Knowledge Graph-Based Mold Opening Machine Fault Early Warning System

Abstract

This invention relates to the field of intelligent manufacturing and industrial equipment fault diagnosis and early warning technology, specifically a knowledge graph-based mold-making machine fault early warning system, comprising: a data acquisition module, which collects time-series data of the operating status of the mold-making machine, multi-field coupling parameters of the mold, and process context data, and connects to an expert knowledge base; an ontology construction module, which constructs a fused knowledge graph composed of an equipment topology subgraph, a spatiotemporal stress subgraph, a process rheology subgraph, and a product quality subgraph, and generates a dynamically instantiated knowledge graph; a penetration simulation module, which calculates the transmission path and evolution state of abnormal features based on a physical information graph neural network; a risk reasoning module, which assesses the severity of the consequences of the affected nodes at the end and the overall risk probability; and a dynamic decision-making module, which generates adaptive intervention strategies, outputs a real-time risk transmission topology graph and intelligent maintenance instructions, and realizes consequence-oriented mold-making machine fault early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial equipment fault diagnosis and early warning technology, specifically a mold opening machine fault early warning system based on knowledge graph. Background Technology

[0002] With the development of continuous molding technology for high-performance composite materials, the mold opening machine, as a key piece of equipment in mold opening and closing, pressure holding and demolding, directly affects mold safety, product quality and the continuity of the entire batch production. In order to achieve reliable operation of the mold opening machine under complex working conditions, real-time monitoring of equipment status, anomaly identification and fault early warning have become important technical requirements in the molding manufacturing field. In early warning of mold making machine failures, most existing methods rely on threshold judgment or trend analysis based on single equipment signals such as vibration, hydraulic pressure, motor current or displacement. Although these methods can detect component abnormalities to a certain extent, the composite material molding process is affected by multiple factors such as mold thermal fatigue, stress concentration, material viscosity changes and curing stage. Relying solely on isolated equipment parameters often makes it difficult to accurately reflect the transmission relationship between the abnormality at the equipment end, mold end and process end. Especially in scenarios such as high temperature holding and continuous batch processing, problems such as false shutdown, missed warnings or inability to judge the severity of consequences are likely to occur. Therefore, it is crucial to ensure equipment safety, mold life, and product yield in the continuous molding process by uniformly processing the timing data of the mold opening machine's operating status, the multi-field coupling parameters of the mold, and the process context information, and by establishing a correlation model that can support risk transmission analysis in combination with the equipment mechanism and material rheological laws, so as to achieve accurate assessment of abnormal consequences and reasonable decision-making on intervention strategies. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a knowledge graph-based fault early warning system for mold-making machines. Specifically, the technical solution of this invention includes: The data acquisition module collects time-series data of equipment operation status, multi-field coupling parameters and process context data of mold, extracts equipment mechanism and material rheology rules from a preset expert knowledge base, and generates a multimodal heterogeneous dataset. The ontology construction module constructs a fusion knowledge graph based on the multimodal heterogeneous dataset, which includes subgraphs of equipment topology, spatiotemporal stress, process rheology, and product quality. It transforms the time-series data of the operating status into dynamic node attributes of the graph and generates a dynamically instantiated knowledge graph. The penetration simulation module converts the material rheology rules into a loss function penalty term to construct a physical information graph neural network model. It calls the dynamically instantiated knowledge graph to identify abnormal nodes whose scalar feature values ​​of attributes deviate from a preset threshold. Based on the model, it calculates the transmission path and evolution state of abnormal node features in the fused knowledge graph, simulates the cross-domain information transmission of mechanical stress and thermodynamic field corresponding to multi-field coupling parameters, and generates risk transmission topology data. The risk reasoning module assesses the severity of the consequences of the affected nodes at the end of the transmission path based on the risk transmission topology data and preset multi-field coupling reasoning rules, determines the comprehensive risk probability of causing the scrapping of the target product and the damage to the core mold, and generates risk reasoning results. The dynamic decision-making module generates an adaptive intervention strategy based on the risk reasoning results and the preset global technical performance optimization principle, outputs a real-time risk transmission topology map, and sends intelligent maintenance instructions to the mold-making machine control system to complete the early warning.

[0004] Optionally, the operating status timing data includes vibration, hydraulic, motor current and displacement sensor data; the multi-field coupling parameters include mold stress concentration points, thermal fatigue and heat capacity parameters; the process context data includes the viscosity of the processed material, curing curve and real-time temperature control data; and the equipment mechanism and material rheology rules include historical maintenance logs, equipment bill of materials and material rheology formulas.

[0005] Optionally, the ontology construction module includes: The graph definition submodule extracts the equipment mechanism and material rheology rules from the multimodal heterogeneous dataset, defines the attributes of equipment entities, mold entities, material entities and product entities and the relationships between entities, and constructs subgraphs of equipment topology, spatiotemporal stress, process rheology and product quality. The graph fusion submodule splices the subgraphs of equipment topology, spatiotemporal stress, process rheology and product quality through entity alignment and relationship extraction to generate the fused knowledge graph. The dynamic instantiation submodule acquires the runtime time-series data, multi-field coupling parameters, and process context data from the multimodal heterogeneous dataset, maps them to the corresponding nodes of the fused knowledge graph according to timestamps, updates the dynamic node attributes of the corresponding nodes, and generates the dynamic instantiation knowledge graph.

[0006] Optionally, the penetration simulation module includes: The node feature extraction submodule calls the dynamically instantiated knowledge graph to identify abnormal nodes whose scalar feature values ​​of the dynamic node attributes deviate from the preset threshold, and extracts the features of the abnormal nodes and the features of their neighboring nodes. The physical constraint loading submodule transforms the material rheological rules in the multimodal heterogeneous dataset into a loss function penalty term, and combines them with the thermodynamic equations constructed based on the multi-field coupling parameters to build the physical information graph neural network model. The information transmission calculation submodule inputs the features of the abnormal node into the physical information graph neural network model, performs feature aggregation and updating along the edges of the fused knowledge graph, calculates the information evolution of the features on the transmission path, and generates the risk transmission topology data.

[0007] Optionally, in the information transmission calculation submodule, the step of calculating the information evolution of features along the transmission path includes: When the material viscosity parameter between nodes is greater than the preset viscosity threshold, it is determined that the feature of the abnormal node is rheologically buffered. The product of the material viscosity parameter and the preset damping coefficient is used as the attenuation index. The negative power of the attenuation index is calculated with the natural constant as the base, and multiplied by the feature of the abnormal node to obtain the feature attenuation amount. When the material viscosity parameter between nodes is less than or equal to a preset viscosity threshold and the thermal fatigue of the target node is greater than a preset critical point, it is determined that the feature of the abnormal node is amplified by thermal stress. The product of the thermal fatigue of the target node and the preset stress concentration factor is used as the feature amplification factor, and multiplied with the feature of the abnormal node to obtain the amplified feature quantity. When the material viscosity parameter between nodes is less than or equal to a preset viscosity threshold, and the thermal fatigue of the target node is less than or equal to a preset critical point, the characteristic linear transmission of the abnormal node is determined, and the characteristic magnitude remains unchanged. The material viscosity parameter is included in the process context data, the target node is a node on the conduction path, and the thermal fatigue degree is included in the multi-field coupling parameters.

[0008] Optionally, the risk reasoning module includes: The transmission path extraction submodule extracts the complete propagation link from the abnormal node to the affected terminal node based on the risk transmission topology data. The rule matching submodule calls the preset multi-field coupled inference rule and inputs the node degradation degree, mold thermal fatigue degree and current process state on the complete propagation link into the rule engine for condition matching. The node degradation degree is the ratio of the absolute value of the difference between the scalar feature value of the abnormal node and the preset safe interval boundary value to the corresponding boundary value deviated from. The mold thermal fatigue degree and the current process state are obtained by parsing the multimodal heterogeneous dataset. The consequence assessment submodule obtains the corresponding basic risk coefficient according to the matching rules, and uses the product of the basic risk coefficient, the degree of node deterioration, and the normalized thermal fatigue degree of the mold as the comprehensive risk probability of target product scrapping and core mold damage, and generates the risk reasoning result that includes the severity of the consequence and the comprehensive risk probability.

[0009] Optionally, the dynamic decision-making module includes: The strategy generation submodule, based on the risk reasoning results, compares the comprehensive risk probability with the preset risk tolerance, and combines the current batch product availability and mold hardware recovery parameters obtained from the process context data parsing to generate the adaptive intervention strategy; The topology visualization submodule graphically renders the abnormal nodes, propagation links, and severity of consequences in the risk transmission topology data, and outputs the real-time risk transmission topology map. The instruction issuing submodule converts the adaptive intervention strategy into equipment control signals and sends the intelligent maintenance instruction to the mold opening machine control system.

[0010] Optionally, the step of generating the adaptive intervention strategy in the strategy generation submodule includes: When the overall risk probability exceeds a preset severe fault threshold, an emergency shutdown strategy is generated. When the overall risk probability is less than or equal to the preset severe failure threshold, and the severity of the consequences indicates that the risk of product scrapping is greater than the preset yield threshold, a parameter compensation strategy is generated to adjust the mold closing parameters and buffer vibration. When the overall risk probability is less than or equal to a preset severe failure threshold, and the severity of the consequences indicates that the risk of product scrapping is less than or equal to a preset yield rate threshold, a delayed maintenance strategy is generated to reduce the operating speed until the end of the current batch before repair.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This system acquires time-series data of equipment operation status, multi-field coupling parameters and process context data of mold through the data acquisition module, and generates a dynamic instantiated knowledge graph containing equipment topology, spatiotemporal stress, process rheology and product quality subgraphs through the ontology construction module. Its beneficial effects are: it breaks the limitation of traditional mold opening machine fault early warning relying only on isolated equipment signals such as vibration and pressure, and incorporates process states such as equipment operation deterioration, mold thermal fatigue and material viscosity into the graph structure, realizing the association and alignment of multimodal heterogeneous data under complex working conditions, and providing a reliable data foundation for accurately tracing the source of anomalies and cross-domain impacts. 2. This system employs a penetration simulation module to transform material rheological rules into a loss function penalty term to construct a physical information graph neural network model. Based on material viscosity parameters and mold thermal fatigue, it calculates the information evolution of abnormal features under three scenarios: rheological buffering, thermal stress amplification, or linear transmission. Its advantages include: overcoming the shortcomings of purely data-driven models that easily deviate from actual physical laws; enabling the system to realistically simulate the cross-domain transmission behavior of abnormal parameters in mechanical stress and thermodynamic fields; and accurately capturing the dynamic process of abnormal features being buffered or nonlinearly amplified, effectively reducing warning distortion caused by neglecting thermodynamic and materials science factors. 3. This system utilizes a risk reasoning module to extract the complete propagation chain from the risk transmission topology data. It then inputs the node degradation level, mold thermal fatigue degree, and current process status into a rule engine for multi-field coupled reasoning, outputting reasoning results that include the severity of consequences and the overall risk probability. Its beneficial effects are: upgrading traditional component-level isolated alarms to consequence warnings oriented towards the final impact; the system can not only detect anomalies, but also accurately deduce the overall probability that the anomaly will lead to the scrapping of the target product or damage to the core mold, enabling operators to clearly grasp the true severity of the current anomaly, and solving the problem of difficulty in judging potential catastrophic consequences in complex continuous production. 4. This system relies on a dynamic decision-making module. Based on risk reasoning results and the batch product availability and mold hardware recovery parameters parsed from process context data, it intelligently generates adaptive intervention strategies such as emergency shutdown, parameter compensation, or delayed maintenance according to the comprehensive risk probability and severity of consequences. Its beneficial effects are: overcoming the technical defects of the past, which caused material solidification imbalance in the mold cavity and the degradation and scrapping of the entire batch structure due to immediate shutdown at the first sign of an anomaly; the system can actively issue parameter adjustment or speed reduction commands within the tolerance boundary to buffer the extension of anomalies, and maximize the maintenance of the production cycle and product yield of the continuous molding process while ensuring the safety of the core mold and key equipment. Attached Figure Description

[0012] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] like Figure 1 As shown, the knowledge graph-based mold-making machine fault early warning system includes: The data acquisition module collects time-series data of equipment operation status, multi-field coupling parameters and process context data of mold, extracts equipment mechanism and material rheology rules from a preset expert knowledge base, and generates a multimodal heterogeneous dataset. The ontology construction module builds a fusion knowledge graph based on a multimodal heterogeneous dataset, which includes subgraphs of equipment topology, spatiotemporal stress, process rheology, and product quality. It transforms the time-series data of the operating status into dynamic node attributes of the graph and generates a dynamically instantiated knowledge graph. The penetration simulation module converts material rheological rules into loss function penalty terms to construct a physical information graph neural network model. It calls dynamically instantiated knowledge graphs to identify abnormal nodes whose scalar feature values ​​of attributes deviate from preset thresholds. Based on this model, it calculates the transmission path and evolution state of abnormal node features in the fused knowledge graph, simulates the cross-domain information transmission of mechanical stress and thermodynamic field corresponding to multi-field coupling parameters, and generates risk transmission topology data. The risk reasoning module assesses the severity of consequences for affected nodes at the end of the transmission path based on risk transmission topology data and preset multi-field coupling reasoning rules, determines the comprehensive risk probability of causing the scrapping of the target product and damage to the core mold, and generates risk reasoning results. The dynamic decision-making module generates adaptive intervention strategies based on risk reasoning results and the preset global technical performance optimization principle, outputs a real-time risk transmission topology map, and sends intelligent maintenance instructions to the mold making machine control system to complete the early warning.

[0015] This embodiment provides a knowledge graph-based fault early warning mechanism for mold-making machines. Specifically, for the continuous molding scenario of high-performance composite material shells, the mold-making machine is performing high-temperature pressure holding and subsequent mold-making processes for the same batch of products. The mold has accumulated multiple thermal cycles. If the machine is stopped based solely on a single equipment malfunction, it may easily cause material solidification imbalance within the mold cavity and lead to the degradation and scrapping of the entire batch of structures. If operation continues blindly, it may cause core inserts to fracture under the superposition of thermal fatigue and stress. Therefore, this embodiment integrates equipment degradation events and multi-field states of mold and material rheological states of process into a dynamic knowledge graph, and assesses whether it will evolve into technical consequences such as product degradation or mold damage along the graph propagation path. Specifically, the data acquisition module continuously collects operational status timing data from the mold-making machine, acquires multi-field coupling parameters from the mold, and collects process context data from the process control end. The operational status timing data can be written to a buffer at fixed sampling periods, for example... Using a time window, the root mean square value of vibration, hydraulic pressure fluctuation, average motor current, and displacement deviation are summarized in a windowed manner to form a time-based summary. Device feature vectors The multi-field coupling parameters at the mold end can include the current temperature of the hot spot area of ​​the mold, the thermal fatigue index of the cavity, the stress concentration area number and its strength coefficient; the process context data can include the current viscosity of the processed material, the curing progress, the pressure holding stage identifier and the real-time temperature control curve; at the same time, the system also connects to the expert knowledge base to extract equipment mechanism rules, material rheology rules, historical maintenance association rules and equipment component hierarchical relationships; The ontology construction module constructs a fused knowledge graph based on the aforementioned multimodal heterogeneous dataset. The fused knowledge graph includes at least four types of subgraphs: an equipment topology subgraph expressing the connection relationships between hydraulic valves, cylinders, connecting rods, sliders, and mold mounting bases; a spatiotemporal stress subgraph expressing the spatial and stress relationships between cavity regions, inserts, hot spots, and cooling circuits; a process rheology subgraph expressing the process relationships between material viscosity, curing stage, temperature control zone, and flow channel resistance; and a product quality subgraph expressing the mapping relationships between product dimensional accuracy, surface defects, and scrap criteria. After fusion, the graph can be represented as a graph structure, with the node set containing equipment entities, mold entities, material entities, and product entities, and the edge set containing structural connection edges, spatiotemporal influence edges, process dependency edges, and quality mapping edges. Furthermore, the ontology construction module maps the real-time collected time-series data into the dynamic attributes of nodes; for example, the dynamic attributes of node hydraulic cylinder 1 include pressure fluctuation coefficient and vibration intensity; the dynamic attributes of node cavity hotspot A include surface temperature and thermal fatigue degree; the dynamic attributes of node material batch B include real-time viscosity and degree of curing; during the graph instantiation process, time-stamped data is written to the corresponding node and the state of the previous moment is retained to form a dynamically instantiated knowledge graph. The penetration simulation module performs risk transmission simulation based on a physical information graph neural network. Specifically, it first identifies anomalous nodes whose dynamic properties deviate from a threshold. For example, when the rate of change of pressure fluctuation of a hydraulic valve node relative to the baseline reference pressure exceeds a preset pressure anomaly judgment threshold, the hydraulic valve is marked as an anomalous node. The material rheological rules are transformed into a physical penalty term in the loss function and jointly trained with the supervised loss of the graph neural network. The total loss can take the following form: The total loss is calculated using the prediction error term of the abnormal node evolution state. Conduct basic supervision and introduce weighted coefficients. Adjusted penalty term used to constrain the consistency between viscosity change and propagation attenuation trend and through weighting coefficients Adjusted penalty term used to constrain the consistency of temperature field, thermal fatigue degree, and stress amplification trend. The graph neural network aggregates information at each layer by edge, that is, it uses a preset trainable parameter weight matrix to perform linear transformation on the feature vectors of the target node and its neighboring node sets respectively, and combines the propagation weights calculated by edge type, material viscosity and thermal stress state to perform weighted summation of the features of neighboring nodes, and outputs the feature representation of the next layer through activation function mapping; in this way, the transmission path and evolution state of abnormal features from the equipment domain to the mold domain and then to the product quality domain can be obtained. The risk reasoning module does not directly output whether a fault has occurred, but rather assesses the severity of the consequences based on the risk propagation topology data. The system extracts the propagation path from the anomaly source node to the affected end node, such as hydraulic valve pressure fluctuation → mold closing cylinder response lag → slider displacement slight deviation → local thermal stress increase in the cavity → product edge dimension deviation. It calls multi-field coupling reasoning rules to jointly judge the degree of node deterioration, mold thermal fatigue, and current process status, and obtains the comprehensive probability of the target product scrap risk and the core mold damage risk. After obtaining the risk inference results, the dynamic decision-making module does not default to shutdown, but generates an adaptive intervention strategy based on the principle of optimal global technical performance. This principle of optimal global technical performance considers at least three dimensions of technical impact: the availability metric of the current batch of products, i.e., the specific technical manifestation of the product's technical availability in the process context data; the hardware reconstruction resource consumption of the mold, i.e., the specific technical manifestation of the mold hardware recovery parameters; and the expected risk degradation introduced by continuous operation. The system can express the expected technical loss function as the product of the overall risk probability and the impact of single product degradation and mold deterioration, plus the production process disruption loss caused by downtime; its specific mathematical expression is as follows: ,in, Expected technology loss, To consider the overall risk probability, To determine the extent of the impact of a single product downgrade, To determine the degree of impact of mold deterioration, The production process disruption and loss factor caused by downtime; the above , , and The system adopts a unified dimensional system; the system compares the expected technical loss values ​​under different strategies, selects the intervention strategy with the smallest value, and simultaneously outputs a real-time risk transmission topology diagram and intelligent maintenance instructions. As a fault-tolerant measure, if sensor data is missing at a certain moment, this embodiment prioritizes using the most recent valid value for short-term retention. When the continuous missing time exceeds the preset tolerance period, the corresponding node attribute is marked as low confidence, and its propagation weight is reduced during graph propagation to prevent misjudgment caused by abnormal collection. If an entity cannot be aligned in the knowledge graph, it is first attached as a temporary node to the corresponding equipment segment or process segment, and the graph is reconstructed after subsequent data is completed. If the risk transmission path is not unique, the first preset number of paths with the highest cumulative propagation intensity are retained in the risk reasoning module to avoid long-tail paths interfering with real-time decision-making. During the pressure holding stage of the 18th product in a certain batch of composite material shells, the hydraulic valve node detected a pressure fluctuation deviation higher than the judgment threshold and was therefore identified as an abnormal node. After the graph propagation, it was found that when the fluctuation was transmitted to the cavity hot spot A through the mold closing mechanism, the propagation characteristics were amplified because the current mold thermal fatigue degree was greater than the preset critical point and the material viscosity had dropped to below the preset viscosity threshold. This amplified the propagation characteristics and ultimately mapped them to the edge warping risk node and insert damage risk node in the product quality subgraph. Based on this, the system no longer outputs a single equipment alarm, but instead gives an executable conclusion that the remaining workpieces in the current batch can be completed under the condition of deceleration compensation, otherwise direct blocking will lead to the material imbalance and scrapping of the entire mold cavity. The purpose of this step is to elevate equipment failure early warning from isolated component judgment to consequence-oriented cross-domain reasoning, enabling the system to distinguish between anomalies that can continue to be compensated for under local parameter deviations and catastrophic precursors that must be stopped immediately, in high-performance mold and continuous production batch scenarios, thereby achieving failure early warning that is more in line with actual production. In this embodiment, the operating status timing data includes vibration, hydraulic, motor current and displacement sensor data; the multi-field coupling parameters include mold stress concentration points, thermal fatigue and heat capacity parameters; the process context data includes the viscosity of the processed material, curing curve and real-time temperature control data; and the equipment mechanism and material rheology rules include historical maintenance logs, equipment bill of materials and material rheology formulas.

[0016] The ontology building module includes: The graph definition submodule extracts the equipment mechanism and material rheology rules from the multimodal heterogeneous dataset, defines the attributes of equipment entities, mold entities, material entities and product entities and the relationships between entities, and constructs subgraphs of equipment topology, spatiotemporal stress, process rheology and product quality. The graph fusion submodule stitches together equipment topology, spatiotemporal stress, process rheology and product quality subgraphs through entity alignment and relation extraction to generate a fused knowledge graph. The dynamic instantiation submodule acquires the runtime time-series data, multi-field coupling parameters, and process context data from the multimodal heterogeneous dataset, maps them to the corresponding nodes of the fused knowledge graph according to the timestamp, updates the dynamic node attributes of the corresponding nodes, and generates a dynamically instantiated knowledge graph.

[0017] The penetration simulation module includes: The node feature extraction submodule calls the dynamically instantiated knowledge graph to identify abnormal nodes whose scalar feature values ​​of dynamic node attributes deviate from a preset threshold, and extracts the features of abnormal nodes and their neighboring nodes. The physical constraint loading submodule transforms the material rheological rules in the multimodal heterogeneous dataset into loss function penalty terms, and combines them with thermodynamic equations constructed based on multi-field coupling parameters to build a physical information graph neural network model. The information transmission calculation submodule inputs the features of abnormal nodes into the physical information graph neural network model, performs feature aggregation and updating along the edges of the fused knowledge graph, calculates the information evolution of features along the transmission path, and generates risk transmission topology data.

[0018] This embodiment provides a comprehensive processing mechanism for the propagation of multimodal heterogeneous data to dynamically instantiated knowledge graphs and then to physical constraints. Specifically, in the aforementioned continuous molding scenario, if only equipment vibration or pressure curves are saved without simultaneously introducing mold stress concentration points, material viscosity, and curing curves, the graph can only reflect where the machine is abnormal, but cannot explain why the abnormality becomes dangerous under the current mold and the current batch. To address this deficiency, this embodiment combines the data boundaries, graph construction methods, and penetration simulation processes of the previous embodiment, giving the anomaly propagation a clear data source, graph structure carrier, and physical constraint foundation. Specifically, during data acquisition, signals from different sources are aligned using a unified time reference; based on time... For example, the frequency domain characteristics of the vibration sensor output. Hydraulic circuit output pressure characteristics Motor current output load characteristics Displacement sensor output position error characteristics The above features are preprocessed and then concatenated into a device state vector. The synchronous multi-field coupling vector of the mold can be expressed as: This vector is determined by the intensity at the stress concentration point. Local heat capacity parameters and thermal fatigue state Composition; simultaneously extracting the viscosity of the included materials. Curing curve stage parameters and real-time temperature control process context vector Historical maintenance logs, equipment bills of materials, and material rheology formulas serve as inputs to the structured knowledge base. They are not updated on a second-by-second basis, but they participate in the definition of the atlas ontology and rule mapping. The knowledge graph definition submodule first determines the equipment entities and their levels based on the equipment mechanism, such as hydraulic system—hydraulic valve—mold closing cylinder—connecting rod—slider; then it defines the mold entities based on the results of computer-aided engineering analysis of the mold, such as cavity hotspot A—insert B—cooling circuit C; it defines material entities and their associated states based on material rheology knowledge, such as material batch—viscosity range—curing stage; and it defines product entities, such as shell edge area—dimensional deviation—surface defect; the relationships between entities at least include being connected to, acting on, controlled by, affecting quality indicators, and being in the process stage; after obtaining these four subgraphs, they are spliced ​​together into a fused knowledge graph through entity alignment and relationship extraction; When aligning entities, to avoid map breaks caused by inconsistent naming across different data sources, a dual-path strategy of rule priority and similarity supplementation can be adopted. If the mold-closing cylinder 1 in the equipment bill of materials and the main mold-closing cylinder in the maintenance log have the same component code, they are directly aligned; if a unified code is lacking, alignment is based on name similarity. Location similarity and topological similarity And combined with their respective weight allocation coefficients , and Calculate the overall similarity: in, and These represent the two entities to be aligned; when the overall similarity... Greater than or equal to the preset entity alignment similarity threshold When they are identified as the same entity, the equipment topology sub-graph, spatiotemporal stress sub-graph, process rheology sub-graph, and product quality sub-graph can be combined into a unified graph structure. Dynamically instantiated submodules will be timestamped Map them to the corresponding nodes respectively; for example, at time... The hydraulic pressure fluctuation value is written to the hydraulic valve 1 node, the hot spot temperature and thermal fatigue degree are written to the cavity hot spot A node, and the current viscosity and curing stage are written to the material batch B node. If a certain process parameter is constant between two sampling times, the same attribute value is maintained along that time period. If the sensor refresh frequency is different, the time reference of the highest sampling frequency is used for resampling or interpolation so that all nodes have comparable dynamic attributes at the same time. The node feature extraction submodule identifies abnormal nodes on a dynamic graph; based on nodes For example, its scalar properties can be... and safety threshold range In comparison, when or At that time, the deviation was calculated and extracted as the anomaly intensity. : Among them, when hour, ;when hour, ;when hour, ; Indicates the first Each sampling time, This is the lower boundary of the preset safety threshold range. The upper boundary of the preset safety threshold range is defined as follows: when the value is within the safety threshold range, the anomaly intensity is 0. The features of the node and its neighboring nodes are extracted to form a local subgraph input; the physical constraint loading submodule writes the material rheological rules and thermodynamic equations into the training objective of the graph neural network; for the consistency between material viscosity and propagation attenuation, it can be based on the corresponding edges learned by the model. propagation weight Material viscosity parameters associated with nodes And the target weight trend function based on material viscosity given by the material rheology rules. Set penalty items: in, To integrate the edge set in the knowledge graph; for the thermodynamic field, constraint terms can be given based on the preset thermal equilibrium or thermal diffusion equations to limit the evolution relationship between temperature gradient, heat capacity, and thermal fatigue. Specifically, the thermodynamic penalty term is defined as follows: ,in, To integrate the node sets in the knowledge graph, For nodes Temperature characteristic value, Let be the magnitude of the temperature gradient at node i. These are local heat capacity parameters. Based on thermal fatigue The preset thermodynamic evolution objective function, and this function The dimensions of the output are constrained to be the same as those of the output. The dimensions are consistent, thus avoiding propagation results in the model that violate the process mechanism; The information transmission calculation submodule inputs the abnormal node features into the physical information graph neural network model and performs feature aggregation along the edges of the graph; for ease of disclosure, a simplified calculation model can be used for illustration: assuming a hydraulic valve node... The abnormal characteristic amplitude is Its adjacent mold closing cylinder node The propagation weight is And nodes To the hot spot node of the cavity The propagation weight is affected by both material viscosity and thermal state. Then after one propagation, the node The received anomalous evolution can be approximated as: If there are parallel paths from the link node to the slider node and then to the cavity hotspot node, the propagation amounts of each path can be accumulated or weighted to obtain the total risk activation level of the hotspot. As a fault-tolerance mechanism, if the maintenance record of a certain component is missing in the historical maintenance log, the graph definition can still create the node based on the equipment bill of materials and topology, but its prior fault tendency weight is reset to the default value; if the entity alignment similarity is lower than the threshold, it will not be forcibly merged, but a new independent node will be added and a suspected weak relationship edge will be established, and a re-determination will be made after more evidence arrives; if the deviation between the output propagation result and the thermodynamic constraint of the graph neural network on a certain batch of materials is greater than the preset deviation threshold, a re-verification will be triggered, and the conservative conclusion of the rule side will be given priority to avoid outputting control commands with excessive deviation due to model drift in critical operating conditions; In the same batch of composite material shell production, the system collected data showing increased vibration characteristics of the hydraulic valve, an upward trend in motor current that did not exceed the rated safety range, and a shift in displacement error. Simultaneously, the heat capacity parameter of mold hotspot A was lower than the preset benchmark value, thermal fatigue was close to the critical value, and the material viscosity decreased due to prolonged heat preservation time. After the atlas was defined, the hydraulic valve, mold closing cylinder, slider, cavity hotspot A, and product edge dimension nodes were connected in series as a main path. After dynamic instantiation, the node feature extraction submodule discovered that the hydraulic valve node abnormally deviated from the threshold. The physical constraint loading submodule corrected the edge propagation weights based on the current material rheological rules and thermodynamic constraints. Finally, the information transmission calculation showed that the abnormal evolution of cavity hotspot A was significantly higher than that of the regular batch, thus providing interpretable risk transmission topology data for subsequent risk reasoning. The purpose of this step is to unify data standards, clarify ontological relationships, and load physical constraints so that the propagation of the graph no longer stops at empirical edge connections, but forms a dynamic mechanism mapping that can penetrate the equipment domain, mold domain, and process domain, thereby achieving a more reliable prediction of abnormal consequences. In this embodiment, the step of calculating the information evolution of a feature along the transmission path in the information transmission calculation submodule includes: When the material viscosity parameter between nodes is greater than the preset viscosity threshold, the characteristics of the abnormal node are determined to be rheologically buffered. The product of the material viscosity parameter and the preset damping coefficient is used as the attenuation index. The negative power of the attenuation index is calculated with the natural constant as the base, and then multiplied with the characteristics of the abnormal node to obtain the characteristic attenuation amount. When the material viscosity parameter between nodes is less than or equal to the preset viscosity threshold and the thermal fatigue of the target node is greater than the preset critical point, it is determined that the feature of the abnormal node is amplified by thermal stress. The product of the thermal fatigue of the target node and the preset stress concentration factor is used as the feature amplification factor, and multiplied with the feature of the abnormal node to obtain the amplified feature quantity. When the material viscosity parameter between nodes is less than or equal to the preset viscosity threshold, and the thermal fatigue of the target node is less than or equal to the preset critical point, the characteristic linear propagation of the abnormal node is determined, and the characteristic magnitude remains unchanged. The material viscosity parameter is included in the process context data, the target node is the node on the conduction path, and the thermal fatigue degree is included in the multi-field coupling parameters.

[0019] The risk reasoning module includes: The transmission path extraction submodule extracts the complete propagation link from the abnormal node to the affected node at the end based on the risk transmission topology data. The rule matching submodule calls the preset multi-field coupled inference rules and inputs the node degradation degree, mold thermal fatigue degree and current process state on the complete propagation link into the rule engine for condition matching. The node degradation degree is the ratio of the absolute value of the difference between the scalar feature value of the abnormal node and the preset safe interval boundary value to the corresponding boundary value deviated from. The mold thermal fatigue degree and the current process state are obtained by parsing the multimodal heterogeneous dataset. The consequence assessment submodule obtains the corresponding basic risk coefficient based on the matching rules. The product of the basic risk coefficient, the degree of node deterioration, and the normalized mold thermal fatigue is used as the comprehensive risk probability of the target product scrapping and the core mold damage, generating a risk reasoning result that includes the severity of the consequences and the comprehensive risk probability.

[0020] This embodiment provides a mechanism linking information evolution calculation with consequence risk reasoning. Specifically, in the previous embodiment, the graph neural network was able to output the risk propagation topology, but if only the propagation strength is abstracted, there may still be two shortcomings at the review and implementation levels: first, there is a lack of verifiable branch logic as to why anomalies are buffered or amplified; second, after propagation to the product end, how to quantify the risk probability as whether it can continue to operate or must be shut down still requires a clear reasoning process. To solve this problem, this embodiment combines the three-branch information evolution rule of the previous embodiment with the consequence reasoning rule. Specifically, the information transmission calculation submodule does not uniformly use a fixed weight on each transmission edge, but instead performs branching processing based on the material viscosity parameter and the thermal fatigue degree of the target node; The first scenario is rheological buffering; when the viscosity parameter of the material between nodes is greater than a preset viscosity threshold, it is assumed that the melt or processed material has an absorption effect on mechanical fluctuations and displacement disturbances; in this case, the attenuation amount after the abnormal feature is transmitted to the target node is the product of the original feature amplitude and the power calculated with the natural constant as the base and the attenuation exponent as the negative exponent; for example, suppose the feature amplitude of the abnormal source node is The material viscosity parameter is The damping coefficient is After the attenuation exponent is calculated, the attenuation amount is... Multiply by the negative of the natural constant The power of, approximately The results indicate that the anomaly is significantly buffered by the action of high-viscosity materials. The second scenario involves thermal stress amplification. When the material viscosity parameter is less than or equal to a preset viscosity threshold and the thermal fatigue degree of the target node is greater than a preset critical point, it is assumed that the material no longer provides sufficient buffering, and local thermal fatigue of the mold makes the anomaly more likely to evolve into structural damage. In this case, the propagation result is the product of the original anomaly characteristic and the amplification factor. For example, suppose the characteristic amplitude of the anomaly source node is... The stress concentration factor is Thermal fatigue degree is Then the amplified feature quantity is Multiply ,equal This result indicates that although the abnormal source at the equipment end itself does not exceed the second preset threshold, it will be nonlinearly amplified in the critical mold region of thermal fatigue. The third scenario is linear transmission; when the material viscosity parameter and thermal fatigue degree do not reach the above branch activation conditions, it is considered that there is neither obvious rheological buffering nor thermal amplification conditions. At this time, the anomaly is transmitted in a linear manner, keeping the characteristic magnitude unchanged. Based on the above branching rules, the terminal evolution quantity on a complete propagation link can be obtained by recursively multiplying the characteristic amplitude of the starting node with the propagation factor of each side along the way. If there are multiple paths from the abnormal node to the affected terminal node, the risk reasoning module extracts the complete propagation link and can obtain the comprehensive activation degree of the terminal node by multiplying the evolution quantity of each single propagation path by the confidence weight of that path and then summing them. The risk reasoning module extracts the complete propagation link from the risk propagation topology using the propagation path extraction submodule. Taking a certain anomaly as an example, the link can be represented as hydraulic valve 1 → mold closing cylinder → slider → cavity hotspot A → product edge dimension node. Then, the rule matching submodule calculates the node degradation degree of the anomaly node. The node degradation degree is the ratio of the absolute value of the difference between the scalar feature value of the anomaly node and the preset safety interval boundary value to the boundary value. For example, if the hydraulic fluctuation value is 11.2, since it exceeds the upper boundary, the upper boundary value is taken as 10 for calculation, and the degradation degree is 0.12. This degradation degree, along with the mold thermal fatigue degree and the current process status, is input into the rule engine. The rule engine can operate using a condition matching-basic risk coefficient output method; for example, one of the preset rules is: when the node degradation degree is within the first preset parameter range, the mold thermal fatigue degree is greater than the preset critical value, and the current process is in the high-pressure mold closing or high-temperature mold demolding stage, the severe mold structure failure risk rule is hit, and the basic risk coefficient is output. When the degree of node degradation is less than the preset mild degradation threshold, the thermal fatigue degree is within the preset normal operating range, and the process is in the final stage of pressure holding, the product micro-deformation tendency rule is met, and the basic risk coefficient is output. The consequence assessment submodule calculates the overall risk probability based on the hit rules. This is achieved by multiplying the corresponding basic risk coefficient, the degree of node degradation, and the normalized mold thermal fatigue degree. If both product scrap risk and core mold damage risk are considered simultaneously, their respective overall risk probability components can be calculated separately and then synthesized according to project importance, i.e., using a weighted formula. Calculations are performed, in which, As a probability component of product scrap risk, The core mold damage risk probability component, and These are the weighting coefficients for the importance of the corresponding product and mold engineering, and they satisfy the following conditions: This forms a comprehensive risk probability; As a fault-tolerant measure, if multiple rules are hit simultaneously on the same propagation path, the rule with the more severe consequences at the evaluation level is selected first, or the maximum risk probability is used as the safe output. If the boundary value is zero or missing when calculating the degree of node degradation, the historical average or rated value of the indicator is used as the normalization benchmark to prevent abnormal calculation errors in the denominator. If the material viscosity is close to the threshold and accompanied by measurement noise, a hysteresis interval can be set. Within the hysteresis interval, the average result of the features in the last few frames is used as the judgment basis to avoid repeated jumps in the state of buffering and amplification branches. If the propagation path is long and some intermediate node states are missing, a preset low confidence parameter is assigned to the missing segment and the corresponding path weight is reduced. In the aforementioned batch of composite material housings, the initial characteristic amplitude of the hydraulic valve anomaly source node was: When this feature propagates to the mold clamping cylinder, because the material viscosity is greater than the corresponding preset rheological buffer threshold, the rheological buffer condition is met. Therefore, after negative exponential decay calculation, the feature quantity is reduced to approximately When the stress continues to propagate to hot spot A in the cavity, the viscosity decreases in the later stage of pressure holding and the thermal fatigue of the hot spot has exceeded the critical point, satisfying the thermal stress amplification condition. The characteristic value is amplified to approximately [value missing]. After path extraction, the rule engine identifies a combination of moderate equipment malfunction, high thermal fatigue, and proximity to the mold opening process, and outputs a basic risk coefficient. If the degradation level of the abnormal node is Normalized thermal fatigue degree is The overall risk probability, calculated by multiplying together, is approximately... If the quality evaluation weight corresponding to product scrap is higher than the hardware degradation weight corresponding to mold damage, the system can further break down the severity of the consequences so that the subsequent decision-making module can choose to run compensation or interrupt execution. The purpose of this step is to further transform the graph propagation results into evolutionary quantities with physical interpretation and risk probabilities with engineering feasibility, so that the risk conclusions are not only calculable, but also explain why the anomaly will evolve or converge under this condition, thereby achieving more reliable consequence-driven early warning. In this embodiment, the dynamic decision-making module includes: The strategy generation submodule, based on the risk reasoning results, compares the comprehensive risk probability with the preset risk tolerance, and combines the availability of the current batch of products and the mold hardware recovery parameters obtained from the process context data parsing to generate an adaptive intervention strategy. The topology visualization submodule graphically renders the abnormal nodes, propagation links, and severity of consequences in the risk transmission topology data, and outputs a real-time risk transmission topology map. The instruction issuing submodule converts the adaptive intervention strategy into equipment control signals and sends intelligent maintenance instructions to the mold making machine control system.

[0021] In the strategy generation submodule, the steps for generating an adaptive intervention strategy include: When the overall risk probability exceeds the preset severe fault threshold, an emergency shutdown strategy is generated. When the overall risk probability is less than or equal to the preset severe failure threshold, and the severity of the consequences indicates that the risk of product scrapping is greater than the preset yield threshold, a parameter compensation strategy is generated to adjust the mold closing parameters and buffer vibration. When the overall risk probability is less than or equal to the preset severe failure threshold, and the severity of the consequences indicates that the risk of product scrapping is less than or equal to the preset yield rate threshold, a delayed maintenance strategy is generated to reduce the operating speed until the end of the current batch before repair.

[0022] This embodiment provides a dynamic decision-making and execution mechanism for optimizing global process continuity and equipment health. Specifically, after the risk reasoning is completed, if the system still uses the traditional approach and triggers a shutdown intervention lacking differentiation based solely on a single equipment indicator, two technical defects may occur: firstly, it may lead to erroneous shutdowns causing performance imbalances in all workpieces within the mold cavity; secondly, it may continue to operate even when the component abnormality does not exceed the limit, but continuing to operate on a thermally fatigued mold may actually amplify structural damage. To resolve this contradiction, this embodiment combines the strategy generation, topology visualization, and instruction issuance of the embodiment with the hierarchical intervention rules of the embodiment, thereby upgrading the system output from fault alarms to technical intervention instructions with specific working conditions adaptability. Specifically, the strategy generation submodule receives the risk reasoning results, which include at least the comprehensive risk probability, the severity of the consequences, and the availability of the current batch of products and the mold hardware recovery parameters obtained from the process context data parsing. It should be noted that, for the step of obtaining the availability of the current batch of products and the mold hardware recovery parameters from the process context data parsing, since the process context data intuitively mainly represents the physical characteristics of the processed material viscosity, curing curve, and real-time temperature control, the system has pre-established a joint mapping logic in the global material and process configuration table: by extracting the material batch identifier and curing curve stage parameters from the process context, the complexity of the currently executed process and the proportion of material input are evaluated, and the cumulative quality assessment weight equivalent to the availability of the current batch of products is dynamically calculated. Simultaneously, by analyzing the mold level corresponding to the process requirements, the system derives the equipment hardware degradation assessment base equivalent to the mold hardware recovery parameters, thereby achieving a smooth conversion from physical process data to quantitative technical evaluation indicators. When it is necessary to select the best among multiple strategies, the system establishes a comprehensive intervention evaluation function. By calculating the sum of the expected product quality degradation degree and the expected mold hardware degradation degree corresponding to the specific strategy, multiplying it by the comprehensive risk probability, and then adding the production cycle interruption loss factor introduced by shutdown or parameter adjustment, the total technical loss value of the strategy is obtained. The strategy generation submodule selects the strategy that minimizes the total loss value as the adaptive intervention strategy. According to further limitations of the embodiments, the system executes at least the following three types of strategy branches; The first type is the emergency shutdown strategy. When the overall risk probability is greater than the preset serious fault threshold, it means that the abnormality is very likely to evolve into serious mold damage or extreme quality accident under the current working conditions. At this time, the strategy generation submodule immediately generates a shutdown command, and the command sending submodule converts it into equipment control signals, and sequentially executes decompression, stops the mold closing action, maintains the controlled temperature, and locks dangerous execution parts to avoid a simple power outage causing greater thermal shock or mechanical shock. The second category is parameter compensation strategy; when the overall risk probability is less than or equal to the preset severe failure threshold and the severity of the consequences indicates that the risk of product scrap is greater than the preset yield threshold, it means that the abnormality has not yet reached the catastrophic level, but if the process parameters are not corrected, the current batch size deviation will be significantly out of tolerance. It should be further explained that, in order to ensure the logical consistency of the rule engine when matching conditions, the system automatically converts the preset yield rate threshold into the maximum allowable quality deviation boundary rate corresponding to the target process yield through a built-in equivalent mapping algorithm when parsing the preset yield rate threshold. When the probability of product scrap risk is greater than the preset yield rate threshold, the system does not take the action of interrupting execution, but generates parameter compensation strategy, such as reducing the mold closing speed, increasing the buffer section stroke, fine-tuning the holding pressure, or limiting the mold opening acceleration, so as to weaken the subsequent evolution of the anomaly in the graph. The expected risk level after compensation can be regarded as the original comprehensive risk probability minus the risk reduction coefficient calibrated by the compensation intervention model. If the evaluation after compensation is still higher than the allowable safety limit, the system will upgrade to a shutdown strategy. The third category is delayed maintenance strategies. When the overall risk probability is less than or equal to the preset severe fault threshold and the severity of the consequences indicates that the risk of product scrap is less than or equal to the preset yield rate threshold, it means that based on the equivalent transformation logic, the probability of the workpiece exceeding tolerance caused by this anomaly is strictly constrained within the allowable tolerance boundary, and is within a safe and tolerable control domain under the current mold, current material, and current process combination. At this time, the system generates a delayed maintenance strategy of slowing down the operation until the end of the current batch before maintenance, such as reducing the equipment cycle time to the rated value. to Simultaneously increase the sampling frequency of key nodes and automatically trigger the mandatory maintenance mechanism after the processing batch is completed; The topology visualization submodule graphically renders the risk transmission topology data. Specifically, it can display the anomaly source node, key propagation links, and affected terminal nodes on the same interface. The node color represents the risk characteristic level, the edge thickness represents the information transmission weight, and the terminal nodes display the product quality deterioration tendency and mold structure fatigue evolution values. For operators, what they see is not abstract algorithm parameters, but a real-time topology network of anomaly source, transmission path, and terminal technology impact. The instruction issuing submodule is responsible for converting adaptive intervention strategies into executable control signals; if it is an emergency shutdown strategy, it outputs a sequence shutdown control word; if it is a parameter compensation strategy, it outputs the corrected servo mold closing speed or vibration buffer parameters; if it is a delayed maintenance strategy, it issues a frequency reduction operation instruction and simultaneously writes the module to be inspected and the suggested adjustment compensation window to the manufacturing execution system; in this way, a process closed loop is formed between the early warning calculation end and the equipment execution end. As a fault-tolerance mechanism, if the overall risk probability is exactly at the critical fault threshold boundary, a dual verification judgment mechanism can be set up. Only when multiple consecutive sampling cycles meet the high-risk conditions will the blocking intervention be triggered to prevent the execution misjudgment caused by instantaneous sensor fluctuations. If the process mapping parameters cannot be accurately analyzed temporarily, the historical safety assessment base of the corresponding product series will be used as a substitute. If the topology visualization shows that some node locations are lost, at least the three core objects of abnormal triggering source, feature transmission main path and affected terminal will be highlighted to ensure that the timeliness of the issuance of the underlying servo control command is not affected. If the issued control command does not receive a reply from the underlying control unit of the mold opening machine, the basic safety intervention mode will be automatically switched to, the overload circuit pressure will be released first and the audible and visual alarm device will be driven to provide physical prompts. In the later stages of production of the same batch of composite material shells, the system, based on the aforementioned propagation and reasoning results, concludes that the overall risk probability is less than the preset conventional tolerance threshold and lower than the severe fault threshold. However, due to the high process complexity of the current batch and the activation degree of the product edge dimension nodes being greater than the preset risk activation threshold, the severity of the consequences indicates that the risk of product scrapping exceeds the tolerance boundary after equivalent mapping. Therefore, the system generates a parameter compensation strategy, automatically lowers the mold closing servo command and loads vibration buffer parameter settings, and highlights the main transmission path of hydraulic valve 1 → mold closing cylinder → cavity hotspot A → edge dimension deviation in the topology diagram. If the dynamic node evolution characteristics of subsequent workpieces are attenuated after compensation, the current production batch is allowed to be completed continuously. If the evolution characteristics rise again and exceed the severe fault threshold, the system automatically switches to an emergency shutdown strategy. For example, in a scenario where the damping parameters of another batch of materials are greater than the preset high damping threshold and the mold does not show signs of thermal fatigue, the same hydraulic fluctuations only drive a lower overall risk probability, and the product size deviation does not trigger the threshold condition. In this case, the system outputs a slowdown operation until the end of this batch before performing maintenance, thus avoiding unnecessary shutdown and blocking operations. The purpose of this step is to translate the risk reasoning results into executable and verifiable servo adjustments and maintenance actions, transforming the early warning system from a simple condition monitoring tool into an adaptive decision-making node that takes into account mold safety, product tolerance, and processing continuity, thereby achieving a more robust intelligent fault management closed loop.

[0023] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A knowledge graph-based fault early warning system for mold-making machines, characterized in that, include: The data acquisition module collects time-series data of equipment operation status, multi-field coupling parameters and process context data of mold, extracts equipment mechanism and material rheology rules from a preset expert knowledge base, and generates a multimodal heterogeneous dataset. The ontology construction module constructs a fusion knowledge graph based on the multimodal heterogeneous dataset, which includes subgraphs of equipment topology, spatiotemporal stress, process rheology, and product quality. It transforms the time-series data of the operating status into dynamic node attributes of the graph and generates a dynamically instantiated knowledge graph. The penetration simulation module converts the material rheology rules into a loss function penalty term to construct a physical information graph neural network model. It calls the dynamically instantiated knowledge graph to identify abnormal nodes whose scalar feature values ​​of attributes deviate from a preset threshold. Based on the model, it calculates the transmission path and evolution state of abnormal node features in the fused knowledge graph, simulates the cross-domain information transmission of mechanical stress and thermodynamic field corresponding to multi-field coupling parameters, and generates risk transmission topology data. The risk reasoning module assesses the severity of the consequences of the affected nodes at the end of the transmission path based on the risk transmission topology data and preset multi-field coupling reasoning rules, determines the comprehensive risk probability of causing the scrapping of the target product and the damage to the core mold, and generates risk reasoning results. The dynamic decision-making module generates an adaptive intervention strategy based on the risk reasoning results and the preset global technical performance optimization principle, outputs a real-time risk transmission topology map, and sends intelligent maintenance instructions to the mold-making machine control system to complete the early warning.

2. The knowledge graph-based mold-opening machine fault early warning system according to claim 1, characterized in that, The operational status timing data includes vibration, hydraulic, motor current and displacement sensor data; the multi-field coupling parameters include mold stress concentration points, thermal fatigue and heat capacity parameters; the process context data includes the viscosity of the processed material, curing curve and real-time temperature control data; and the equipment mechanism and material rheology rules include historical maintenance logs, equipment bill of materials and material rheology formulas.

3. The knowledge graph-based mold-opening machine fault early warning system according to claim 1, characterized in that, The ontology construction module includes: The graph definition submodule extracts the equipment mechanism and material rheology rules from the multimodal heterogeneous dataset, defines the attributes of equipment entities, mold entities, material entities and product entities and the relationships between entities, and constructs subgraphs of equipment topology, spatiotemporal stress, process rheology and product quality. The graph fusion submodule splices the subgraphs of equipment topology, spatiotemporal stress, process rheology and product quality through entity alignment and relationship extraction to generate the fused knowledge graph. The dynamic instantiation submodule acquires the runtime time-series data, multi-field coupling parameters, and process context data from the multimodal heterogeneous dataset, maps them to the corresponding nodes of the fused knowledge graph according to timestamps, updates the dynamic node attributes of the corresponding nodes, and generates the dynamic instantiation knowledge graph.

4. The knowledge graph-based mold-opening machine fault early warning system according to claim 3, characterized in that, The penetration simulation module includes: The node feature extraction submodule calls the dynamically instantiated knowledge graph to identify abnormal nodes whose scalar feature values ​​of the dynamic node attributes deviate from the preset threshold, and extracts the features of the abnormal nodes and the features of their neighboring nodes. The physical constraint loading submodule transforms the material rheological rules in the multimodal heterogeneous dataset into a loss function penalty term, and combines them with the thermodynamic equations constructed based on the multi-field coupling parameters to build the physical information graph neural network model. The information transmission calculation submodule inputs the features of the abnormal node into the physical information graph neural network model, performs feature aggregation and updating along the edges of the fused knowledge graph, calculates the information evolution of the features on the transmission path, and generates the risk transmission topology data.

5. The knowledge graph-based mold-opening machine fault early warning system according to claim 4, characterized in that, In the information transmission calculation submodule, the step of calculating the information evolution of features along the transmission path includes: When the material viscosity parameter between nodes is greater than the preset viscosity threshold, it is determined that the feature of the abnormal node is rheologically buffered. The product of the material viscosity parameter and the preset damping coefficient is used as the attenuation index. The negative power of the attenuation index is calculated with the natural constant as the base, and multiplied by the feature of the abnormal node to obtain the feature attenuation amount. When the material viscosity parameter between nodes is less than or equal to a preset viscosity threshold and the thermal fatigue of the target node is greater than a preset critical point, it is determined that the feature of the abnormal node is amplified by thermal stress. The product of the thermal fatigue of the target node and the preset stress concentration factor is used as the feature amplification factor, and multiplied with the feature of the abnormal node to obtain the amplified feature quantity. When the material viscosity parameter between nodes is less than or equal to a preset viscosity threshold, and the thermal fatigue of the target node is less than or equal to a preset critical point, the characteristic linear transmission of the abnormal node is determined, and the characteristic magnitude remains unchanged. The material viscosity parameter is included in the process context data, the target node is a node on the conduction path, and the thermal fatigue degree is included in the multi-field coupling parameter.

6. The knowledge graph-based mold-opening machine fault early warning system according to claim 1, characterized in that, The risk reasoning module includes: The transmission path extraction submodule extracts the complete propagation link from the abnormal node to the affected terminal node based on the risk transmission topology data. The rule matching submodule calls the preset multi-field coupled inference rule and inputs the node degradation degree, mold thermal fatigue degree and current process state on the complete propagation link into the rule engine for condition matching. The node degradation degree is the ratio of the absolute value of the difference between the scalar feature value of the abnormal node and the preset safe interval boundary value to the corresponding boundary value deviated from. The mold thermal fatigue degree and the current process state are obtained by parsing the multimodal heterogeneous dataset. The consequence assessment submodule obtains the corresponding basic risk coefficient according to the matching rules, and uses the product of the basic risk coefficient, the degree of node deterioration, and the normalized thermal fatigue degree of the mold as the comprehensive risk probability of target product scrapping and core mold damage, and generates the risk reasoning result that includes the severity of the consequence and the comprehensive risk probability.

7. The knowledge graph-based mold-opening machine fault early warning system according to claim 6, characterized in that, The dynamic decision-making module includes: The strategy generation submodule, based on the risk reasoning results, compares the comprehensive risk probability with the preset risk tolerance, and combines the current batch product availability and mold hardware recovery parameters obtained from the process context data parsing to generate the adaptive intervention strategy; The topology visualization submodule graphically renders the abnormal nodes, propagation links, and severity of consequences in the risk transmission topology data, and outputs the real-time risk transmission topology map. The instruction issuing submodule converts the adaptive intervention strategy into equipment control signals and sends the intelligent maintenance instruction to the mold opening machine control system.

8. The knowledge graph-based mold-opening machine fault early warning system according to claim 7, characterized in that, The strategy generation submodule includes the following steps for generating the adaptive intervention strategy: When the overall risk probability exceeds a preset severe fault threshold, an emergency shutdown strategy is generated. When the overall risk probability is less than or equal to the preset severe failure threshold, and the severity of the consequences indicates that the risk of product scrapping is greater than the preset yield threshold, a parameter compensation strategy is generated to adjust the mold closing parameters and buffer vibration. When the overall risk probability is less than or equal to a preset severe failure threshold, and the severity of the consequences indicates that the risk of product scrapping is less than or equal to a preset yield rate threshold, a delayed maintenance strategy is generated to reduce the operating speed until the end of the current batch before repair.

Images