Intelligent analysis method for operation state of injection molding production line equipment

Abstract

The invention relates to the technical field of equipment state monitoring, and discloses an intelligent analysis method for the running state of injection molding production line equipment. The method comprises the step of synchronously collecting sensor data, alarm logs and quality reports. Semantic disassembly and event coding are performed on an alarm log to generate a structured event sequence, working condition stage division and feature extraction are performed on a sensor data stream to generate a time sequence feature matrix, and defect classification is performed on a quality report to generate a state vector. And dynamically constructing an equipment operation knowledge graph of the current time window, wherein graph nodes represent equipment, parameters and events, and edges represent association relationships among the equipment, the parameters and the events. According to the method, semantic-level deep fusion of multi-source heterogeneous data and explicit modeling of a complex incidence relation are achieved, internal relations among equipment abnormity, process parameters and product quality can be systematically revealed, and accurate state diagnosis and root cause analysis are supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment state monitoring, in particular to an intelligent analysis method for the running state of an injection molding production line. BACKGROUND

[0002] In the intelligent operation and maintenance of injection molding production, accurate analysis of the running state of the equipment is the key. The current common technical means mainly rely on monitoring and alarm of time series data of sensors, independent alarm log recording and periodic product quality sampling. These conventional methods usually independently or simply parallelly process various data sources.

[0003] The prior art has defects. The sensor data, alarm text and quality inspection results differ greatly in data structure, time scale and semantic level, forming a "data island". The conventional parallel processing method cannot realize the deep alignment and fusion of multi-dimensional data in the semantic level, and the specific process background of the alarm event, the specific production stage corresponding to the sensor reading change, and the correlation between the quality defect and the running parameters in a specific time period are fragmented. This results in that the state analysis stays at the single-point alarm response or post-event statistical level, and cannot reveal the potential complex causal relationship between the equipment abnormality, process parameter fluctuation and product quality problem.

[0004] A technical solution is needed to solve the problem of deep fusion and deep relationship mining of multi-source heterogeneous data. It is necessary to convert unstructured alarm text, continuous changing sensor time series signal and discrete quality sampling result into unified structured information rich in semantics; it is necessary to construct a networked model based on these structured information, which can dynamically reflect the complex correlation between equipment entities, process parameters, alarm events and quality states, to support comprehensive research and judgment and deep diagnosis of the running state of the equipment. SUMMARY

[0005] The purpose of the present application is to provide an intelligent analysis method for the running state of an injection molding production line to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides an intelligent analysis method for the running state of an injection molding production line, which comprises: Synchronously collecting device native running data in a predetermined time window from a control server and a historical database of a target injection molding production line, wherein the device native running data integrates time series arranged sensor reading stream, event-driven alarm log item set and periodically recorded product quality sampling report; The multi-dimensional deep analysis of the device native operation data includes: semantic disassembly and event coding of the alarm log entry set to output a structured alarm event sequence; working condition stage division and stage-in statistical feature calculation of the sensor reading stream to output a multi-dimensional time sequence feature matrix; defect mode classification and quality grade marking of the product quality sampling inspection report to output a quality state vector; According to the structured alarm event sequence, the multi-dimensional time sequence feature matrix and the quality state vector, a device operation knowledge graph facing a current time window is dynamically constructed, and nodes of the device operation knowledge graph represent device entities, process parameters and alarm events, and edges represent time sequence, causal or logical correlation between nodes.

[0007] Preferably, the multi-dimensional deep analysis of the device native operation data includes the following steps: Each text record in the alarm log entry set is analyzed, and device component names, alarm codes and state description words are extracted therefrom by named entity recognition technology; The extracted device component names, alarm codes and state description words are mapped and matched with a predefined device ontology library, and unstructured text records are converted into structured alarm events with uniform semantics; All structured alarm events are sorted and merged according to timestamps to generate the structured alarm event sequence, wherein adjacent and same-type structured alarm events are merged into a composite event with a continuous time length.

[0008] Preferably, the multi-dimensional deep analysis of the device native operation data further includes the following steps: A specific signal pattern representing the start and end of an injection molding cycle is identified in the sensor reading stream, and the continuous sensor reading stream is cut into continuous single injection molding cycle data segments based on the specific signal pattern; Each single injection molding cycle data segment is further divided into preset process stages of injection, pressure maintaining, cooling and mold opening; Statistical features of readings of each sensor channel in each preset process stage are calculated, the statistical features include mean, variance, peak value and trend slope, and all statistical features are organized by stage and sensor channel to form the multi-dimensional time sequence feature matrix.

[0009] Preferably, the dynamic construction of a device operation knowledge graph facing a current time window according to the structured alarm event sequence, the multi-dimensional time sequence feature matrix and the quality state vector includes the following steps: Device physical components, sensor measuring points, alarm event types and product quality defects are taken as node types to create an initial graph node set; traversing the structured alarm event sequence, adding each event as an event node to the graph node set, and establishing "triggered by" and "associated with" edges between the event node and related device component node, sensor measurement point node; traversing the multi-dimensional time-series feature matrix, creating a feature anomaly node for statistical features exceeding a preset threshold, and establishing "occurred in" edges between the feature anomaly node and corresponding sensor measurement point node, process stage node; based on the quality state vector, creating a quality defect node, and establishing "may be caused by" edges between the quality defect node and the event node and feature anomaly node within a specific time interval according to the time alignment relationship.

[0010] Preferably, the method further comprises the process of enhancing and reasoning the device operation knowledge graph: In the device operation knowledge graph, multi-hop path search is performed along the "triggered by", "associated with" and "may be caused by" edges to identify common sub-paths connecting multiple feature anomaly nodes, event nodes and ultimately pointing to the same quality defect node; For each of the identified common sub-paths, a path confidence is calculated, which is determined based on the historical frequency and conditional probability of node co-occurrence on the common sub-path; The common sub-paths with confidence exceeding a threshold value are solidified as new causal edges with stronger semantics and added to the device operation knowledge graph, thereby forming an enhanced device operation causal knowledge graph.

[0011] Preferably, the method further comprises the process of state backtracking analysis based on virtual simulation: From the enhanced device operation causal knowledge graph, all high-confidence causal paths ending with a specific quality defect node are extracted to form a candidate root cause chain set; For each causal path in the candidate root cause chain set, the source abnormal node at the beginning of the path is located, which corresponds to a specific sensor feature anomaly or alarm event; In the injection molding process simulation model, device parameter deviations or disturbance signals corresponding to the source abnormal node are injected, and the injection molding process simulation model is run to obtain a virtual production result; Comparing the virtual production result with the actually recorded quality state vector, if the simulated defect matches the actual defect pattern, this causal path is marked as a verified root cause path.

[0012] Preferably, the injection of device parameter deviations or disturbance signals corresponding to the source abnormal node in the injection molding process simulation model comprises: When the source abnormal node is a sensor feature abnormal node, resolve the sensor measurement point, process stage and abnormal feature type associated with the sensor feature abnormal node; In the injection molding process simulation model, locate the corresponding simulated sensor and process stage, modify the model input parameters or internal state to reproduce the reading deviation behavior described by the abnormal feature type; When the source abnormal node is an alarm event node, resolve the corresponding device component and alarm type of the alarm event node; In the injection molding process simulation model, trigger the degradation or failure logic of the corresponding device component to simulate the abnormal device condition represented by the alarm type.

[0013] Preferably, the method further comprises a process of generating targeted maintenance and adjustment strategies: Conduct inductive analysis on the root cause path that passes the verification, extract the device component nodes, key process parameter nodes and their abnormal patterns that repeatedly appear in the path; Combine the device maintenance manual and the process expert rule base to match one or more basic countermeasures for each abnormal pattern, forming an original strategy set; Evaluate the implementation feasibility of the measures in the original strategy set under the current production order constraints, including production rhythm, material attributes and device physical limits; According to the implementation feasibility evaluation results, sort and optimize the combination of measures, and output the final executable strategy instruction set.

[0014] Preferably, the evaluation of the implementation feasibility under the current production order constraints includes: Obtain the process requirement file of the current production order and the real-time health index of each device on the production line; Conduct conformity check on each measure in the original strategy set with the parameter range in the process requirement file, and select the non-conflicting candidate measures; Match the candidate measures with the real-time health index of the device, evaluate the potential impact on the health of the device and the risk level of executing the candidate measures; Integrate the conformity check results and the risk level to calculate the feasibility score for each measure, and filter and sort the measures according to the feasibility score.

[0015] Preferably, the path confidence is calculated for each of the identified common sub-paths, including performing the following steps: Extract the node sequence contained in the common sub-path, which points from the starting source abnormal node or event node to the terminal quality defect node through intermediate nodes; Based on the historical equipment operation knowledge graph, the number of co-occurrence of each pair of adjacent nodes in the node sequence in a historical time window is counted, and the co-occurrence frequency of each pair of adjacent nodes is calculated. For the node sequence, starting from the starting node, the conditional probability of the occurrence of the downstream node under the condition that the upstream node occurs is calculated hop by hop based on Bayes theorem. The co-occurrence frequency of each pair of adjacent nodes is multiplied by the conditional probability to obtain the correlation strength of the pair of adjacent nodes, and the correlation strengths of all hops on the path are multiplied to obtain the original confidence score of the common sub-path. The original confidence scores of all identified common sub-paths are normalized so that the path confidence falls within the value interval of 0 to 1.

[0016] Compared with the prior art, the beneficial effects of the present application are: The alarm logs in the device native operation data are semantically disassembled and event coded, and the unstructured natural language description is converted into a structured sequence containing event type, component, severity, etc. The sensor reading stream is divided into working condition stages according to the process, and the statistical characteristics are calculated and counted in each stage to form a multi-dimensional time sequence feature matrix with clear working condition context. The product quality sampling report is classified and labeled by defect mode, and a quantitative quality state vector is generated. The originally heterogeneous and multi-modal raw data are converted into structured data with clear semantics and aligned dimensions, breaking the semantic barriers between sensor data, alarm information and quality results, and providing a unified and high-quality data foundation for subsequent deep correlation analysis. The depth and accuracy of data fusion surpasses the conventional simple data juxtaposition or time stamp alignment method.

[0017] According to the structured alarm event sequence, multi-dimensional time sequence feature matrix and quality state vector obtained after deep analysis, a device operation knowledge graph is dynamically constructed, which takes device entities, process parameters and alarm events as nodes, and takes time sequence, causal or logical relationship as edges. The graph is not a static device information library, but a relationship network dynamically generated for the current time window operation data. It can intuitively present the related process parameter characteristics when a specific alarm occurs, clearly trace the parameter abnormal course leading to a certain quality defect mode, and reveal the co-occurrence or triggering relationship between different alarm events. This relationship network representation based on the graph changes the device operation state from a series of discrete data points to a relationship network, realizes the transition from isolated index monitoring to systematic correlation reasoning, and supports the traceability analysis of complex fault root cause and quality fluctuation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The working principle diagram of the injection molding production line equipment operation state intelligent analysis method described in the present application; Figure 2 Flowchart for parsing and structuring of alarm log entries set; Figure 3 Flowchart for knowledge graph enhancement and inference of equipment operation; Figure 4 Pressure sensor PS-205 normal and variance large pressure curve comparison chart for holding stage; Figure 5 Sensor mean heat map for each stage of injection molding process. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0020] Please refer to Figure 1 The present application provides an intelligent analysis method for the running state of an injection molding production line equipment, which comprises: synchronously collecting equipment native running data within a predetermined time window from a control server and a historical database of a target injection molding production line, the equipment native running data integrating time-series arranged sensor reading streams, event-driven alarm log entry sets and periodically recorded product quality sampling reports; performing multi-dimensional deep analysis on the equipment native running data, which specifically comprises: performing semantic disassembly and event coding on the alarm log entry sets and outputting a structured alarm event sequence, performing working condition stage division and stage internal statistical feature calculation on the sensor reading streams and outputting a multi-dimensional time-series feature matrix, and performing defect mode classification and quality grade marking on the product quality sampling reports and outputting a quality state vector; dynamically constructing an equipment operation knowledge graph facing the current time window according to the structured alarm event sequence, the multi-dimensional time-series feature matrix and the quality state vector obtained after analysis, the nodes of the equipment operation knowledge graph representing equipment entities, process parameters and alarm events, and the edges representing the time-series, causal or logical correlation between the nodes.

[0021] Embodiment 1: refer to Figure 2In the alarm log entry set parsing, the following steps are performed: parsing each text record in the alarm log entry set, extracting the device component name, alarm code and state description word therein through named entity recognition technology; mapping and matching the extracted device component name, alarm code and state description word with the predefined device ontology library, converting the unstructured text record into a structured alarm event with unified semantics; sorting and merging all structured alarm events according to the time stamp, generating a structured alarm event sequence, wherein adjacent and same type structured alarm events are merged into a composite event with a continuous time length.

[0022] In the sensor reading stream parsing, the following steps are performed: identifying a specific signal pattern representing the start and end of an injection molding cycle in the sensor reading stream, cutting the continuous sensor reading stream into continuous single injection molding cycle data segments based on the specific signal pattern; for each single injection molding cycle data segment, further dividing it into preset process stages of injection, pressure maintaining, cooling and mold opening; calculating the statistical features of each sensor channel reading in each preset process stage, the statistical features including mean, variance, peak value and trend slope, organizing all statistical features by stage and sensor channel to form a multi-dimensional time sequence feature matrix.

[0023] In a specific implementation, it is assumed that the target injection molding production line includes an injection molding machine with a model of MM-200, a control server runs a manufacturing execution system, and a historical database records production data. The alarm log entry set within a predetermined time window is collected from the control server, and a text record of an alarm log entry is “2024-05-10 14:22:15, MM-200_HeaterZone3TempHigh, ALM-3072, Heating barrel zone 3 temperature exceeds 235C”. The readings of multiple sensors within the same time window are synchronously collected from the historical database to form a sensor reading stream, and the sensor reading stream includes millisecond-level sampling data of a temperature sensor PT-101, a pressure sensor PS-205 and a displacement sensor LS-400. At the same time, the product quality sampling report corresponding to the same time window is collected from the quality database, and the product quality sampling report contains records of product number, size out-of-tolerance defects and appearance shrinkage defects.

[0024] In the alarm log entry set parsing process, each text record is semantically decomposed. The named entity recognition technique is applied to the alarm log entry text record to extract the device component name "Heating barrel zone 3", the alarm code "ALM-3072", and the state description words "Temp High" and "temperature exceeds 235 C". The extracted device component name, alarm code, and state description words are mapped and matched with the predefined device ontology library, which defines the standard component name "Heating barrel third zone", the standard alarm type "Temperature upper limit alarm", and the standard state "Exceeds limit". The mapping and matching convert the unstructured text record into a structured alarm event with unified semantics, and the structured alarm event contains standardized fields: timestamp "2024-05-10 14:22:15", device component "Heating barrel third zone", event type "Temperature upper limit alarm", and event description "Temperature exceeds 235 degrees Celsius". All structured alarm events are sorted by timestamp, and adjacent and same-type structured alarm events are merged. For example, if "Temperature upper limit alarm" about "Heating barrel third zone" occurs in three consecutive injection cycles, the three independent events are merged into a composite event with a duration, and the start time of the composite event is the occurrence time of the first event, and the end time is the occurrence time of the last event, thereby generating a structured alarm event sequence.

[0025] In the sensor reading stream parsing process, specific signal patterns representing the start and end of an injection cycle are identified in the sensor reading stream. The specific signal pattern is a signal jump of the mold locking pressure sensor reading from zero to a set pressure threshold. Based on the jump point of the mold locking pressure signal, the continuous sensor reading stream is cut into continuous single injection cycle data segments. For each single injection cycle data segment, it is further divided into preset process stages of injection, holding, cooling, and opening according to the injection machine process timing template. The division of the process stage is determined according to the phase change of the screw position signal and the mold cavity pressure signal. The statistical features of the readings of each sensor channel in each preset process stage are calculated. For the temperature sensor PT-101 reading in the injection stage, its mean value is calculated as 225.3 degrees Celsius, its variance is calculated as 4.7, and its peak value is identified as 231.8 degrees Celsius. The calculation of the trend slope is obtained by linear fitting the sequence of temperature readings changing with time in this stage, and the calculation formula is as follows:

[0026] wherein: represents the trend slope, represents the number of sampling points in the injection stage, represents the time offset of the th sampling point relative to the start of the stage, This indicates that the temperature sensor PT-101 is in the... The readings at each sampling point. The trend slope of the PT-101 temperature sensor readings during the injection phase. The calculation result is 0.15 degrees Celsius per second. The mean, variance, peak value, and trend slope of the temperature sensor PT-101 readings during the injection stage are used as a set of statistical characteristics. The same calculations are performed on the pressure sensor PS-205 and the displacement sensor LS-400 during the injection stage. Statistical characteristic calculations are performed on all sensors at all preset process stages, and all statistical characteristics are organized by stage and sensor channel to form a multidimensional time-series characteristic matrix.

[0027] In practice, the multi-dimensional deep analysis of the alarm log entry set and the sensor reading stream is performed independently. Parsing the alarm log entry set outputs a structured alarm event sequence, while parsing the sensor reading stream outputs a multi-dimensional time-series feature matrix. The structured alarm event sequence and the multi-dimensional time-series feature matrix are synchronously aligned on the time axis, and both cover the same predetermined time window. It can be understood that the parsing of the alarm log entry set focuses on the structuring of discrete events, while the parsing of the sensor reading stream focuses on the extraction of quantized features from continuous signals. Optionally, named entity recognition technology can employ sequence labeling methods based on pre-trained language models. The specific implementation includes loading a language model pre-trained on a general text corpus, such as BERT or a similar architecture, which has the ability to understand natural language context; taking each text record in the alarm log entries as the input sequence, first performing word segmentation and adding special tags, then inputting it into the pre-trained language model to obtain the contextual embedding representation of each word; superimposing a sequence labeling layer on these embedding representations, the model outputs a label sequence for each word, and through post-processing steps such as label decoding and entity block merging, finally extracting structured device component names, alarm codes, and status description words; the entire process supports end-to-end training, and the pre-trained model can be fine-tuned using labeled alarm log data to optimize entity recognition accuracy in specific industrial text domains. Optionally, the working condition stage division can adopt a template matching algorithm based on dynamic time warping.

[0028] In an embodiment, in constructing the knowledge graph of the device operation, the following steps are performed: taking the device physical components, sensor measurement points, alarm event types, and product quality defects as node types, an initial graph node set is created; traversing the structured alarm event sequence, each event is taken as an event node to join the graph node set, and "triggered by" and "associated with" edges between the event node and related device component nodes, sensor measurement point nodes are established; traversing the multi-dimensional time series feature matrix, a feature anomaly node is created for a statistical feature exceeding a preset threshold, and "occurred in" edges between the feature anomaly node and corresponding sensor measurement point nodes, process stage nodes are established; based on the quality state vector, a quality defect node is created, and according to the time alignment relationship, "may be caused by" edges between the quality defect node and event nodes and feature anomaly nodes in a specific time interval are established.

[0029] In an embodiment, the construction of the knowledge graph is based on a graph database framework which supports dynamic creation and attribute association of nodes and edges. In an embodiment, taking the device physical components, sensor measurement points, alarm event types, and product quality defects as node types, an initial graph node set is created. The device physical component nodes include "injection molding machine MM-200", "injection screw", "heating cylinder third zone", and "mold cavity A". The sensor measurement point nodes include "temperature sensor PT-101", "pressure sensor PS-205", and "displacement sensor LS-400". The alarm event type nodes are predefined according to the event type field in the structured alarm event sequence, including "temperature upper limit alarm" and "pressure fluctuation alarm". The product quality defect nodes are predefined according to the defect mode field in the quality state vector, including "dimensional out-of-tolerance" and "appearance shrink mark". The initial graph node set contains all these entities with unique identifiers and node type attributes.

[0030] Traverse the structured alarm event sequence, and add each event in the structured alarm event sequence as an event node to the graph node set. A specific event node contains attributes: event ID, timestamp, description, duration. For example, an event node is created for the composite event "heating cylinder third zone temperature upper limit alarm", with event ID "E_20240510_142215", timestamp "2024-05-10 14:22:15", description "heating cylinder third zone temperature continuously over limit", and duration "180 seconds". Edges are established between this event node and related device component nodes, sensor measurement point nodes. In a specific implementation, a "triggered by" edge is created from event node "E_20240510_142215" to device component node "heating cylinder third zone", with edge attributes recording the triggering relationship. An "associated with" edge is created from event node "E_20240510_142215" to sensor measurement point node "temperature sensor PT-101", with edge attributes indicating that this alarm event is associated with the abnormal reading of this sensor. Traverse all events in the structured alarm event sequence, and repeat the creation of event nodes and the establishment of "triggered by" edges and "associated with" edges.

[0031] Traverse the multi-dimensional time series feature matrix, and create a feature anomaly node for statistical features that exceed a preset threshold. The preset threshold is set separately for different statistical feature types, for example, the temperature mean threshold of the injection stage is set to 220 degrees Celsius, and the pressure variance threshold of the injection stage is set to 5.0 MPa. Scan the multi-dimensional time series feature matrix, and identify that the mean value of temperature sensor PT-101 in the injection stage is 225.3 degrees Celsius, which exceeds the corresponding mean threshold of 220 degrees Celsius, and thus a feature anomaly node is created. The attributes of the feature anomaly node include: feature ID, associated sensor, associated process stage, anomaly feature type, and anomaly feature value. The node feature ID is "F_PT101_Injection_Mean", the associated sensor is "temperature sensor PT-101", the associated process stage is "injection", the anomaly feature type is "high mean value", and the anomaly feature value is "225.3". An "occurred in" edge is established between this feature anomaly node and the corresponding sensor measurement point node and process stage node. Specifically, a "occurred in" edge is created from feature anomaly node "F_PT101_Injection_Mean" to sensor measurement point node "temperature sensor PT-101". At the same time, a "occurred in" edge is created from feature anomaly node "F_PT101_Injection_Mean" to the process stage node representing the "injection" stage. Traverse all statistical features in the multi-dimensional time series feature matrix, and create feature anomaly nodes and establish "occurred in" edges for all features that exceed their respective thresholds.

[0032] Based on the quality state vector, a quality defect node is created. The quality state vector indicates that there is a "dimension out of tolerance" defect in a batch of products produced within a predetermined time window. A quality defect node is created with node attributes including: defect ID, defect type, time window of occurrence, and affected batch. The defect ID of this node is "Q_Dim_Out_001", the defect type is "dimension out of tolerance", the time window of occurrence is "2024-05-10 14:00:00 to 2024-05-10 15:00:00", and the affected batch is "Batch_2456". According to the time alignment relationship, "may be derived from" edges are established between the quality defect node and the event nodes and feature anomaly nodes within a specific time interval. The time alignment relationship means that only those anomaly nodes with timestamps within a set process delay before the quality defect occurrence time window are considered to be connected. The initial confidence score of the "may be derived from" edge is calculated according to the following formula:

[0033] wherein: represents the initial confidence score of the "may be derived from" edge, represents the weight coefficient of the conditional probability, represents the conditional probability of the occurrence of the target quality defect given the occurrence of the source node, represents the weight coefficient of the co-occurrence frequency, represents the frequency of co-occurrence of the source node and the target quality defect within a similar time window in historical data. For example, a "may be derived from" edge is created from the feature anomaly node "F_PT101_Injection_Mean" to the quality defect node "Q_Dim_Out_001", with the edge attributes including the value and value calculated according to historical data, and the initial confidence score is calculated according to the formula . In some embodiments, the weight coefficients and are set by domain experts. Optionally, the process delay for time alignment can be configured according to the production cycle of the specific product. It can be understood that through the above steps, a device operation knowledge graph for the current time window is dynamically constructed, containing multiple node types and relationship edges. In some embodiments, the physical storage of the graph can use the property graph model, and be implemented in the graph database through the node table NodeTable and the edge table EdgeTable. Optionally, the construction of the graph is incremental, and new nodes and edges are added to the existing graph structure as new time window data is analyzed.

[0034] Embodiment 3: refer to Figure 3The process of enhancing and reasoning the equipment operation knowledge graph is as follows: in the equipment operation knowledge graph, multi-hop path search is performed along the "triggered by", "associated with" and "may be caused by" edges, a common sub-path connecting multiple feature anomaly nodes, event nodes and finally pointing to the same quality defect node is identified; the path confidence of each identified common sub-path is calculated, the path confidence is determined based on the historical frequency and conditional probability of node co-occurrence on the common sub-path; the common sub-path with confidence exceeding the threshold value is solidified as a new causal edge with stronger semantics and added to the equipment operation knowledge graph, thereby forming an enhanced equipment operation causal knowledge graph.

[0035] wherein the path confidence of each identified common sub-path is calculated, including performing the following steps: extracting the node sequence contained in the common sub-path, the node sequence starting from the source abnormal node or event node, passing through the intermediate node, and pointing to the terminal quality defect node; based on the historical equipment operation knowledge graph, the number of co-occurrences of each pair of adjacent nodes in the node sequence within the historical time window is counted, and the co-occurrence frequency of each pair of adjacent nodes is calculated; for the node sequence, starting from the starting node, the conditional probability of the occurrence of the downstream node under the occurrence of the upstream node is calculated hop by hop based on Bayes theorem; the co-occurrence frequency of each pair of adjacent nodes is multiplied by the conditional probability to obtain the association strength of this pair of adjacent nodes, and the association strengths of all hops on the path are multiplied to obtain the original confidence score of the common sub-path; the original confidence scores of all identified common sub-paths are normalized so that the path confidence falls within the value interval of 0 to 1.

[0036] In a specific implementation, the equipment operation knowledge graph contains an event node of "heating cylinder third zone" temperature upper limit alarm, a feature anomaly node of injection stage temperature sensor PT-101 mean value being too high, a feature anomaly node of hold stage pressure sensor PS-205 variance being too large, and a quality defect node of "dimension out of tolerance". In the equipment operation knowledge graph, multi-hop path search is performed along the "triggered by", "associated with" and "may be caused by" edges, and the search algorithm uses a graph traversal method based on depth-first search. Multi-hop path search identifies a common sub-path connecting multiple feature anomaly nodes, event nodes and finally pointing to the same quality defect node. For example, a common sub-path is identified, which starts from the feature anomaly node "F_PS205_Hold_Var", passes through the event node "E_20240510_142215", and finally points to the quality defect node "Q_Dim_Out_001", which indicates that "pressure variance anomaly" and "temperature alarm" are jointly associated with the "dimension out of tolerance" defect.

[0037] In practice, a path confidence score is calculated for each identified common sub-path. The path confidence score is determined based on the historical frequency and conditional probability of node co-occurrence on the common sub-path. The node sequence contained in the common sub-path is extracted, starting from the source anomalous node or event node, passing through intermediate nodes, and pointing to the terminating quality defect node. Taking the common sub-path "Feature Anomalous Node A -> Event Node B -> Quality Defect Node C" as an example, the node sequence includes Feature Anomalous Node A, Event Node B, and Quality Defect Node C. Based on the historical equipment operation knowledge graph, the number of times each pair of adjacent nodes in the node sequence co-occurs within the historical time window is statistically analyzed. The historical equipment operation knowledge graph is constructed from production data from the past thirty days. The co-occurrence frequency of each pair of adjacent nodes is calculated; the co-occurrence frequency is the ratio of the number of co-occurrences to the total number of times the source node appears within the historical time window. For the node sequence, starting from the initial node, the conditional probability of the downstream node appearing given the occurrence of the upstream node is calculated hop-by-hop based on Bayes' theorem. The formula for calculating the original confidence score of the common sub-path is as follows:

[0038] in: Indicates the first The original confidence scores of the common sub-paths. This indicates the total number of nodes on this common sub-path. Indicates the sequence number along the path. Indicates the first The first common sub-path The co-occurrence frequency of adjacent nodes, Indicates the first The first common sub-path The conditional probability of a downstream node jumping to an upstream node given that the upstream node has already occurred. For example, for the common subpath "featured anomaly node A -> event node B -> quality defect node C", first calculate the co-occurrence frequency of feature anomaly node A and event node B. and conditional probability Then calculate the co-occurrence frequency of event node B and quality defect node C. and conditional probability Then The product of these factors serves as the original confidence score for this common sub-path. The original confidence scores of all identified common sub-paths are normalized using a maximum-minimum scaling method to ensure that the confidence scores of each path fall within the range of 0 to 1.

[0039] In a specific implementation, the common sub-paths with confidence exceeding a threshold value are solidified as new causal edges with stronger semantics. The threshold value is set to 0.65. For example, the normalized path confidence of the common sub-path “feature abnormality node A -> event node B -> quality defect node C” is calculated to be 0.72, which exceeds the threshold value 0.65. This common sub-path is solidified as a new causal edge directly pointing from the starting feature abnormality node A to the terminating quality defect node C, with the semantic label of the new causal edge being “causes”, and the attribute of the new causal edge recording its provenance from the original common sub-path and its confidence score 0.72. This new “causes” edge is added to the equipment operation knowledge graph. Similarly, the same solidification operation is performed on all other common sub-paths with confidence exceeding the threshold value, and the solidification process is to summarize a multi-hop path relationship into a direct causal relationship. In some embodiments, the new causal edge can be defined as “conducted to” or “triggered by” in addition to the “causes” relationship, according to the types of intermediate nodes in the path. Optionally, the threshold value can be differentiated according to different quality defect types. It can be understood that through this enhancement process, the indirect associations between nodes in the equipment operation knowledge graph are explicitly expressed as direct causal hypotheses, forming an enhanced equipment operation causal knowledge graph, and the edges in the enhanced equipment operation causal knowledge graph include both the “triggered by”, “associated with”, “occurred in”, “possibly originated from” edges directly established from the original data, and the “causes” edges with higher confidence generated through reasoning.

[0040] In some embodiments, the process of state backtracking analysis based on virtual simulation is as follows: from the enhanced equipment operation causal knowledge graph, all high-confidence causal paths with a specific quality defect node as the terminal are extracted to form a candidate root cause chain set; for each causal path in the candidate root cause chain set, the source abnormality node at the start of the causal path is located, which corresponds to a sensor feature abnormality or an alarm event; in the injection molding process simulation model, the equipment parameter deviation or disturbance signal corresponding to the source abnormality node is injected, and the injection molding process simulation model is run to obtain a virtual production result; the virtual production result is compared with the actual recorded quality state vector, and if the simulated defect matches the actual defect mode, this causal path is marked as a verified root cause path.

[0041] In the injection molding process simulation model, inject the equipment parameter deviation or disturbance signal corresponding to the source abnormal node, including: when the source abnormal node is a sensor feature abnormal node, analyze the sensor feature abnormal node associated sensor measurement point, process stage and abnormal feature type; in the injection molding process simulation model, locate the corresponding simulation sensor and process stage, modify the model input parameters or internal state to reproduce the reading deviation behavior described by the abnormal feature type; when the source abnormal node is an alarm event node, analyze the device component and alarm type corresponding to the alarm event node; in the injection molding process simulation model, trigger the degradation or failure logic of the corresponding device component, simulate the abnormal working condition of the device represented by the alarm type.

[0042] In specific implementation, the state backtracking analysis process based on virtual simulation takes the enhanced equipment operation causal knowledge graph as input, and the enhanced equipment operation causal knowledge graph contains the "cause" edges from the feature abnormal node or event node to the quality defect node generated by reasoning. From the enhanced equipment operation causal knowledge graph, all high-confidence causal paths with a specific quality defect node as the terminal are extracted, and the filtering condition of the high-confidence causal path is that the confidence score of the terminal "cause" edge is greater than the preset path threshold 0.70. A candidate root cause chain set is constructed. For example, for the quality defect node "Q_Dim_Out_001", three causal paths with this node as the terminal are extracted, and the confidence of the terminal edge of the path is 0.72, 0.68 and 0.81 respectively. The two paths with confidence of 0.72 and 0.81 are included in the candidate root cause chain set, and the path with confidence of 0.68 is excluded because it does not reach the threshold.

[0043] In specific implementation, for each causal path in the candidate root cause chain set, locate its starting source abnormal node. The node sequence of the first causal path is "feature abnormal node: F_PS205_Hold_Var" -> "event node: E_20240510_142215" -> "quality defect node: Q_Dim_Out_001", and the starting source abnormal node is "feature abnormal node: F_PS205_Hold_Var", which corresponds to the variance abnormality of the pressure sensor PS-205 in the holding stage. The node sequence of the second causal path is "event node: E_20240510_144530" -> "quality defect node: Q_Dim_Out_001", and the starting source abnormal node is "event node: E_20240510_144530", which corresponds to the "low mold cooling water flow alarm" event.

[0044] In the injection molding process simulation model, the device parameter deviation or disturbance signal corresponding to the source abnormal node is injected. The injection molding process simulation model is a digital twin model based on physical laws and material properties, which can simulate the whole process of the injection molding machine from mold closing, injection, holding to mold opening. When the source abnormal node is a sensor feature abnormal node, the sensor measurement point, process stage and abnormal feature type associated with the sensor feature abnormal node are analyzed. For the source abnormal node "F_PS205_Hold_Var", it is analyzed that the associated sensor measurement point is "pressure sensor PS-205", the associated process stage is "holding", and the abnormal feature type is "large variance". In the injection molding process simulation model, the corresponding simulation sensor "PS-205_Sim" and simulation process stage "Hold_Phase" are located, and the model input parameters or internal state are modified to reproduce the reading deviation behavior described by the abnormal feature type. Specifically, by adjusting the pressure control logic in the holding phase of the simulation model, a Gaussian noise signal with a mean of 0 and a standard deviation of 1.8 times the original set value is introduced to superimpose on the holding pressure setting curve, so that the feature of large variance is reproduced in the output of the simulation sensor "PS-205_Sim".

[0045] Referring to Table 1, when the source abnormal node is an alarm event node, the device component and alarm type corresponding to the alarm event node are analyzed. For the source abnormal node "event node: E_20240510_144530", it is analyzed that the corresponding device component is "mold cooling water circuit C" and the alarm type is "low flow alarm". In the injection molding process simulation model, the degradation or failure logic of the corresponding device component is triggered to simulate the abnormal working condition of the device represented by the alarm type. Specifically, the heat exchange coefficient corresponding to "mold cooling water circuit C" in the simulation model is reduced to 60% of the normal value to simulate the heat exchange efficiency reduction caused by insufficient cooling water flow. The injection molding process simulation model is run to obtain a virtual production result. The virtual production result includes the product size data of the simulated production. The virtual production result is compared with the actual recorded quality state vector, which records the occurrence rate of the actual product "size out-of-tolerance" defect. The basis for determining whether the simulated defect matches the actual defect pattern is that the difference between the proportion of the simulated product size distribution exceeding the tolerance range and the actual recorded proportion is less than the set tolerance. If the simulated defect matches the actual defect pattern, mark this causal path as a verified root cause path.

[0046] Table 1: Comparison table of virtual simulation backtracking analysis results

[0047] In some embodiments, after running the injection molding process simulation model to obtain a virtual production result, the matching degree of the simulated defect and the actual defect pattern is quantitatively scored. The quantitative scoring function is as follows:

[0048] wherein: represents the matching score, represents the occurrence rate of the target defect in the simulated defect pattern, represents the occurrence rate of the target defect in the actual defect pattern. If the matching score is greater than or equal to the matching threshold 0.85, it is determined to be matched. For the first path in Table 1, the matching score is calculated to be greater than the threshold, so it is marked as verified. It can be understood that, by state backtracking analysis through virtual simulation, the controllability and repeatability of the simulation model can be used to preliminarily verify the causal hypothesis inferred from the knowledge graph.

[0049] Referring to Figure 4 , the readings of the injection pressure sensor PS-205 in the parameter perturbation injection scene are compared. The blue curve in the figure is the normal pressure curve, the red curve is the variance large pressure curve after perturbation injection, and the light blue area is the normal pressure fluctuation range. Specifically, the normal pressure curve (blue) presents a stable periodic fluctuation around the mean value within the holding time, and is always within the normal fluctuation range; while the variance large pressure curve (red) shows a significant deviation feature: its reading fluctuation amplitude is much larger than the normal range, and the peak value exceeds 160 MPa several times, and the valley value is as low as 130 MPa, which is significantly different from the stable fluctuation of the normal curve. This curve feature corresponds to the pressure variance large abnormality represented by the feature abnormal node "F_PS205_Hold_Var", which is the sensor reading deviation behavior reproduced after injecting Gaussian noise perturbation in the injection process simulation model, and also provides a measured data basis for subsequent virtual simulation verification of the root cause path.

[0050] In example 5, the process of generating a targeted maintenance and adjustment strategy is as follows: the verified root cause path is analyzed to extract the repeatedly appearing device component nodes, key process parameter nodes and their abnormal patterns in the path; one or more basic measures are matched for each abnormal pattern according to the device maintenance manual and the process expert rule library, forming an original strategy set; the measures in the original strategy set are evaluated for their implementation feasibility under the current production order constraints, including production rhythm, material properties and device physical limits; the measures are sorted and optimized according to the implementation feasibility evaluation results, and the final executable strategy instruction set is output.

[0051] The implementation feasibility of each measure in the original strategy set under the current production order constraints is evaluated, including: obtaining the process requirement file of the current production order and the real-time health index of each device on the production line; performing compliance checking on each measure in the original strategy set and the parameter range in the process requirement file to screen out candidate measures that do not conflict; matching the candidate measures with the real-time health index of the device to evaluate the potential impact of the candidate measures on the health of the device and the risk level; and comprehensively checking the compliance and the risk level to calculate a feasibility score for each measure and to screen and sort the measures according to the feasibility score.

[0052] In a specific implementation, the process of generating the targeted maintenance and adjustment strategy takes the verified root cause path as input, and the verified root cause path is derived from the result of virtual simulation backtracking analysis. The verified root cause path is analyzed to extract repeatedly occurring device component nodes, key process parameter nodes, and abnormal patterns in the path. For example, for the verified root cause path with the "F_PS205_Hold_Var" as the source abnormal node, the analysis extracts the repeatedly occurring device component node "hydraulic pressure holding system", the key process parameter node "pressure holding pressure", and the abnormal pattern "pressure control fluctuation beyond limit in the pressure holding stage".

[0053] One or more basic countermeasures are matched for each abnormal pattern in combination with the device maintenance manual and the process expert rule base. The device maintenance manual is stored in a structured document form and contains an entry for the "hydraulic pressure holding system", and the suggested measures include "check and clean the pressure holding pressure proportional valve" and "calibrate the pressure sensor PS-205". The process expert rule base contains a series of rules in the form of "IF <abnormal condition> THEN <adjustment action>", and for the condition of "pressure control fluctuation beyond limit in the pressure holding stage", the matched rule suggests the adjustment action of "lower the pressure holding pressure set value by 5% and enable pressure curve smoothing filtering". All the matched basic countermeasures are collected to form the original strategy set, and the original strategy set includes the measure A "check and clean the pressure holding pressure proportional valve", the measure B "calibrate the pressure sensor PS-205", and the measure C "lower the pressure holding pressure set value by 5% and enable pressure curve smoothing filtering".

[0054] In a specific implementation, the feasibility of implementing each measure in the original set of measures under the current production order constraints is evaluated. The process requirement file of the current production order is obtained, which specifies that the material for this production is polycarbonate and the product wall thickness is 2.5 mm, and clearly requires that the pressure holding pressure range is 75 to 85 MPa. At the same time, the real-time health index of each device on the production line is obtained, which is read from the device predictive maintenance system, including "hydraulic pressure holding system health" is 82%, "pressure sensor PS-205 calibration state" is "has been running for 1200 hours since the last calibration". Each measure in the original set of measures is checked for compliance with the parameter range in the process requirement file, and candidate measures that do not conflict are selected. Measure C suggests "down 5% of the pressure holding pressure set value", assuming that the current set value is 80 MPa, after down 5% it is 76 MPa, which is still within the range of 75 to 85 MPa specified in the process requirement file, so measure C passes the compliance check. Measures A and B do not involve direct modification of process parameters and are defaulted to pass the compliance check. The candidate measures are matched with the real-time health index of the device to evaluate the potential impact on the health of the device and the risk level of executing the candidate measures. Measure A "check and clean the pressure holding pressure proportional valve" is evaluated in combination with the "hydraulic pressure holding system health" index of 82%, which determines that this maintenance operation has moderate potential benefits and low risk, and is assigned a risk level of "low". Measure B "calibrate pressure sensor PS-205" is evaluated in combination with the "pressure sensor PS-205 calibration state" index of "has been running for 1200 hours since the last calibration", which determines that this calibration operation has high potential benefits and low risk, and is assigned a risk level of "low". Measure C "adjust the pressure holding pressure setting and enable filtering" is a software parameter adjustment that does not involve hardware intervention, and based on process knowledge it is determined to have high potential benefits and extremely low risk, and is assigned a risk level of "extremely low".

[0055] The compliance check results and risk levels are combined to calculate the feasibility score for each measure. The feasibility score calculation formula is as follows:

[0056] Wherein: represents the feasibility score of the th measure, represents the compliance check result, when the check is passed , when it is not passed , represents the numerical value of the risk level quantification, the risk level "extremely low" corresponds to , "low" corresponds to , "medium" corresponds to , and "high" corresponds to , Risk impact coefficient, set as 0.1. Calculate the feasibility score of Measure A . Calculate the feasibility score of Measure B . Calculate the feasibility score of Measure C . Filter and rank the measures according to the feasibility scores, with the ranking rule being the feasibility score from high to low.

[0057] According to the implementation feasibility evaluation results, the measures are ranked and optimized, and the final executable strategy instruction set is output. The executable strategy instruction set adopts a structured JSON format, including strategy ID, target device, specific operation instruction, expected execution time window, and priority. For example, the output instruction set contains three instructions: instruction 1 has a priority of 1, and the content is “immediately adjust the holding pressure set value from 80 MPa to 76 MPa in the control interface, and enable the software filtering function of the pressure control loop”; instruction 2 has a priority of 2, and the content is “schedule offline calibration of pressure sensor PS-205 after the end of this batch production”; instruction 3 has a priority of 3, and the content is “plan to disassemble and clean the proportional valve of the hydraulic holding pressure system during the routine maintenance this weekend”. It can be understood that the executable strategy instruction set converts the analysis conclusion into an operation task that can be directly understood and executed by the on-site personnel or the automatic system. In some embodiments, for software parameter adjustment type instructions, they can be directly issued to the injection molding machine controller through the interface of the manufacturing execution system.

[0058] Referring to Figure 5 In the intelligent analysis of the running state of the injection molding production line equipment, the average value distribution of the readings of the pressure sensor PS-205, the temperature sensor TS-108, the flow sensor FS-302, and the position sensor LS-401 in the four core process stages of injection, holding pressure, cooling, and mold opening is presented. Specifically, in the figure, the average level of each sensor in different process stages is quantified by color gradient (corresponding to the right value scale): the average value of the temperature sensor TS-108 is significantly high (deep purple, close to 250) in the injection and holding pressure stages, the average value decreases in the cooling stage (red, about 150-200), and further decreases in the mold opening stage (orange, about 100-150); the average values of the remaining sensors (pressure, flow, and position) in each process stage are in a relatively low range (light yellow to orange, corresponding to 50-100).

[0059] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0060] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent analysis of the operating status of injection molding production line equipment, characterized in that, Perform the following procedure: The system synchronously collects native equipment operation data within a predetermined time window from the control server and historical database of the target injection molding production line. The native equipment operation data integrates time-series arranged sensor reading streams, event-driven alarm log entries, and periodically recorded product quality spot check reports. The process of performing multi-dimensional in-depth analysis on the native operating data of the device includes: semantic decomposition and event encoding of the alarm log entry set to output a structured alarm event sequence; dividing the sensor reading stream into operating stages and calculating statistical features within each stage to output a multi-dimensional time series feature matrix; and classifying the product quality spot check report into defect patterns and marking quality levels to output a quality status vector. Based on the structured alarm event sequence, multidimensional temporal feature matrix and quality status vector, a device operation knowledge graph oriented to the current time window is dynamically constructed. The nodes of the device operation knowledge graph represent device entities, process parameters and alarm events, and the edges represent the temporal, causal or logical relationships between nodes.

2. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 1, characterized in that, The multi-dimensional deep analysis of the device's native operating data includes the following steps: Each text record in the alarm log entry set is parsed, and the device component name, alarm code, and status description are extracted using named entity recognition technology. The extracted device component names, alarm codes, and status descriptors are mapped and matched with a predefined device ontology library, transforming unstructured text records into structured alarm events with unified semantics. All the structured alarm events are sorted and merged according to their timestamps to generate the structured alarm event sequence, wherein adjacent and similar structured alarm events are merged into a composite event with a duration.

3. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 1, characterized in that, The multi-dimensional in-depth analysis of the device's native operating data also includes the following steps: Identify specific signal patterns in the sensor reading stream that characterize the start and end of the injection cycle, and cut the continuous sensor reading stream into continuous single injection cycle data segments based on the specific signal patterns; For each of the aforementioned individual injection molding cycle data segments, it is further divided into preset process stages: injection, holding pressure, cooling, and mold opening; Calculate the statistical characteristics of the readings of each sensor channel within each preset process stage. The statistical characteristics include mean, variance, peak value, and trend slope. Organize all statistical characteristics by stage and sensor channel to form the multidimensional time series feature matrix.

4. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 1, characterized in that, The process of dynamically constructing a device operation knowledge graph oriented towards the current time window based on the structured alarm event sequence, multi-dimensional time-series feature matrix, and quality status vector includes the following steps: An initial set of graph nodes is created using equipment physical components, sensor measurement points, alarm event types, and product quality defects as node types. Traverse the structured alarm event sequence, add each event as an event node to the graph node set, and establish "triggered by" and "associated" edges between the event node and the related device component node and sensor measurement point node; Traverse the multidimensional time series feature matrix, create feature anomaly nodes for statistical features that exceed a preset threshold, and establish "occurred at" edges between the feature anomaly nodes and the corresponding sensor measurement point nodes and process stage nodes; Based on the quality state vector, a quality defect node is created, and according to the time alignment relationship, a "possible origin" edge is established between the quality defect node and the event node and feature anomaly node within a specific time interval.

5. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 4, characterized in that, The method also includes a process of enhancing and reasoning the device operation knowledge graph: In the device operation knowledge graph, multi-hop path search is performed along the edges of "triggered by", "associated with", and "may originate from" to identify common sub-paths that connect multiple feature abnormal nodes and event nodes and ultimately point to the same quality defect node; For each of the identified common sub-paths, a path confidence score is calculated, which is determined based on the historical frequency and conditional probability of node co-occurrence on the common sub-path. The common sub-paths with confidence levels exceeding a threshold are solidified into new causal edges with stronger semantics and added to the device operation knowledge graph, thereby forming an enhanced device operation causal knowledge graph.

6. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 5, characterized in that, The method also includes a process of performing state backtracking analysis based on virtual simulation: From the enhanced device operation causal knowledge graph, extract all high-confidence causal paths ending at a specific quality defect node to form a candidate root cause chain set; For each causal path in the candidate root cause chain set, locate its starting source anomalous node, which corresponds to a specific sensor feature anomaly or alarm event. In the injection molding process simulation model, the equipment parameter deviation or disturbance signal corresponding to the source abnormal node is injected, and the injection molding process simulation model is run to obtain virtual production results; The virtual production results are compared with the actual recorded quality state vector. If the simulated defect matches the actual defect pattern, this causal path is marked as a verified root cause path.

7. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 6, characterized in that, The process of injecting equipment parameter deviation or disturbance signals corresponding to the source abnormal node into the injection molding process simulation model includes: When the source abnormal node is a sensor feature abnormal node, the sensor measurement point, process stage and abnormal feature type associated with the sensor feature abnormal node are analyzed. In the injection molding process simulation model, the corresponding simulated sensors and process stages are located, and the model input parameters or internal states are modified to reproduce the reading deviation behavior described by the abnormal feature type. When the source abnormal node is an alarm event node, the device component and alarm type corresponding to the alarm event node are parsed. In the injection molding process simulation model, the degradation or fault logic of the corresponding equipment component is triggered to simulate the abnormal operating conditions of the equipment represented by the alarm type.

8. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 6, characterized in that, The method also includes a process for generating targeted maintenance and adjustment strategies: The root cause paths that have passed verification are analyzed and the recurring equipment component nodes, key process parameter nodes and their abnormal patterns are extracted from the paths. By combining the equipment maintenance manual and the process expert rule base, one or more basic countermeasures are matched for each abnormal mode to form an original strategy set; For the measures in the original strategy set, evaluate their feasibility of implementation under the current production order constraints, including production cycle time, material attributes, and equipment physical limits; Based on the feasibility assessment results, the measures are ranked and optimized to output the final set of executable strategy instructions.

9. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 8, characterized in that, The assessment of its feasibility under current production order constraints includes: Obtain the process requirements documents for the current production order, as well as the real-time health indicators of each piece of equipment on the production line; Each measure in the original strategy set is checked for compliance with the parameter range in the process requirement document, and candidate measures that do not conflict are selected. The candidate measures are matched with the real-time health indicators of the equipment to assess the potential impact and risk level of implementing the candidate measures on the health of the equipment. Based on the combined compliance verification results and risk level, a feasibility score is calculated for each measure, and the measures are then screened and ranked according to the feasibility scores.

10. The intelligent analysis method for the operating status of injection molding production line equipment according to claim 5, characterized in that, The step of calculating the path confidence for each identified common sub-path includes performing the following steps: Extract the node sequence contained in the common sub-path, the node sequence starting from the source abnormal node or event node, passing through intermediate nodes, and pointing to the terminating quality defect node; Based on the historical equipment operation knowledge graph, the number of times each pair of adjacent nodes in the node sequence co-occurs within the historical time window is counted, and the co-occurrence frequency of each pair of adjacent nodes is calculated. For the node sequence, starting from the starting node, the conditional probability of the downstream node occurring given the occurrence of the upstream node is calculated hop-by-hop based on Bayes' theorem. Multiply the co-occurrence frequency of each pair of adjacent nodes by the conditional probability to obtain the association strength of this pair of adjacent nodes, and multiply the association strengths of all hops on the path together to obtain the original confidence score of the common sub-path. The original confidence scores of all identified common sub-paths are normalized so that the confidence scores of each path fall within the range of 0 to 1.

Images