to a low-dimensional feature representation;

[0169] S302, the low-dimensional feature representation is physically decomposed and extracted to separate the thermal hysteresis state feature component, the mechanical vibration state feature component and the environmental disturbance state feature component.

[0170] S303, combine the thermal hysteresis state characteristic component, the mechanical vibration state characteristic component and the environmental disturbance state characteristic component to generate a state characteristic component;

[0171] S304, the state feature components are input into the long short-term memory network autoencoder, the reconstruction error is extracted, and the judgment conclusion that the reconstruction error exceeds the distribution threshold set based on the historical normal process distribution is determined as a deviation result;

[0172] S305, based on the deviation result, extract the state feature components within the current time window, inject anomaly identifiers into the extracted state feature components and perform combination encapsulation processing to generate structured perception events.

[0173] In this embodiment, the aligned residual sequence represents the difference between the control behavior and feedback changes that remains after time alignment. This difference result has eliminated most of the influence caused by differences in sampling rhythm, response delay, and time misalignment. Therefore, the change components it contains are closer to the actual state deviations that exist during equipment operation. The aligned residual sequence usually mixes thermal change deviations, mechanical response deviations, and environmental disturbance deviations. Deviations from different sources are superimposed on each other in the same sequence. If the judgment is made directly based on the overall residual amplitude, it is easy to mix the changes from different sources together, making it difficult to distinguish whether the deviation comes from thermal process anomalies, mechanical process anomalies, or environmental condition disturbances. Therefore, feature decoupling processing is required.

[0174] The manifold learning module undertakes the dimensionality reduction mapping task. After receiving the aligned residual sequence, instead of simply compressing the original sequence, it maps the data, originally located in the high-dimensional residual space, to a lower-dimensional representation space, based on the local neighborhood relationships and global distribution relationships between samples. This approach preserves the stable structural change patterns in the residual sequence while reducing the impact of noise and redundancy. The low-dimensional feature representation corresponds to the compact expression output by the manifold learning module. This expression compresses the main change patterns in the aligned residual sequence into fewer dimensions, allowing subsequent splitting operations to focus on clearer state differences. The source of the low-dimensional feature representation is not isolated numerical values, but a comprehensive mapping result of local temporal changes, overall amplitude changes, and structural relationships between adjacent time segments in the aligned residual sequence, thus preserving the main directions of state changes.

[0175] After the low-dimensional feature representation is formed, it needs to be decomposed and extracted into physical dimensions. This physical dimension decomposition and extraction corresponds to breaking down the mixed changes in the low-dimensional feature representation into several components with independent physical meanings according to the changing patterns of different physical processes. The thermal hysteresis state feature component is used to characterize the inconsistency between heating input, heat diffusion, and heat accumulation. When the actual temperature response fails to keep pace with changes in the equipment's heat input, or when the temperature change amplitude deviates continuously from the set trend, this change will be reflected in the thermal hysteresis state feature component. The mechanical vibration state feature component is used to characterize high-frequency or mid-frequency fluctuations caused by moving parts, transmission structures, or processing contact states. When changes in speed, turning movements, or load disturbances lead to an increase in vibration amplitude, enhanced periodicity, or increased impact fluctuations, this change will be reflected in the mechanical vibration state feature component. The environmental disturbance state feature component is used to characterize the unstable effects caused by changes in external temperature, cooling conditions, power supply fluctuations, or surrounding interference. When external disturbances occur during equipment operation that have no direct correlation with the control intent but continuously affect the response results, this change will be reflected in the environmental disturbance state feature component. This splitting method allows multiple types of changes that were originally superimposed in the same residual sequence to be separated into different expression spaces, so that subsequent judgments no longer rely solely on the magnitude of a single residual, but instead shift to the separate identification of biases from different physical sources.

[0176] After the thermal hysteresis state feature components, mechanical vibration state feature components, and environmental disturbance state feature components are extracted, they need to be combined to generate state feature components. State feature components are not simply a concatenation of the three components, but rather an integration of them within the same time window according to a unified structure, allowing multiple state changes within a single moment or time slice to be expressed in a consistent form. This approach preserves the differences between different physical sources while ensuring the overall state serves as a unified input for subsequent discrimination processes. Once the state feature components are formed, the thermal response state, mechanical response state, and environmental influence state of the equipment within the current time window are synchronously expressed, providing subsequent processing with a well-defined and semantically clear state representation.

[0177] The Long Short-Term Memory (LSTM) autoencoder (ASC) network is responsible for discriminating deviations from historical normal process distributions. After the state feature components are input into the LSM network ASC, it reconstructs these components based on the temporal memory developed during training with normal process samples. The reconstruction error corresponds to the difference between the current input state feature components and the model's reconstruction result. If the current state aligns with the historical normal process distribution, the model can reconstruct the input well, and the reconstruction error remains within a normal range. If the current state deviates from the historical normal process distribution, the model's reconstruction ability decreases, and the reconstruction error increases accordingly. The historical normal process distribution corresponds to the statistical results of state feature components formed by the equipment under stable processing conditions, including state distribution information under different material conditions, different operating ranges, and different normal operating conditions. A distribution threshold set based on the historical normal process distribution is used to define the boundary between normal and deviated states. Comparing the reconstruction error with the distribution threshold yields the deviation result. The deviation result is not a judgment on whether a single parameter exceeds its limit, but rather a holistic conclusion on whether the current comprehensive state deviates from the normal process distribution, thus more accurately reflecting the occurrence of anomalies under complex conditions.

[0178] After the deviation results are generated, state feature components need to be extracted within the current time window, and anomaly identifiers are injected into these components before a combined encapsulation process is performed. The current time window corresponds to the time range for generating the deviation results. Extracting state feature components within the current time window ensures that the event representation is consistent with the time of anomaly occurrence. The anomaly identifier indicates that the current state no longer falls within the historical normal process distribution range. The anomaly identifier can include the anomaly category, deviation level, deviation direction, and persistence attribute. The combined encapsulation process organizes the state feature components and anomaly identifiers into a unified data representation, so that subsequent processing no longer directly deals with scattered residual data and individual deviation judgments, but with a structured result containing state descriptions and anomaly semantics. This is how structured awareness events are formed. Structured awareness events contain the source of state change, the current time window, the degree of anomaly, and the content of state representation, thus converting the originally scattered numerical deviations into a unified, identifiable, transferable, and associative representation. The resulting structured awareness events retain the specific details of the current abnormal state and provide a clear data entry point for subsequent correlation analysis.

[0179] For example, the formula for the reconstruction error of state feature components is:

[0180]

[0181] in, This indicates the reconstruction error. This represents the state feature vector input at the current moment. This represents the reconstruction result of the feature vector of the current state by the autoencoder of the Long Short-Term Memory network. `Enc()` represents the encoder mapping function. `Dec()` represents the decoder reconstruction function. This represents the square norm, used to measure the difference between the input state features and the reconstructed state features.

[0182] The formula for determining the deviation threshold is:

[0183]

[0184] in, This represents the reconstruction error corresponding to the current state feature components. This represents the average reconstruction error obtained based on historical normal process data. This represents the standard deviation of the reconstruction error obtained based on historical normal process data. This represents the distribution threshold.

[0185] This embodiment reduces redundancy in mixed deviations by inputting the aligned residual sequence into the manifold learning module and generating a low-dimensional feature representation, making subsequent state separation more targeted. By physically decomposing the low-dimensional feature representation and extracting thermal hysteresis state feature components, mechanical vibration state feature components, and environmental disturbance state feature components, and then combining them to generate state feature components, thermal process changes, mechanical process changes, and environmental influence changes can be distinguished from mixed residuals. By inputting the state feature components into a long short-term memory network autoencoder and generating deviation results based on the relationship between reconstruction error and distribution threshold, the difference between the current state and the historical normal process distribution can be accurately expressed. By injecting anomaly markers into the state feature components within the current time window and performing combined encapsulation processing, numerical deviations can be converted into clearly structured perceptual events, thereby improving the interpretability, consistency, and processability of anomaly identification results.

[0186] In one embodiment, step S40 above includes:

[0187] S401, the structured perception event and the running context are input into the attention neural network layer of the graph to obtain the entity embedding vector;

[0188] S402, the entity embedding vector is mapped to the abnormal appearance entity node in the mechanism knowledge graph, and the abnormal appearance entity node is established as the starting point of logical deduction;

[0189] S403, Starting from the logical deduction starting point, backtrack the source path along the reverse direction of the causal edge in the mechanism knowledge graph, extract all potential causal nodes that directly or indirectly point to the logical deduction starting point, and form a set of potential causal nodes;

[0190] S404, based on the causal edge propagation relationship between the set of potential causal nodes and the starting point of logical deduction and the prior probability of the nodes, perform inverse probability deduction to derive the posterior conditional probability of each potential causal node in the set of potential causal nodes, and determine the potential causal node with the largest posterior conditional probability as the target physical root cause.

[0191] In this embodiment, structured sensing events are used to characterize abnormal states identified during equipment operation, while the operational context is used to characterize the corresponding control behaviors, processing sections, and operating condition changes when the abnormality occurs. Structured sensing events provide the results of abnormal behavior, while the operational context provides the environmental and behavioral conditions at the time of the abnormality. Analyzing only based on structured sensing events can yield the category and intensity of the abnormal phenomenon, but it is difficult to distinguish the influence of different process conditions on the abnormal result during the abnormality formation process. Analyzing only based on the operational context can yield information on control behaviors and operating condition changes, but it is difficult to determine which type of physical process triggered the actual abnormality. Introducing structured sensing events and operational context together into a mechanistic knowledge graph allows for the establishment of a correspondence between abnormal behavior, operating conditions, process parameters, and physical causes in a unified expression space, enabling subsequent processing to shift from judging the appearance of abnormalities to locating the causes of abnormalities.

[0192] The graph attention neural network layer undertakes the joint encoding task. The structured perception event contains anomaly category, anomaly severity, anomaly time window, and related state component information. The operational context contains the current processing segment identifier, action type identifier, and operating condition switching identifier. After receiving these two types of inputs, the graph attention neural network layer does not process them separately but rather weights and aggregates the anomaly and operating condition information based on the degree of correlation between different input items. The weight updates in the graph attention layer reflect the differences in the importance of different input information for root cause localization. For example, under high heat load conditions, the weight of temperature anomaly-related information will increase; under high acceleration conditions, the weight of vibration anomaly-related information will increase. After weighted encoding, an entity embedding vector is obtained. The entity embedding vector is used to compress the structured perception event and operational context into a unified vector space, so that the subsequent mapping process no longer relies on single-field matching but on multi-dimensional semantic expression for matching. The entity embedding vector simultaneously contains anomaly behavior information, operating condition constraint information, and the correlation information between the two, thus supporting the subsequent location of the closest entity node in the knowledge graph.

[0193] In the mechanistic knowledge graph, anomalous entity nodes correspond to observable or identifiable abnormal states during equipment operation. The sources of these anomalous entity nodes include describable results such as temperature shifts, enhanced vibration, abnormal current, extrusion instability, localized accumulation, and path distortion. They can also include thermal hysteresis anomalies, mechanical response anomalies, and environmental disturbance anomalies derived from state feature components. Mapping entity embedding vectors to anomalous entity nodes in the mechanistic knowledge graph corresponds to locating the current observation result to the closest anomalous entity node based on the similarity relationship between the entity embedding vector and the node representation in the graph. This mapping process transforms the data representation from the operational process into explicit node representations in the knowledge graph, allowing subsequent inferences to be performed directly on the graph structure. Establishing anomalous entity nodes as the starting point for logical inference means that root cause localization will revolve around the currently confirmed anomalous result, rather than conducting an unconstrained search across all possible causes. This narrows the candidate range, improves reverse inference efficiency, and ensures that the inference direction aligns with the actual observation results.

[0194] After establishing the starting point of logical deduction, it is necessary to trace back the source path along the causal edges in the mechanistic knowledge graph. Causal edges are used to represent the causal transmission relationship between nodes, with the direction pointing from the cause node to the result node. Tracing back in the reverse direction corresponds to searching upstream from the abnormal phenomenon entity node for the potential cause node that led to the abnormal result. Nodes that directly point to the starting point of logical deduction usually represent first-order causes, such as local temperature deviations directly caused by heating imbalance or cooling anomalies; nodes that indirectly point to the starting point of logical deduction usually represent more upstream causes, such as changes in the thermal conductivity of materials further causing heating imbalance by affecting the temperature distribution, ultimately forming the abnormal phenomenon. By extracting all nodes that directly or indirectly point to the starting point of logical deduction, a set of potential causal nodes can be formed. Each node in the set of potential causal nodes is not added arbitrarily, but rather satisfies the existence of a causal path in the graph, and that the causal path can lead to the current abnormal phenomenon entity node. The candidate set formed in this way covers both direct and indirect causes, so that root cause identification will not stop at the surface-level anomaly precursor nodes, but can extend to deeper causes.

[0195] Once the set of potential causal nodes is formed, inverse probabilistic deduction needs to be performed. Inverse probabilistic deduction does not simply rely on Boolean judgments based on the existence of causal edges in the graph; instead, it combines causal edge transitivity and node prior probabilities to probabilistically rank candidate nodes. Causal edge transitivity reflects the strength and direction of the influence of different candidate causes on the current anomalous result. Some causal edges correspond to strong coupling relationships, such as the strong transitivity between thermal management imbalance and temperature deviation; others correspond to weak coupling relationships, such as the impact of environmental temperature disturbances on local current anomalies requiring multiple intermediate processes to manifest. Node prior probabilities reflect the frequency of candidate causes in historical operation, their occurrence under current operating conditions, and their prior credibility in the current state of the equipment. Combining causal edge transitivity with node prior probabilities avoids determining cause priority solely based on topological distance and avoids ignoring current causal structural constraints based solely on historical frequency.

[0196] The derivation of posterior conditional probability corresponds to calculating the probability that each potential causal node in the set of potential causal nodes is the true cause, given the current abnormal phenomenon entity nodes and the current operating context. A high posterior conditional probability indicates that a potential causal node not only has a clear causal path to the current abnormal phenomenon but also has higher explanatory power and rationality under the current operating conditions. A low posterior conditional probability indicates that although the node can lead to the logical deduction starting point in the graph structure, it does not have strong causal explanatory power under the constraints of the current operating conditions. By comparing the posterior conditional probabilities of each potential causal node in the set of potential causal nodes, the potential causal node with the highest posterior conditional probability can be selected. This node is determined as the target physical root cause. The target physical root cause is no longer a simple description of the abnormal phenomenon but an expression of physical cause that can provide a causal explanation for the current abnormal phenomenon, such as heat transfer imbalance, increased material flow resistance, amplified local vibration, abnormal driving load, or mismatch in operating condition switching.

[0197] For example, the formula for the posterior conditional probability of a target physical root cause is:

[0198]

[0199] in, This represents the i-th potential causative node. This represents the observed abnormal entity node, corresponding to the starting point of logical deduction. This represents the posterior conditional probability that the i-th potential causative node becomes the true root cause, given that abnormal manifestation entity nodes have been observed. This represents the conditional probability of the current abnormal entity node occurring when the i-th potential causative node occurs. This represents the prior probability of the i-th potential causative node. This represents the sum of the normalized probabilities of all potential causative nodes to the current abnormal entity node.

[0200] This embodiment combines structured perception events and operational context input graph attention neural network layers to form entity embedding vectors, enabling the joint expression of abnormal state information and operating condition information. By mapping the entity embedding vectors to abnormal manifestation entity nodes in the mechanistic knowledge graph and establishing them as the starting point for logical deduction, abnormal observation results can enter an interpretable causal expression space. By tracing back the source path along the causal edges in the mechanistic knowledge graph and forming a set of potential causal nodes, candidate causes can be limited to those causally related to the current abnormal result. By combining the causal edge propagation relationship and the prior probability of nodes to perform inverse probability deduction and comparing the posterior conditional probability, the node with the strongest explanatory power can be screened from multiple candidate causes, thereby improving the accuracy of target physical root cause localization and the consistency of causal explanation.

[0201] In one embodiment, step S50 above includes:

[0202] S501 reads the thermal decomposition threshold, thermal conduction constraint, and melting zone constraint from the material property boundary, and reads the upper limit of velocity, upper limit of acceleration, and motion stroke constraint from the equipment motion boundary;

[0203] S502, construct physical safety boundary conditions based on the thermal decomposition threshold, the thermal conduction constraint, the melting zone constraint, the upper limit of velocity, the upper limit of acceleration, and the motion stroke constraint;

[0204] S503, extract the parameter items to be adjusted based on the target physical root cause, generate multiple candidate parameter combinations around the parameter items to be adjusted, and gather the multiple candidate parameter combinations to form a candidate strategy set;

[0205] S504, Substitute each candidate strategy in the candidate strategy set into the physical security boundary conditions for boundary verification, eliminate candidate strategies that do not meet the physical security boundary conditions, and obtain the filtered candidate strategy set.

[0206] S505, calculate the geometric accuracy evaluation value, manufacturing time evaluation value, energy consumption cost evaluation value and boundary violation penalty evaluation value for each candidate strategy in the filtered candidate strategy set, and input the geometric accuracy evaluation value, the manufacturing time evaluation value, the energy consumption cost evaluation value and the boundary violation penalty evaluation value into the value evaluation function to calculate the comprehensive evaluation value corresponding to each candidate strategy;

[0207] S506, Based on the comprehensive evaluation value corresponding to each candidate strategy, determine the target adjustment strategy from the set of candidate strategies after screening.

[0208] In this embodiment, the material property boundaries correspond to the range of thermal states and heat transfer states that the material can withstand during processing. The thermal decomposition threshold limits the upper temperature at which irreversible decomposition occurs after heating; exceeding this threshold will cause material performance degradation, gas precipitation, or deposition instability. The thermal conductivity constraint limits the material's thermal diffusion capacity under heat input, keeping the spatial and temporal distribution of heat within a controllable range and avoiding localized heat accumulation or insufficient thermal diffusion. The melting range constraint limits the temperature range in which the material transitions from a solid state to a processable state, maintaining a balance between the flowable and formable states. The material property boundaries do not act on any single parameter alone, but rather collectively define the processable range of the material during heating, flow, and cooling processes.

[0209] The motion boundaries of the equipment correspond to the dynamically permissible range of the actuator during motion control. The upper speed limit restricts the allowable displacement change per unit time, preventing discontinuous deposition, increased vibration, or increased path-following errors due to excessive speed. The upper acceleration limit limits the slope range of velocity changes, avoiding mechanical shock, localized vibration, and accumulation of inertial errors caused by excessive acceleration or deceleration. Motion stroke constraints limit the spatial range reachable by the actuator, ensuring the motion path always remains within the permissible motion range of the equipment structure. The motion boundaries of the equipment are not isolated constraints but, together with material property boundaries, determine the effective range of various control parameters during equipment operation.

[0210] The physical safety boundary conditions consist of a thermal decomposition threshold, thermal conduction constraints, melting zone constraints, upper velocity limits, upper acceleration limits, and motion stroke constraints. The construction process involves converting each individual boundary parameter into a unified and verifiable constraint expression, allowing boundary information from different sources to enter the same constraint system. The thermal decomposition threshold and melting zone constraints form the temperature allowable range, the thermal conduction constraints form the thermal diffusion allowable range, the upper velocity and upper acceleration limits form the dynamic allowable range, and the motion stroke constraints form the spatial allowable range. The physical safety boundary conditions not only restrict the upper and lower bounds of individual parameters but also limit the combined range of multiple parameters changing simultaneously. For example, when the temperature is close to the thermal decomposition threshold, the range of velocity and acceleration needs to be narrowed; when thermal conductivity is low, the combination relationship between heat input and motion cycle needs to be re-constrained. The constructed physical safety boundary conditions enable unified verification of subsequent adjustment results, ensuring that any adjustment result must be within the material and equipment allowable ranges.

[0211] Target physical causes are used to characterize the identified sources of anomalies. Target physical causes are not merely apparent anomalies, but rather expressions of the physical reasons that trigger abnormal results, such as enhanced thermal hysteresis, increased local flow resistance, amplified mechanical vibration, or intensified environmental disturbances. The role of target physical causes is to define the direction of adjustment, ensuring that subsequent parameter changes are not blindly searched for all parameters, but rather adjusted around the set of parameters directly related to the current cause of the anomaly. The adjustable parameters correspond to the adjustable objects extracted from the target physical causes. If the target physical cause is related to heat transfer, the adjustable parameters may include temperature setpoints, heat input rhythm, and cooling intensity; if the target physical cause is related to material flow, the adjustable parameters may include extrusion rate, nozzle temperature, and movement speed; if the target physical cause is related to mechanical disturbance, the adjustable parameters may include acceleration, velocity variation amplitude, and path smoothing parameters. The extraction of adjustable parameters establishes a clear correspondence between the adjustable objects and the causes of the anomalies, preventing the subsequent parameter generation process from deviating from the actual source of the problem.

[0212] Candidate parameter combinations correspond to multiple sets of parameter results formed by combining the parameter to be adjusted within multiple value ranges. Each candidate parameter combination represents a possible adjustment method. Multiple candidate parameter combinations are aggregated to form a candidate strategy set. The candidate strategy set is used to carry all the candidate adjustment results generated around the target physical root cause, enabling subsequent screening and evaluation among multiple alternatives. Differences in the candidate strategy set may be reflected in aspects such as temperature adjustment range, speed adjustment range, cooling intensity adjustment range, and motion behavior adjustment methods. The generation process of the candidate strategy set establishes a calculable parameter mapping relationship between the target physical root cause and specific adjustment actions.

[0213] After the candidate strategy set is generated, it needs to be validated against physical safety boundary conditions. Boundary validation involves evaluating each candidate strategy under material, thermal, dynamic, and spatial constraints. If a candidate strategy causes the temperature to exceed the thermal decomposition threshold, it does not meet the physical safety boundary conditions. Similarly, if a candidate strategy causes acceleration to exceed the allowable range or the trajectory to exceed the travel limit, it cannot be retained. The purpose of boundary validation is not to rank the results, but to directly eliminate all candidate strategies with physical risks. The resulting filtered candidate strategy set has eliminated parameter combinations that could cause material damage, mechanical instability, or spatial exceedances, leaving only candidate strategies within the executable range. This filtered candidate strategy set then becomes the input set for subsequent comprehensive evaluation.

[0214] The geometric accuracy assessment value characterizes the impact of candidate strategies on forming dimensions, path following error, edge integrity, and local deposition accuracy. The manufacturing time assessment value characterizes the impact of candidate strategies on total task execution time, local cycle time variation, and overall efficiency. The energy cost assessment value characterizes the changes in heat input, drive power, and auxiliary component power consumption during the execution of candidate strategies. The boundary violation penalty assessment value characterizes the risk level of candidate strategies when they approach physical safety boundary conditions. Even if a candidate strategy is not directly eliminated by boundary checks, if it approaches the thermal decomposition threshold, the acceleration upper limit, or the motion stroke constraint, its risk level needs to be expressed through the boundary violation penalty assessment value. The geometric accuracy assessment value, manufacturing time assessment value, energy cost assessment value, and boundary violation penalty assessment value describe the performance of candidate strategies from four dimensions: quality, efficiency, cost, and risk, respectively.

[0215] The value assessment function receives the four types of assessment values ​​mentioned above and generates a comprehensive assessment value. The value assessment function is not a single-indicator judgment, but rather a unified measure of geometric accuracy, manufacturing time, energy cost, and out-of-bounds penalty. The comprehensive assessment value corresponds to the overall performance of each candidate strategy under multi-dimensional constraints and multi-dimensional objectives. If a candidate strategy has a high geometric accuracy assessment value but poor manufacturing time and energy cost assessment values, the comprehensive assessment value will not be simply determined by a single indicator. If a candidate strategy is highly efficient but has a large out-of-bounds penalty assessment value, the comprehensive assessment value will be suppressed. The existence of the value assessment function ensures that the selected candidate strategy set no longer relies on a single dimension for selection, but rather achieves a comprehensive balance among quality, time, energy consumption, and risk. A better comprehensive assessment value indicates that the candidate strategy is more suitable for the adjustment needs corresponding to the current objective's physical root causes, while meeting safety conditions.

[0216] The target adjustment strategy corresponds to the final adjustment result determined from the selected candidate strategy set based on a comprehensive evaluation value. The determination of the target adjustment strategy is not an unconditional selection of all candidate strategies, but rather the optimal result obtained by comprehensively evaluating the validated candidate strategies, provided that the physical safety boundary conditions have taken effect. Therefore, the target adjustment strategy simultaneously satisfies two conditions: first, it is within the allowable range of the physical safety boundary conditions; second, it achieves a better balance between geometric accuracy, manufacturing time, energy consumption cost, and risk level. The resulting target adjustment strategy is both feasible to execute and has reasonable overall performance, and can be directly used as input for subsequent pre-simulation and control writing.

[0217] For example, the formula for physical safety boundary condition constraints is:

[0218]

[0219] This represents the physical safety boundary conditions. This represents the control input corresponding to the candidate strategy at the current moment. Indicates control input The target temperature under action. This indicates the lower limit temperature at which a material enters a processable molten state. This indicates the upper limit temperature corresponding to the material's thermal decomposition threshold. It represents the rate of temperature change over time and is used to characterize the intensity of temperature change during heat conduction. This represents the maximum permissible rate of temperature change corresponding to the thermal conduction constraint. Indicates control input The feed rate under action. This indicates the maximum feed rate. Indicates control input The acceleration of motion under action. This represents the maximum acceleration. Indicates control input The position or stroke value of the motion under action. This indicates the minimum permissible travel limit of the device. This indicates the maximum allowable travel range of the device.

[0220] The formula for comprehensively evaluating candidate strategies is:

[0221]

[0222] in, This represents the overall evaluation value of the candidate strategies. Indicates the strategy parameters. Indicates by parameters Determined candidate strategies. [] indicates the candidate strategy The expected evaluation results are as follows. This indicates the weight for geometric accuracy evaluation. This represents the geometric accuracy assessment value. This indicates the weighting of manufacturing time in the evaluation. This indicates the increment in manufacturing time. This indicates the weighting of energy consumption cost assessment. This represents the energy consumption cost assessment value. This indicates the weighting of the penalty assessment for crossing the boundary. This indicates the penalty assessment value for crossing the boundary.

[0223] This embodiment extracts thermal decomposition thresholds, thermal conduction constraints, melting zone constraints, velocity limits, acceleration limits, and motion stroke constraints from material property boundaries and equipment motion boundaries, and constructs physical safety boundary conditions. This unifies the allowable ranges of materials and equipment into the same constraint system. By extracting the parameters to be adjusted based on the target physical root cause and generating a candidate strategy set, the adjustment results can be generated in a targeted manner around the cause of the anomaly. By using the physical safety boundary conditions to screen the candidate strategy set, candidate strategies with material risks, mechanical risks, and spatial risks can be eliminated. By calculating geometric accuracy evaluation values, manufacturing time evaluation values, energy consumption cost evaluation values, and boundary violation penalty evaluation values ​​for the screened candidate strategy set, and inputting them into a value evaluation function to generate a comprehensive evaluation value, a unified evaluation result can be formed among quality, efficiency, cost, and risk. By determining the target adjustment strategy based on the comprehensive evaluation value, the final adjustment result can simultaneously meet the requirements of executability and comprehensive rationality, thereby improving the matching degree between the adjustment result and the current cause of the anomaly and reducing the subsequent execution risk.

[0224] In one embodiment, prior to step S50, the method further includes:

[0225] S5001, reads the thermal decomposition threshold, melting temperature constraint and thermal conduction parameters from the material property database;

[0226] S5002, reads the maximum motion acceleration, maximum feed rate and equipment motion constraint parameters from the equipment specification configuration file;

[0227] S5003 combines the read thermal decomposition threshold, melting temperature constraint and thermal conduction parameters to form the material property boundary;

[0228] S5004 combines the read maximum motion acceleration, maximum feed rate, and equipment motion constraint parameters to form the equipment motion boundary.

[0229] In this embodiment, before constructing the physical safety boundary conditions, the material property database and equipment specification configuration file are read. This corresponds to first establishing a set of source parameters for constraint parameters, and then converting the source set into a boundary expression that can be directly invoked for subsequent control judgments. The material property database is used to store parameters related to the material's thermal behavior, flow behavior, and heat transfer behavior. The thermal decomposition threshold describes the upper temperature limit at which the material undergoes irreversible decomposition during heating. When the processing temperature exceeds this range, the material may experience structural damage, gas evolution, decreased adhesion, or deterioration of deposition quality. The melting temperature constraint describes the temperature range at which the material transitions from a solid state to a processable molten state. This range affects both whether the material can be stably extruded and its ability to retain its morphology after deposition. The thermal conductivity parameters describe the material's ability to transfer heat during heating, deposition, and cooling. These parameters determine the diffusion rate of local heat within and around the material, thus affecting the uniformity of temperature distribution, interlayer bonding, and the degree of local heat accumulation. When reading the material property database, instead of simply extracting a single value, the system retrieves records related to thermal decomposition threshold, melting temperature constraint, and heat conduction parameters from the database based on the current material type, current process conditions, and current processing stage. It then performs consistency checks to eliminate missing, conflicting, or mismatched data, ensuring that the parameters participating in subsequent boundary combinations have a unified material semantics.

[0230] The equipment specification configuration file stores parameters related to the equipment's motion and execution capabilities. Maximum acceleration limits the upper limit of velocity variation for moving parts; this parameter directly affects inertial impact, mechanical vibration, and path deviation during turning movements. Maximum feed rate limits the equipment's motion capability in the processing direction per unit time; this parameter determines both processing efficiency and whether material deposition can match the motion rhythm. Equipment motion constraint parameters characterize the comprehensive limitations of the equipment in terms of motion space, motion mode, power output, and structural allowable range. These parameters may include allowable ranges for motion stroke, angle changes, drive response time, local motion prohibition ranges, and continuous operation load limits. When reading the equipment specification configuration file, the maximum acceleration, maximum feed rate, and equipment motion constraint parameters are extracted not only to obtain static specification values ​​but also to establish the set of currently available motion capabilities for the equipment, ensuring that subsequent control decisions always revolve around the equipment's actual operating range.

[0231] The combination of thermal decomposition threshold, melting temperature constraint, and thermal conductivity parameters forms the material property boundary, corresponding to the transformation of dispersed material parameters into a unified material boundary expression that can be used for constraint judgment. The thermal decomposition threshold defines the upper limit of heat input, the melting temperature constraint defines the processable melting range, and the thermal conductivity parameters define the temperature diffusion capability and heat accumulation characteristics. When these three types of parameters exist individually, they can only express the local limitations of the material during the heating process. Only when combined can they fully describe the thermal environment in which the material is allowed to be processed, under what thermal diffusion conditions it can still maintain stable deposition, and at what rate of temperature change it will not produce abnormal changes. After the material property boundary is formed, subsequent processing does not deal with isolated temperature parameters and isolated thermal conductivity parameters, but with a comprehensive boundary result that reflects the overall thermal processing capability of the material.

[0232] The maximum acceleration, maximum feed rate, and equipment motion constraint parameters combine to form the equipment motion boundary, corresponding to the unified transformation of dispersed equipment dynamics and mechanism constraints into a boundary expression usable for constraint judgment. Maximum acceleration limits the actuator's speed variation capability, maximum feed rate limits the motion cycle capability, and equipment motion constraint parameters limit the spatial range, action mode, and operational limitations. The combination of these three parameters collectively expresses the range of motion states the equipment can achieve during processing. For example, although a certain speed level may not exceed the maximum feed rate, if combined with a large acceleration change and a short transition interval, it may still exceed the motion capability allowed by the equipment motion constraint parameters. After the equipment motion boundary is formed, subsequent processing can use this unified motion boundary expression to determine whether a set of control parameters will cause overspeed, overshoot, amplified jitter, structural collision, or trajectory deviation.

[0233] Material property boundaries and equipment motion boundaries play different constraint roles in subsequent processing. Material property boundaries are responsible for limiting the range of thermal and material states during processing, ensuring that adjustments to control parameters do not compromise the material's machinability. Equipment motion boundaries are responsible for limiting the range of motion and speed of the actuators, ensuring that adjustments to control parameters do not exceed the equipment's dynamic and mechanical capabilities. Forming and independently expressing these two types of boundaries allows for targeted constraint references in subsequent processing when facing different anomaly types. Thermally related anomalies rely more on material property boundaries, motion-related anomalies rely more on equipment motion boundaries, and anomalies involving thermal-motion coupling problems rely on both types of boundaries, thus ensuring that subsequent control decisions have clear sources and objects of constraint.

[0234] This embodiment reads thermal decomposition thresholds, melting temperature constraints, and heat conduction parameters from the material property database and combines them to form material property boundaries, thus uniformly expressing the allowable range of materials during heating, melting, and heat transfer. Similarly, by reading maximum acceleration, maximum feed rate, and equipment motion constraint parameters from the equipment specification configuration file and combining them to form equipment motion boundaries, it uniformly expresses the allowable range of equipment speed variations, feed cycle time, and motion capabilities. Once the material property boundaries and equipment motion boundaries are formed, subsequent control decisions no longer rely on isolated parameters but can directly call boundary results with clear sources and complete semantics, thereby improving the completeness of constraint expression and the targeting of subsequent adjustment processes.

[0235] In one embodiment, step S60 above includes:

[0236] S601, The target adjustment strategy is input into the virtual simulation engine to deduce the equipment evolution response trajectory during the execution of the target adjustment strategy, so as to complete the physical feasibility pre-simulation;

[0237] S602, extract the predicted physical quantity from the device evolution response trajectory, and verify that the predicted physical quantity is within the range of the physical safety boundary conditions to complete the thought chain verification;

[0238] S603, based on the underlying communication interaction protocol, converts the target adjustment strategy through the physical feasibility pre-simulation and the thought chain verification into underlying control commands;

[0239] S604, the underlying control commands are written to the device control terminal through the communication interface connected to the device control terminal.

[0240] In this embodiment, the target adjustment strategy corresponds to the parameter adjustment result obtained after boundary screening and comprehensive evaluation. This result already has an adjustment direction for the current abnormal state, but it still belongs to the parameter expression of the control layer and has not yet been verified before execution, nor has it been converted into a control format that the device control terminal can directly recognize. Therefore, before writing it into the device control terminal, four consecutive processing steps need to be completed around the target adjustment strategy: physical feasibility pre-drilling, thought chain verification, control protocol conversion, and communication writing, so that the target adjustment strategy is transformed from an adjustment expression into an executable expression.

[0241] The virtual simulation engine is used to perform physical feasibility simulations. After receiving the target adjustment strategy, the virtual simulation engine does not perform static checks on the parameters, but rather dynamically extrapolates the evolution of the equipment's state during the execution of the target adjustment strategy based on the equipment's structural parameters, heat transfer parameters, motion parameters, and material parameters. The target adjustment strategy includes temperature adjustment, velocity adjustment, acceleration adjustment, cooling intensity adjustment, or path adjustment. Changes in these parameters collectively affect the equipment's thermal response, mechanical response, and material response during execution. After the target adjustment strategy is input into the virtual simulation engine, the engine calculates the equipment's state changes in subsequent control cycles in a time-progressive manner, establishing a calculable correspondence between changes in control parameters and changes in the equipment's physical state.

[0242] The equipment evolution response trajectory corresponds to the dynamic results output by the physical feasibility pre-simulation. It does not represent the state at a single point in time, but rather the continuous result of the equipment state changing over time under the action of the target adjustment strategy. The equipment evolution response trajectory can include temperature evolution trajectory, vibration evolution trajectory, current evolution trajectory, displacement evolution trajectory, and load evolution trajectory. The temperature evolution trajectory characterizes the temperature change trend in the processing area, nozzle area, or key component area after a change in heat input. The vibration evolution trajectory characterizes the fluctuation of the mechanical response after changes in motion parameters. The current evolution trajectory characterizes the degree of change in the driving load after control adjustment. The displacement evolution trajectory characterizes whether the executed action can be completed along the predetermined path. The load evolution trajectory characterizes the dynamic changes experienced by each key component during the execution of the target adjustment strategy. Through the equipment evolution response trajectory, the target adjustment strategy can be transformed from the adjustment intention at the parameter level into the state result at the operational level, allowing subsequent judgments to move beyond the parameters themselves and focus on the actual effects after parameter execution.

[0243] Physical feasibility simulations pre-analyze the potential equipment conditions that may arise after the execution of a target adjustment strategy. This process focuses on whether the target adjustment strategy can be jointly withstood by the equipment and materials under real operating conditions. If the target adjustment strategy causes the temperature evolution trajectory to approach the material decomposition boundary, or causes an abnormal amplification of the vibration evolution trajectory, or causes the displacement evolution trajectory to deviate from the allowable path, then the target adjustment strategy cannot be directly executed. The value of physical feasibility simulations lies in identifying potential risks before execution, avoiding problems that only become apparent after the adjustment parameters are applied to the actual equipment.

[0244] After the equipment evolution response trajectory is formed, predictive physical quantities need to be extracted from it. These predictive physical quantities correspond to key physical results in the equipment evolution response trajectory that directly characterize the safe, stable, and executable states. Predictive physical quantities can include predicted temperature values, predicted vibration amplitudes, predicted drive currents, predicted accelerations, predicted displacement deviations, and predicted heat accumulation. These predictive physical quantities are not arbitrarily selected but rather chosen based on equipment safety and process continuity. Temperature-related predictive physical quantities are used to determine whether thermal conditions will cause material anomalies; vibration-related predictive physical quantities are used to determine whether mechanical responses will cause motion instability; current and load-related predictive physical quantities are used to determine whether the actuator will enter a high-risk operating range; and displacement-related predictive physical quantities are used to determine whether trajectory execution will deviate.

[0245] The thought chain verification further validates the logical and constraint consistency of the execution results of the target adjustment strategy. It is not a simple repetition of physical feasibility simulation, but rather a segmented verification of the causal process from parameter changes to response changes in the target adjustment strategy, based on the predicted physical quantities in the equipment's evolution response trajectory. This verification process requires a consistent relationship between the changes in adjustment parameters, simulation results, and boundary conditions. For example, after lowering the temperature setpoint, the predicted temperature should decrease accordingly; after increasing the motion speed, the predicted acceleration and vibration amplitude should not exceed the allowable range; after increasing the cooling intensity, the predicted heat accumulation should show a decreasing trend. If the parameter changes are inconsistent with the changes in predicted physical quantities, or if the predicted physical quantities respond to the parameter changes but exceed the limits of the physical safety boundary conditions, the target adjustment strategy fails the thought chain verification. The key judgment criterion in the thought chain verification is the physical safety boundary conditions. These conditions transform material property boundaries and equipment motion boundaries into execution constraints, ensuring that the determination of predicted physical quantities does not rely on empirical manual judgment but on clearly defined boundary ranges. With this approach, physical feasibility simulation addresses whether an executable process can be deduced, while thought chain verification addresses whether the deduction results meet boundary constraints and maintain logical consistency between parameters and responses.

[0246] The target adjustment strategy, verified through physical feasibility simulation and thought chain validation, is still a high-level parameter expression oriented towards control logic. It lacks a format suitable for direct writing to the device control terminal, thus requiring control format conversion based on the underlying communication interaction protocol. This protocol defines the control fields, parameter order, data length, encoding format, and transmission format that the device control terminal can recognize. The temperature change, speed change, acceleration change, and execution cycle change included in the target adjustment strategy need to be converted into specific control codes such as temperature setting instructions, speed setting instructions, motion control instructions, and execution timing instructions according to the underlying communication interaction protocol. This conversion process transforms the target adjustment strategy from an abstract adjustment expression into low-level control instructions that the device control terminal can directly parse and execute.

[0247] After the low-level control commands are generated, they need to be written through a communication interface connected to the device control terminal. The communication interface carries the transmission of control results from the edge side to the device control terminal; the interface type can be a serial interface, bus interface, or network communication interface. The writing process requires that the low-level control commands be consistent with the current control cycle, ensuring that the newly generated low-level control commands enter the device control terminal at the correct time and position. The writing process also requires that the format of the low-level control commands be consistent with the parsing rules of the device control terminal to avoid field misalignment, parameter truncation, or execution order conflicts. After being written through the communication interface, the target adjustment strategy completes the entire transformation process from high-level adjustment expression to pre-performance verification, to low-level control encoding, and finally to the device execution entry point.

[0248] In one application example, the device is a fused deposition modeling (FDM) machine for printing high-temperature resistant polymer structural parts. The machine includes a control unit, an edge computing terminal, a sensor acquisition unit, and actuators. The control unit outputs control commands for nozzle movement, extrusion, heating, cooling, and layer switching. The sensor acquisition unit collects response data such as nozzle temperature, platform temperature, extrusion motor current, frame vibration, motion axis position, and cooling airflow. The edge computing terminal deploys a PMA (Process Monitoring Agent), an EDA (Equipment Diagnostic Agent), and a DSA (Decision Support Agent). These agents run as independent processes and communicate asynchronously via a message broker using a publish-subscribe approach. The PMA continuously performs high-frequency data acquisition and status identification; the EDA performs causal inference upon receiving anomalies; and the DSA executes safety decisions and control deployments upon receiving diagnostic conclusions. Even when the EDA is performing computationally intensive knowledge graph inference, the PMA continues to acquire control command and response data streams, preventing diagnostic computation from blocking the high-frequency monitoring process.

[0249] Before the printing task begins, the cloud-based evolutionary side constructs a thermally coupled digital twin environment. This digital twin environment establishes a virtual model based on the printhead structure, heating area, cooling air duct, motion mechanism, and material delivery path of the printing equipment. Friction coefficient, thermal conductivity, density, and ambient temperature disturbances are introduced into the virtual model to form multiple sets of unsteady virtual operating conditions. For example, in one set of virtual operating conditions, the friction coefficient between the printhead and the consumable increases, the material's thermal conductivity decreases, and the ambient temperature fluctuates; in another set of virtual operating conditions, the material density changes, cooling conditions are enhanced, and the acceleration of the motion axis changes frequently. The cloud-based evolutionary side trains a high-dimensional time-series model based on these multiple sets of unsteady virtual operating conditions and designates the trained high-dimensional time-series model as the teacher model. Subsequently, the teacher model receives all physical field data and generates teacher features, while the student model receives sensor time-series data and generates student features. The cloud-based evolutionary side performs knowledge distillation training based on the teacher and student features to obtain a lightweight edge model. This lightweight edge model is then quantized and compressed, and deployed to the edge computing terminal. In this way, the edge computing terminal can perform subsequent monitoring and inference processing on-site using a lightweight edge model, without continuously relying on cloud computing resources.

[0250] During printing, the device control terminal continuously outputs control commands according to the path data of the structural part to be printed. The PMA process monitoring agent continuously reads the control commands through the command communication connection between the device control terminal and the edge computing terminal, and buffers them in the output order to form a control command stream. The control command stream contains motion control fields, energy control fields, and layer switching control fields. The motion control field is used to indicate the nozzle movement position, feed speed, and path turning information; the energy control field is used to indicate the nozzle temperature, platform temperature, extrusion power, and cooling intensity; and the layer switching control field is used to indicate the current printing layer switch, infill area switch, and outer contour processing switch. The PMA process monitoring agent decomposes and parses the control command stream to obtain each control command containing the motion control field, energy control field, and layer switching control field, and writes a corresponding time stamp for each control command.

[0251] Meanwhile, the PMA process monitoring agent reads temperature, vibration, and current sampling values ​​through data communication between the sensor acquisition end and the edge computing terminal, and aggregates them according to the sampling time sequence to form a response data stream. Temperature sampling values ​​reflect the thermal state of the nozzle and deposition areas, vibration sampling values ​​reflect the mechanical disturbances of the motion mechanism and printing platform, and current sampling values ​​reflect the load changes of the extrusion motor and motion drive components. Based on the motion control, energy control, and layer switching control fields in each control command with corresponding time identifiers, the PMA process monitoring agent extracts the current processing segment identifier, action type identifier, and operating condition switching identifier, and combines these identifiers to form the operating context. For example, the current processing segment identifier indicates that a high fill density area is being printed, the action type identifier indicates continuous extrusion with path turning, and the operating condition switching identifier indicates that the equipment has just switched from outer contour printing to inner fill printing.

[0252] The PMA process monitoring agent performs dynamic time warping on the control command stream and response data stream to generate an aligned residual sequence. Since the nozzle temperature response, motor load response, and structural vibration response have varying degrees of delay relative to the control commands, the PMA process monitoring agent segments the control command stream and response data stream according to a unified time window, resulting in a set of control command time segments and a set of response data time segments. The PMA process monitoring agent extracts action switching features and setpoint change features from the control command time segment set to form a control command feature sequence, and extracts response fluctuation features and response peak features from the response data time segment set to form a response data feature sequence. The PMA process monitoring agent calculates a distance matrix based on the control command feature sequence and the response data feature sequence, and then calculates a cumulative distance matrix based on the distance matrix. Based on the cumulative distance matrix, it determines the mapping path corresponding to the minimum cumulative distance, and then aligns the control command time segment set and the response data time segment set according to the mapping path, resulting in an aligned control command segment sequence and an aligned response data segment sequence. The PMA process monitoring agent calculates the timing difference based on the aligned control command segment sequence and the aligned response data segment sequence to generate an aligned residual sequence.

[0253] During continuous printing, the PMA process monitoring agent further decouples the alignment residual sequence, separating state feature components and generating structured sensing events based on the deviations of these state feature components from the historical normal process distribution. Specifically, the PMA process monitoring agent inputs the alignment residual sequence into the manifold learning module to perform feature decoupling, reducing the dimensionality of the alignment residual sequence to a low-dimensional feature representation. Subsequently, the PMA process monitoring agent extracts the physical dimension of the low-dimensional feature representation, separating thermal hysteresis state feature components, mechanical vibration state feature components, and environmental disturbance state feature components. These components are then combined to generate state feature components. The PMA process monitoring agent inputs these state feature components into a long short-term memory network autoencoder to extract reconstruction errors. The conclusion that the reconstruction error exceeds a distribution threshold set based on the historical normal process distribution is determined as a deviation result. If, within a certain time window, the extrusion motor current continuously increases, the nozzle temperature response lags, and the vibration amplitude increases, the PMA process monitoring agent extracts the state feature components within the current time window, injects anomaly identifiers into the extracted state feature components, and performs combination and encapsulation processing to generate a structured sensing event. The PMA process monitoring agent encapsulates the structured sensing event into an anomaly event message and publishes it to the message broker's anomaly event topic. The attached document also explains that the PMA will publish an anomaly event message to the message broker after identifying a state deviation.

[0254] The EDA device diagnostic agent continuously monitors abnormal event topics. When the EDA device diagnostic agent receives a structured perception event from the PMA process monitoring agent, it inputs the structured perception event and the runtime context into the graph attention neural network layer to obtain an entity embedding vector. This entity embedding vector contains both abnormal state information and operating background information, such as high fill density areas, continuous extrusion, increased current, temperature hysteresis, and enhanced vibration. The EDA device diagnostic agent maps the entity embedding vector to abnormal appearance entity nodes in the mechanistic knowledge graph and establishes these abnormal appearance entity nodes as the starting point for logical deduction. Subsequently, starting from the logical deduction starting point, the EDA device diagnostic agent backtracks along the causal edges in the mechanistic knowledge graph, extracting all potential causal nodes that directly or indirectly point to the logical deduction starting point, forming a set of potential causal nodes. This set of potential causal nodes may include nodes such as partial nozzle blockage, material slippage due to moisture, throat thermal creep, insufficient cooling, and increased extrusion path resistance. The EDA device diagnostic agent, based on the causal edge propagation relationship between the set of potential causal nodes and the starting point of logical deduction, and the prior probabilities of the nodes, performs inverse probabilistic deduction to derive the posterior conditional probability of each potential causal node in the set of potential causal nodes, and identifies the potential causal node with the highest posterior conditional probability as the target physical root cause. If the posterior conditional probability shows that the probability corresponding to throat thermal creep is the highest, the EDA device diagnostic agent identifies throat thermal creep as the target physical root cause and publishes the diagnostic conclusion message to the diagnostic topic of the message broker. The attached document clearly states that EDA listens for abnormal events and combines them with a physical mechanism knowledge graph to perform reverse tracing of the anomalies, outputting a root cause determination result with clear physical semantics.

[0255] The DSA (Decision Support Agent) continuously monitors and diagnoses topics. When the DSA agent receives a diagnostic conclusion that the target physical root cause is throat thermal creep, it reads the material property boundaries and equipment motion boundaries, and constructs physical safety boundary conditions based on these boundaries. The material property boundaries can be formed by a combination of thermal decomposition thresholds, melting temperature constraints, and heat conduction parameters; the equipment motion boundaries can be formed by a combination of maximum acceleration, maximum feed rate, and equipment motion constraint parameters. The DSA agent extracts adjustable parameters based on the target physical root cause, such as nozzle temperature, cooling airflow, extrusion speed, and motion speed, and generates multiple candidate parameter combinations around these parameters. These candidate combinations are then aggregated to form a candidate strategy set. The candidate strategy set may include strategies such as reducing nozzle temperature and increasing cooling airflow, reducing feed rate and maintaining nozzle temperature, and slightly reducing extrusion speed and adjusting motion speed.

[0256] The DSA (Decision Support Agent) substitutes each candidate strategy in the candidate strategy set into the physical safety boundary conditions for boundary verification, eliminating candidate strategies that do not meet the physical safety boundary conditions, thus obtaining a filtered candidate strategy set. If a candidate strategy causes the nozzle temperature to approach the material thermal decomposition threshold, or causes the motion acceleration to exceed the equipment motion boundary, then the candidate strategy is eliminated. For each candidate strategy in the filtered candidate strategy set, the DSA decision support agent calculates the geometric accuracy evaluation value, manufacturing time evaluation value, energy consumption cost evaluation value, and boundary violation penalty evaluation value, respectively, and inputs these values ​​into a value evaluation function to calculate the comprehensive evaluation value corresponding to each candidate strategy. Based on the comprehensive evaluation value corresponding to each candidate strategy, the DSA decision support agent determines the target adjustment strategy from the filtered candidate strategy set. For example, in the case of throat thermal creep, the target adjustment strategy may include moderately reducing the nozzle temperature, increasing the cooling airflow, reducing the extrusion speed, and limiting the acceleration variation. The attached document states that after receiving the diagnostic results, DSA will read physical boundary parameters such as the material's thermal decomposition temperature and maximum allowable acceleration, and perform virtual pre-simulation verification on the candidate strategies under boundary constraints. Once the feasibility is confirmed, a correction instruction will be issued.

[0257] After the target adjustment strategy is determined, the DSA decision support agent inputs the strategy into the virtual simulation engine to deduce the equipment evolution response trajectory during the execution of the strategy, thus completing a physical feasibility pre-simulation. The equipment evolution response trajectory can include predicted temperature, vibration, current, and position trajectories. The DSA agent extracts the predicted physical quantities from the trajectory and verifies that they are within the limits of physical safety boundary conditions, thus completing the thought chain verification. If the predicted temperature does not exceed the material thermal decomposition threshold, the predicted acceleration does not exceed the maximum acceleration, the predicted feed rate does not exceed the maximum feed rate, and the travel distance does not exceed the limits, the target adjustment strategy passes the physical feasibility pre-simulation and thought chain verification. Based on the underlying communication interaction protocol, the DSA agent converts the target adjustment strategy, which has passed the physical feasibility pre-simulation and thought chain verification, into underlying control commands and writes these commands to the equipment control terminal through a communication interface connected to the terminal. The written underlying control commands can manifest as control codes for adjusting nozzle temperature setpoints, adjusting cooling airflow, adjusting extrusion speed, and limiting acceleration.

[0258] After the device control terminal executes the underlying control commands, the operating status of the printing equipment changes. The PMA process monitoring agent continues to collect the new response data stream generated by the execution of the underlying control commands, and uses this new response data stream as the response data stream for subsequent acquisition of control command streams and response data streams. The new response data stream includes new temperature sampling values, vibration sampling values, and current sampling values. If the adjustment is effective, the extrusion motor current in the new response data stream gradually decreases, the nozzle temperature lag decreases, and the vibration amplitude decreases. The PMA process monitoring agent continues to perform dynamic time warping on the new control command stream and the new response data stream, generating a new aligned residual sequence, and judging whether the deviation has decreased based on the new state feature components. If a deviation still exists, the PMA process monitoring agent reissues structured sensing events, the EDA device diagnostic agent re-executes root cause inference, and the DSA decision support agent re-selects strategies, performs pre-run verification, and issues underlying control commands, thus forming a continuously operating collaborative monitoring and closed-loop control process.

[0259] In the above application example, the PMA process monitoring agent is responsible for continuous data acquisition, dynamic time warping, feature decoupling, and structured perception event publishing; the EDA device diagnostic agent is responsible for mechanistic knowledge graph mapping, causal edge backtracking, posterior conditional probability inference, and target physical root cause output; and the DSA decision support agent is responsible for physical safety boundary construction, candidate policy set generation, candidate policy screening, value assessment, physical feasibility simulation, thought chain verification, and low-level control instruction writing. The three types of agents collaborate in an event-driven manner through a message broker, each independently receiving messages and executing corresponding tasks according to its subscribed topics. Because the PMA process monitoring agent, EDA device diagnostic agent, and DSA decision support agent operate independently, the graph inference computation of the EDA device diagnostic agent will not interrupt the continuous acquisition of sensor data by the PMA process monitoring agent, and the virtual simulation process of the DSA decision support agent will not block the high-frequency monitoring of the PMA process monitoring agent. This allows on-site monitoring, root cause diagnosis, and safety decision-making to continuously and collaboratively perform within the same printing task.

[0260] This embodiment inputs the target adjustment strategy into a virtual simulation engine and simulates the equipment's evolution response trajectory, allowing the acquisition of the dynamic physical response corresponding to the adjustment result before execution. By extracting predicted physical quantities from the equipment's evolution response trajectory and performing a thought chain verification based on physical safety boundary conditions, the execution results of the target adjustment strategy can be judged for constraint consistency and response consistency. By converting the verified target adjustment strategy into low-level control instructions according to the underlying communication interaction protocol and writing them to the equipment control terminal through the communication interface, the adjustment result can enter the field control process in an executable format. After this processing, the target adjustment strategy has undergone dynamic pre-simulation and boundary verification before being written to the equipment, which can reduce execution risks, improve the matching degree between control parameters and equipment operating conditions, and enhance the executability and operational stability of the adjustment results.

[0261] In one embodiment, a monitoring and closed-loop control system based on large model and multi-agent collaboration is provided. This monitoring and closed-loop control system corresponds one-to-one with the monitoring and closed-loop control method based on large model and multi-agent collaboration described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the monitoring and closed-loop control system based on a large model and multi-agent collaboration of the present invention. The modules include a dual-stream acquisition module 10, a timing alignment module 20, a decoupling sensing module 30, a root cause inference module 40, a safety decision-making module 50, a pre-simulation distribution module 60, and a feedback feedback module 70. Detailed descriptions of each functional module are as follows:

[0262] The dual-stream acquisition module 10 is used to acquire control command stream and response data stream, and extract the running context based on the control command stream;

[0263] The timing alignment module 20 is used to perform dynamic time warping on the control command stream and the response data stream to generate an aligned residual sequence;

[0264] The decoupling sensing module 30 is used to perform feature decoupling processing on the aligned residual sequence, separate the state feature components, and generate structured sensing events based on the deviation of the state feature components from the historical normal process distribution.

[0265] The root cause inference module 40 is used to map the structured perception event and the running context to entity nodes in the mechanism knowledge graph, and perform inverse probability inference along the causal edges in the mechanism knowledge graph to determine the target physical root cause;

[0266] The safety decision module 50 is used to construct physical safety boundary conditions based on material property boundaries and equipment motion boundaries, generate a candidate strategy set based on the target physical root cause, filter the candidate strategy set using the physical safety boundary conditions, and determine the target adjustment strategy from the filtered candidate strategy set through a value evaluation function.

[0267] The pre-simulation distribution module 60 is used to perform physical feasibility pre-simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility pre-simulation and thought chain verification into low-level control instructions and write them into the device control terminal.

[0268] The feedback feedback module 70 is used to collect the new response data stream generated by executing the underlying control command, and use the new response data stream as a response data stream for subsequent acquisition of control command stream and response data stream.

[0269] For preferred embodiments of the monitoring and closed-loop control system based on large models and multi-agent collaboration, please refer to the foregoing limitations on the monitoring and closed-loop control method based on large models and multi-agent collaboration, which will not be repeated here. Each module in the aforementioned monitoring and closed-loop control system based on large models and multi-agent collaboration 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 a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0270] In one embodiment, a system is provided for implementing a monitoring and closed-loop control method based on a large model and multi-agent collaboration, which can, as... Figure 4As shown, the system comprises three main parts: a cloud evolution layer, an edge computing layer, and a physical execution layer. The cloud evolution layer handles high-performance computing tasks and includes a digital twin evolution environment and a model training center. The digital twin evolution environment constructs a heat-fluid coupling virtual environment corresponding to the manufacturing process and generates various working condition samples within it. The model training center performs high-dimensional time-series model training and knowledge graph construction. The cloud evolution layer extends downwards with information flows for model weight updates and knowledge distillation, indicating that cloud training results are continuously transmitted to the edge, ensuring the edge model maintains consistent updates with the cloud training results. The edge computing layer contains embedded edge terminals, each deploying a PMA (Process Monitoring Agent), an EDA (Equipment Diagnosis Agent), and a DSA (Decision Support Agent). The PMA agent is responsible for sensing and aligning control data with physical feedback data; the EDA agent is responsible for combining knowledge graphs to complete root cause inference of anomalies; and the DSA agent is responsible for outputting regulatory decisions under physical constraints. The three agents are arranged from left to right in the diagram, representing the sequential processing relationship from process monitoring to diagnostic inference and then to decision support. The physical execution layer includes manufacturing equipment and multi-source sensing modules. The manufacturing equipment in the diagram uses an FDM 3D printer as an example; FDM stands for Fused Deposition Modeling. The multi-source sensing modules collect various types of feedback information, such as vibration, temperature, and current, and send the sensing results upwards to the edge computing layer, forming a sensor feedback stream. After completing analysis and decision-making, the edge computing layer sends control commands back to the physical execution layer, enabling the manufacturing equipment to continue operating according to the updated control commands. The control command write-back, G-code command stream, and sensor feedback stream together constitute a closed-loop operation relationship, illustrating the complete implementation architecture of the system from cloud training and edge inference to on-site execution.

[0271] In one embodiment, a PMA agent workflow is provided in a monitoring and closed-loop control method based on a large model and multi-agent collaboration, which can be as follows: Figure 5 As shown. Figure 5The diagram, from left to right, illustrates the processing steps: control data input, response data input, dynamic time warping and alignment, residual vector calculation, feature decoupling, and structured sensing event output. The upper left of the diagram shows the G-code instruction sequence input area, indicating that control commands output from the device control terminal are input as an ideal control sequence to the edge-side processing unit. The G-code content describes control information such as position changes, feeding behavior, and extrusion behavior. The lower left of the diagram shows the sensor time series input area, indicating that physical response data from acquisition units such as temperature and vibration sensors are synchronously input into the analysis process. Because the output rhythm of control commands differs from the sensor sampling rhythm, a dynamic time warping and alignment layer is set up in the middle to stretch, compress, and correspond the control and response sequences on a time scale, establishing a matching relationship between the two types of data under a unified temporal semantic. After time alignment, the diagram enters the residual vector calculation module, indicating that the system calculates the difference between the aligned control and response results to obtain a residual vector that characterizes the control deviation and physical response deviation. The residual vector is then input into the manifold learning feature decoupling module, indicating that the system performs low-dimensional mapping and component separation on multiple state changes in the mixed residuals. The right side of the diagram shows three output branches: thermal hysteresis component, mechanical vibration component, and environmental disturbance component. This illustrates that the processing can decompose and express abnormal changes from different physical sources, such as temperature hysteresis, vibration enhancement, and external environmental fluctuations. The rightmost module is the structured sensing event output module, indicating that after component extraction and anomaly detection, the system encapsulates the anomaly category, anomaly time window, and state feature information into a unified event result output for subsequent diagnostic processes. The entire diagram reflects the complete workflow of the PMA agent from dual-stream input to structured event generation.

[0272] In one embodiment, an EDA agent logic reasoning process is provided in a monitoring and closed-loop control method based on a large model and multi-agent collaboration, which can be as follows: Figure 6 As shown. Figure 6The left side shows the structured perception event input area. The diagram uses temperature anomalies and vibration waveform anomalies as example inputs, representing abnormal event results from the previous stage being sent to the diagnostic processing module. The middle section is the neural symbolic reasoning engine, which includes two processing units: anomaly pattern matching and causal path inference. Anomaly pattern matching maps the input structured perception events to anomaly patterns in the knowledge graph, enabling the current anomaly to establish a matching relationship with historical mechanism patterns, typical fault patterns, and parameter coupling patterns. Causal path inference traces back from the anomaly manifestation to upstream mechanisms and root causes, searching for candidate causes that can explain the current anomaly. The right side of the diagram shows the additive manufacturing fault knowledge graph library, which is organized hierarchically into a physical phenomenon layer, an intermediate mechanism layer, and a fault root cause layer. The physical phenomenon layer stores directly observable anomalies, such as temperature anomalies, increased vibration, or current fluctuations. The intermediate mechanism layer stores the mechanistic processes connecting phenomena and root causes, such as thermal creep, flow blockage, shear anomalies, or mechanical resonance. The root cause layer stores the final cause nodes, such as nozzle blockage, abnormal feeding, material dampness, or localized overheating. The arrows in the diagram indicate that the EDA agent does not merely stop at anomaly identification; instead, it feeds structured perceived events into a neural symbolic reasoning engine, combines causal relationships from a knowledge graph to perform reverse tracing, and finally outputs an interpretable diagnostic report. The output arrows at the bottom of the diagram clearly indicate that the diagnostic results can be directly invoked by downstream decision-making modules.

[0273] In one embodiment, a DSA agent decision-making control process is provided in a monitoring and closed-loop control method based on a large model and multi-agent collaboration, which can be as follows: Figure 7 As shown. Figure 7The inputs at the top are the diagnostic results and the current state, indicating that decision-making is not based solely on the diagnostic root cause, but rather combines the current equipment state, operating condition, and control state to form the decision input. The dashed box in the diagram represents the GRPO strategy generation module. GRPO stands for Group Relative Policy Optimization. This module generates multiple candidate strategies in parallel based on the diagnostic results and the current state. The example in the diagram shows strategies A, B, and C, representing different types of adjustment directions, such as reducing speed, increasing temperature, or maintaining parameter balance. The left side of the diagram shows the physical conservation law constraint input area, illustrating that physical limits such as maximum acceleration and thermal threshold are input into the physical safety boundary checker. The physical safety boundary checker, located below the multiple candidate strategies, filters each strategy, removing inoperable results that exceed the material or equipment allowable ranges. The arrows pointing from multiple strategies to the physical safety boundary checker in the diagram indicate that different candidate strategies need to enter a unified boundary check process. After boundary check, the value assessment and thought chain check module is accessed below. Value assessment compares the overall performance of retention strategies across dimensions such as quality, efficiency, energy consumption, and security. The thought chain verification checks the logical consistency between adjustment actions and device response results, avoiding the generation of strategies that appear reasonable only numerically but may cause anomalies during execution. The output at the bottom of the figure is a G-code instruction, indicating that the results processed above will ultimately be converted into control codes recognizable by the device control end.

[0274] In one embodiment, a virtual-to-physical migration deployment process is provided in a monitoring and closed-loop control method based on a large model and multi-agent collaboration, which can be as follows: Figure 8 As shown. Figure 8The left side represents the cloud evolution layer, corresponding to the high-performance computing side. It includes a digital twin simulation environment and a high-dimensional temporal model. The digital twin simulation environment generates virtual training samples covering various thermal-fluid coupling perturbation conditions, while the high-dimensional temporal model is used for training and representation learning of high-dimensional physical field data in the cloud. The middle section of the diagram shows the model distillation and quantization pruning module, indicating that the high-dimensional model in the cloud is not directly deployed to the field device. Instead, it obtains transferable knowledge through knowledge distillation and then reduces the number of model parameters and computational overhead through quantization pruning, enabling the model to adapt to edge-side resource conditions. The right side of the diagram represents the edge computing layer, corresponding to the resource-constrained side. The deployment result is a lightweight intelligent agent model. This lightweight intelligent agent model is further deployed down to embedded terminal hardware, which then performs local inference and local control of the 3D printer. The bottom of the diagram also shows arrows indicating the feedback of measured data from the 3D printer back to the cloud evolution layer, showing that the measured data does not remain at the edge but is continuously transmitted back to the cloud for updating the simulation environment, correcting training samples, and optimizing model parameters. This not only reflects the migration process from cloud training to edge deployment, but also the evolutionary relationship of the continuous flow of field test results back to the cloud for retraining and recalibration.

Claims

1. A monitoring and closed-loop control method based on a large model and multi-agent collaboration, characterized in that, Includes the following steps: Acquire control command streams and response data streams, and extract the runtime context based on the control command streams; The control command stream and the response data stream are subjected to dynamic time warping to generate an aligned residual sequence; The aligned residual sequence is subjected to feature decoupling processing to separate state feature components, and structured sensing events are generated based on the deviation of the state feature components from the historical normal process distribution. The structured perception event and the runtime context are mapped to entity nodes in the mechanism knowledge graph, and inverse probability inference is performed along the causal edges in the mechanism knowledge graph to determine the target physical root cause; Physical safety boundary conditions are constructed based on material property boundaries and equipment motion boundaries. A set of candidate strategies is generated based on the target physical root cause. The set of candidate strategies is screened using the physical safety boundary conditions. Finally, a target adjustment strategy is determined from the screened set of candidate strategies using a value evaluation function. Perform physical feasibility simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility simulation and thought chain verification into low-level control instructions and write them into the device control terminal. The newly generated response data stream is collected and used as a response data stream for subsequent acquisition of control command stream and response data stream.

2. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, Acquire control command streams and response data streams, and extract the runtime context based on the control command streams, including: Establish command communication connections between the device control terminal and the edge computing terminal, and establish data communication connections between the sensor acquisition terminal and the edge computing terminal; Control commands are continuously read from the device control terminal and buffered in the order of output to form a control command stream; The control command stream is split and parsed to obtain each control command containing a motion control field, an energy control field, and a layer switching control field, and a corresponding time identifier is written for each control command; Temperature, vibration, and current sampling values ​​are read from the sensor acquisition terminal and aggregated according to the sampling time sequence to form a response data stream; Based on the motion control field, energy control field, and layer switching control field in each control instruction with the corresponding time identifier, the current processing segment identifier, action type identifier, and working condition switching identifier are extracted, and the current processing segment identifier, the action type identifier, and the working condition switching identifier are combined to form the operating context.

3. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, Dynamic time warping is performed on the control command stream and the response data stream to generate an aligned residual sequence, including: The control command stream and the response data stream are segmented according to a unified time window to obtain a set of control command time segments and a set of response data time segments. Action switching features and setpoint change features are extracted from the control command time segment set to form a control command feature sequence, and response fluctuation features and response peak features are extracted from the response data time segment set to form a response data feature sequence; A distance matrix is ​​calculated based on the control command feature sequence and the response data feature sequence, and a cumulative distance matrix is ​​calculated based on the distance matrix. The mapping path corresponding to the minimum cumulative distance is determined based on the cumulative distance matrix. Based on the mapping path, the control command time segment set and the response data time segment set are time-aligned to obtain an aligned control command segment sequence and an aligned response data segment sequence. The timing difference is calculated based on the alignment control instruction fragment sequence and the alignment response data fragment sequence to generate an alignment residual sequence.

4. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, The aligned residual sequence is subjected to feature decoupling processing to separate state feature components, and structured sensing events are generated based on the deviation of the state feature components from the historical normal process distribution, including: The aligned residual sequence is input into the manifold learning module to perform feature decoupling processing, and the aligned residual sequence is dimensionality reduced and mapped to a low-dimensional feature representation; The low-dimensional feature representation is physically decomposed and extracted to separate the thermal hysteresis state feature component, the mechanical vibration state feature component, and the environmental disturbance state feature component. The thermal hysteresis state characteristic component, the mechanical vibration state characteristic component, and the environmental disturbance state characteristic component are combined to generate a state characteristic component; The state feature components are input into the long short-term memory network autoencoder to extract the reconstruction error, and the judgment that the reconstruction error exceeds the distribution threshold set based on the historical normal process distribution is determined as a deviation result. Based on the deviation results, the state feature components within the current time window are extracted, anomaly identifiers are injected into the extracted state feature components, and combined encapsulation processing is performed to generate structured perception events.

5. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, Mapping the structured perceived events and the runtime context to entity nodes in a mechanistic knowledge graph, and performing inverse probabilistic inference along the causal edges in the mechanistic knowledge graph to determine the target physical root cause, including: The structured perception event and the running context are input into the graph attention neural network layer to obtain the entity embedding vector; The entity embedding vector is mapped to the abnormal appearance entity node in the mechanism knowledge graph, and the abnormal appearance entity node is established as the starting point of logical deduction. Starting from the logical deduction starting point, trace back the source path along the reverse direction of the causal edge in the mechanism knowledge graph, extract all potential causal nodes that directly or indirectly point to the logical deduction starting point, and form a set of potential causal nodes. Based on the causal edge propagation relationship between the set of potential causal nodes and the starting point of logical deduction, and the prior probability of the nodes, reverse probability deduction is performed to derive the posterior conditional probability of each potential causal node in the set of potential causal nodes, and the potential causal node with the largest posterior conditional probability is determined as the target physical root cause.

6. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, Physical safety boundary conditions are constructed based on material property boundaries and equipment motion boundaries. A candidate strategy set is generated based on the target physical root cause. The candidate strategy set is then screened using the physical safety boundary conditions. Finally, a target adjustment strategy is determined from the screened candidate strategy set using a value evaluation function, including: Read the thermal decomposition threshold, thermal conduction constraint, and melting range constraint from the material property boundary, and read the upper limit of velocity, upper limit of acceleration, and motion stroke constraint from the equipment motion boundary; Based on the thermal decomposition threshold, the thermal conduction constraint, the melting zone constraint, the upper velocity limit, the upper acceleration limit, and the motion stroke constraint, physical safety boundary conditions are constructed. Based on the target physical root cause, extract the parameter items to be adjusted, generate multiple candidate parameter combinations around the parameter items to be adjusted, and gather the multiple candidate parameter combinations to form a candidate strategy set. Substitute each candidate strategy in the candidate strategy set into the physical security boundary conditions for boundary verification, and eliminate candidate strategies that do not meet the physical security boundary conditions to obtain the filtered candidate strategy set. For each candidate strategy in the filtered candidate strategy set, calculate the geometric accuracy evaluation value, manufacturing time evaluation value, energy consumption cost evaluation value, and boundary violation penalty evaluation value respectively. Then, input the geometric accuracy evaluation value, the manufacturing time evaluation value, the energy consumption cost evaluation value, and the boundary violation penalty evaluation value into the value evaluation function to calculate the comprehensive evaluation value corresponding to each candidate strategy. Based on the comprehensive evaluation value corresponding to each candidate strategy, the target adjustment strategy is determined from the set of candidate strategies after screening.

7. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, Perform physical feasibility simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility simulation and thought chain verification into low-level control instructions and write them into the device control terminal, including: The target adjustment strategy is input into the virtual simulation engine to simulate the device evolution response trajectory during the execution of the target adjustment strategy, so as to complete the physical feasibility simulation. Extract the predicted physical quantities from the device evolution response trajectory and verify that the predicted physical quantities are within the range defined by the physical safety boundary conditions to complete the thought chain verification; Based on the underlying communication and interaction protocol, the target adjustment strategy, which is verified through the physical feasibility simulation and the thought chain, is converted into underlying control commands; The underlying control commands are written to the device control terminal through a communication interface connected to the device control terminal.

8. The monitoring and closed-loop control method based on large model and multi-agent collaboration as described in claim 1, characterized in that, Before constructing physical safety boundary conditions based on material property boundaries and equipment motion boundaries, generating a candidate strategy set based on the target physical root cause, filtering the candidate strategy set using the physical safety boundary conditions, and determining the target adjustment strategy from the filtered candidate strategy set using a value evaluation function, the process further includes: Read the thermal decomposition threshold, melting temperature constraint, and thermal conductivity parameters from the material property database; Read the maximum motion acceleration, maximum feed rate, and equipment motion constraint parameters from the equipment specification configuration file; The read thermal decomposition threshold, melting temperature constraint, and thermal conductivity parameters are combined to form the material property boundary; The maximum acceleration, maximum feed rate, and device motion constraint parameters read are combined to form the device motion boundary.

9. A monitoring and closed-loop control system based on a large model and multi-agent collaboration, characterized in that, The monitoring and closed-loop control system based on large models and multi-agent collaboration includes: The dual-stream acquisition module is used to acquire control command streams and response data streams, and extract the runtime context based on the control command streams; The timing alignment module is used to perform dynamic time warping on the control command stream and the response data stream to generate an aligned residual sequence; The decoupling sensing module is used to perform feature decoupling processing on the aligned residual sequence, separate the state feature components, and generate structured sensing events based on the deviation of the state feature components from the historical normal process distribution. The root cause inference module is used to map the structured perception event and the running context to entity nodes in the mechanism knowledge graph, and perform inverse probability inference along the causal edges in the mechanism knowledge graph to determine the target physical root cause; The safety decision module is used to construct physical safety boundary conditions based on material property boundaries and equipment motion boundaries, generate a set of candidate strategies based on the target physical root causes, filter the set of candidate strategies using the physical safety boundary conditions, and determine the target adjustment strategy from the filtered set of candidate strategies through a value evaluation function. The pre-simulation delivery module is used to perform physical feasibility pre-simulation and thought chain verification on the target adjustment strategy, and convert the target adjustment strategy that has passed the physical feasibility pre-simulation and thought chain verification into low-level control instructions and write them into the device control terminal. The feedback feedback module is used to collect the new response data stream generated by executing the underlying control command, and to use the new response data stream as a response data stream for subsequent acquisition of control command stream and response data stream.

Images