A method for analyzing the defectiveness of a connector terminal

Abstract

This invention relates to a method for tracing and analyzing defects in connector terminals, aiming to achieve high-precision automatic attribution of terminal defects in complex production processes. By standardizing and parsing equipment logs, automatic detection, and work order records, millisecond-level temporal semantic anchors are extracted. Time alignment of multimodal detection data is then used to establish physical causal chains and a lightweight temporal semantic graph. This method generates attribution paths that conform to time, physical, and process constraints through a limited depth-first traversal. Furthermore, natural language generation technology is used to output specific and traceable descriptions of defect causes, enabling multimodal visual interaction between defect analysis and raw data. Combined with dynamic optimization rules and templates based on engineer feedback, the model possesses self-evolution capabilities, significantly improving attribution accuracy and reliability in engineering applications.

Description

Technical Field

[0001] This invention relates to the field of manufacturing process traceability and defect attribution analysis technology, and in particular to a method for tracing and analyzing defects in connector terminals. Background Technology

[0002] Existing defect tracing and attribution technologies in connector terminal production mainly follow the following mainstream solutions: One type is causal rule reasoning technology based on knowledge bases such as expert systems and rule engines. This technology solidifies the engineering causal logic between each link in the form of rule entries by manually sorting out key information such as production process, equipment operating conditions, and process parameter settings. It is suitable for tracing typical defects in stable production environments. Another type is data-driven attribution technology based on models such as Bayesian networks, Markov logic chains, and multivariate conditional probability graphs. This technology relies on large-scale historical defect datasets and feature variable annotation results to infer the probability of equipment anomalies and parameter drift on the final defect through probability statistics. In addition, with the popularization of industrial intelligent manufacturing environments, there are also integrated MES (Manufacturing Execution System) and AOI inspection systems that realize end-to-end production process traceability. These systems store and track real-time collected process parameters, equipment status, and detected defects at the data level, providing data support for subsequent post-event attribution.

[0003] In recent years, considering the typical characteristics of terminal defects, such as multiple causes and single effects, process linkage, and spatiotemporal triggering of anomalies, the industry trend is shifting towards an integrated solution for process traceability and defect attribution that combines multi-source production data, real-time detection data, and intelligent reasoning methods. Some cutting-edge literature has proposed a hybrid reasoning framework combining causal Bayesian models and temporal event streams to improve the system's ability to automatically attribute causes and quantify prediction uncertainties. However, overall, Bayesian causal networks, rule engines, and MES-AOI integration solutions remain the mainstream technologies in the current field.

[0004] In practical applications, while rule-based defect attribution methods have advantages in question-and-answer interpretability, they heavily rely on manual knowledge management and lack the ability to generalize to new and complex causal events, making it difficult to dynamically respond to production line adjustments or product iterations. Attribution systems based on machine learning probabilistic models such as Bayesian networks, due to their highly abstract model structures and the difficulty in intuitively reproducing on-site engineering processes through probabilistic relationships, often exhibit "black box inference" phenomena. This makes it difficult for engineers to obtain reliable interpretations of the attribution paths generated by the system, resulting in a "reliability gap." Post-event data tracking by MES and inspection systems, while enabling the retrieval of historical parameters for defective workpieces, lacks a causal chain description with temporal logic and physical constraints. This forces engineers to rely on manually comparing large amounts of raw data evidence, making it difficult to intuitively understand the evolution path of defects, resulting in low efficiency.

[0005] Existing technologies generally suffer from the following prominent problems: Defect attribution paths lack interpretable reasoning mechanisms tailored to the actual production context; mainstream probabilistic models only output probability correlation weights between events, failing to reflect the precise time, equipment number, and physical action of each process step, making it difficult for engineers to directly locate the root cause based on system attribution operations; existing attribution paths are often described as abstract graph structures or fractional weights, lacking natural language and visual explanation capabilities for engineering decision-making scenarios, and lacking the linked display of multimodal data—such as process events, detection images, and electrical performance signals—resulting in low trust levels; the lack of support for attribution result traceability verification and interactive correction prevents engineers from conducting targeted verification and feedback on the details of attribution, hindering the continuous improvement of the system's self-learning and attribution capabilities.

[0006] In summary, in the current field of connector terminal defect attribution and production process traceability, engineers have raised urgent demands for transparency, understandability, and verifiability of attribution outputs. Existing solutions based on probability statistics and static rules fail to efficiently aggregate and semantically map key physical events, temporal contexts, and defect characteristics inherent in the production process, lacking an end-to-end attribution path generation system characterized by "traceability, understandability, and verifiability." Therefore, a new technical approach is urgently needed that abandons the black-box model, fully explores the temporal semantic anchors of production line events, introduces dynamic data and automatic natural language interpretation, and supports interactive correction by engineers. This approach aims to overcome the key technical obstacles of insufficient explanation, low reliability, and limited efficiency in existing attribution methods. Summary of the Invention

[0007] This application provides a method for tracing and analyzing defects in connector terminals, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0008] This application provides a method for tracing and analyzing defects in connector terminals, specifically including: S1: Extract key process events with clear physical meaning and millisecond-level time accuracy from the programmable logic controller logs of connector equipment, automatic optical inspection timestamps, and manufacturing execution system work order records, and generate a set of time sequence semantic anchor points; S2: Based on the timestamp information in the set of temporal semantic anchor points, multi-source time alignment technology is used to accurately associate the coordinates of the thermal defects, the three-dimensional point cloud deformation area, and the electrical performance abnormality interval of the same workpiece in images collected at different stages with the nearest neighbor of several temporal semantic anchor points to generate a preliminary screening causal chain. S3: Using the temporal semantic anchor points in the initial screening causal chain as nodes and the physical action direction and process logic flow direction as edge types, construct a directed heterogeneous graph to generate a lightweight temporal semantic graph. S4: For the new input defect sample, perform a restricted depth-first traversal along the lightweight temporal semantic graph, retain only the path sequence that simultaneously satisfies the triple checks of temporal sequence constraints, physical operability, and process reachability, and generate a candidate attribution path set that conforms to the manufacturing process ontology logic; S5: Call the pre-trained process semantic generator to process each path sequence in the candidate attribution path set, transform the abstract process relationship into a natural language description, generate attribution result text with spatiotemporal annotations, and complete the source tracing analysis of connector terminal defects. S6: Synchronously highlight the attribution result text and the original temporal semantic anchor point set, establish corresponding jump links, and generate a multimodal visualization interface that supports interactive semantic verification by engineers. S7: Receive feedback instructions submitted by engineers on the multimodal visualization interface, including acceptance, questioning, or supplementary information. Based on the type of feedback instruction, dynamically adjust the definition rules of the time-series semantic anchor points and the parameter configuration of the semantic generation template library in the process semantic generator, and generate an optimized anchor point definition rule set and semantic generation template library. S8: Based on the optimized anchor point definition rule set and semantic generation template library, update the node attributes and edge type descriptions of the lightweight temporal semantic graph to complete the reconstruction of the engineering credibility foundation of the attribution logic.

[0009] The method for tracing and analyzing defects in connector terminals provided in this application has the following beneficial effects: (1) By constructing an attribution reasoning mechanism based on temporal semantic anchors, the interpretability and engineering credibility of the defect tracing path of connector terminals are significantly improved. Traditional methods rely on black-box probability models or static knowledge bases for cause inference, which often makes it difficult to provide attribution basis with physical meaning and temporal certainty, resulting in a lack of on-site verifiability and trust from process personnel. This solution abandons such uncertainty modeling paradigms and instead starts from the manufacturing process ontology, accurately extracting key events with clear physical meaning and millisecond-level temporal accuracy from multi-source data such as PLC logs, AOI inspection, and MES system as semantic anchors, such as "crimping action completed", "current over-limit for 200ms", and "abnormal springback signal record". Each anchor carries a unique identifier, equipment information, process context and original parameter snapshot, forming a traceable, locatable and reproducible data foundation. Based on this, by using multi-source time alignment technology, defect features (such as image hotspots and 3D deformation areas) are reverse-correlated to the nearest neighbor process events, establishing a preliminary causal chain of "defect phenomenon → trigger anchor point → transmission anchor point → final state anchor point". This makes the attribution path no longer rely on statistical correlation inference, but is rooted in the observable and verifiable event sequence in the real production process. This fundamentally overcomes the problem of misattribution caused by traditional methods ignoring the time dynamics and physical action direction, and greatly improves the on-site adaptability and technical persuasiveness of the attribution conclusions.

[0010] (2) Relying on a lightweight temporal semantic graph and a restricted traversal mechanism, the system achieves a dual enhancement of logical rigor in attribution path generation and natural readability in semantic expression. Unlike conventional knowledge graphs that introduce embedded representations or probability weights, which introduce uninterpretable risks, the graph constructed in this scheme uses semantic anchors as nodes and "physical action direction" (such as heat conduction and stress transfer) and "process logic flow" as edge types to form a directed heterogeneous structure. The graph does not store fuzzy confidence levels, but instead solidifies the description of the process influence relationship after calibration (such as "positive reinforcement" and "threshold triggering"), ensuring that the reasoning process is entirely based on known engineering laws. When a new defect sample is input, the system performs a restricted depth-first traversal, retaining only paths that satisfy the triple verification of "time sequence constraint + physical operability + process reachability", effectively eliminating false associations that do not conform to the manufacturing mechanism. Furthermore, by combining a pre-trained process semantic generator, abstract anchor sequence is transformed into natural descriptions that closely resemble engineers' language habits. For example, "Due to a 0.08mm deviation in the X-axis of the No. 3 crimping machine, local stress concentration is triggered, which in turn induces microcrack propagation during tin plating, ultimately resulting in a 12.3% decrease in insertion and extraction force," achieving a semantic leap from data to decision-making recommendations. This mechanism not only ensures the technical rationality of the reasoning path but also significantly lowers the interpretation threshold, enabling process personnel with unconventional backgrounds to quickly understand and adopt the analysis results, effectively promoting the practical application of intelligent diagnostic systems on the production line.

[0011] (3) The introduction of an interactive semantic verification closed loop enhances the system's adaptive evolution capability and human-machine collaboration efficiency. Existing attribution systems generally lack effective channels for integrating expert experience, resulting in models becoming rigid once deployed and difficult to dynamically optimize with process changes or the emergence of new failure modes. This solution provides a visual interactive interface after attribution output, highlighting the generated natural language description along with the original anchor data, image heatmaps, and time waveforms. Engineers can click on any semantic fragment to jump to the corresponding original data source, achieving one-click penetration verification of "conclusion-evidence". At the same time, three types of feedback entry points are set up: "acceptance / questioning / supplementary evidence", allowing users to mark correction opinions for suspicious links. All feedback will be used to dynamically adjust the anchor definition rules and semantic generation template library, forming a virtuous cycle of "data-driven discovery - manual verification and confirmation - continuous rule evolution". This design not only improves the robustness and long-term availability of the system, but also establishes a collaborative paradigm of mutual trust between humans and machines, enabling the system not only to answer "what is the reason", but also to continuously learn "why this judgment is made" in practice, thereby gradually evolving into a process intelligence agent with domain cognition capabilities.

[0012] In summary, this solution, by constructing an interpretable attribution system centered on temporal semantic anchors, achieves a reliable transformation from raw manufacturing data to engineering-level causal explanations. Without introducing complex probabilistic modeling or counterfactual reasoning, it significantly improves the accuracy, transparency, and practicality of defect tracing. It is particularly suitable for high-end electronic manufacturing scenarios with stringent reliability requirements, providing a practical and feasible technical path for achieving interpretability, auditability, and sustainable evolution of intelligent manufacturing systems. Attached Figure Description

[0013] Figure 1 This is the main flowchart of a method for tracing and analyzing defects in connector terminals.

[0014] Figure 2 This is a sub-flowchart of a method for tracing and analyzing defects in connector terminals.

[0015] Figure 3 This is another sub-flowchart of a method for tracing and analyzing defects in connector terminals. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0018] like Figure 1 As shown, this application provides a method for tracing and analyzing defects in connector terminals, specifically including: S1: Extract key process events with clear physical meaning and millisecond-level time accuracy from the programmable logic controller logs of connector equipment, automatic optical inspection timestamps, and manufacturing execution system work order records, and generate a set of time sequence semantic anchor points; S2: Based on the timestamp information in the set of temporal semantic anchor points, multi-source time alignment technology is used to accurately associate the coordinates of the thermal defects, the three-dimensional point cloud deformation area, and the electrical performance abnormality interval of the same workpiece in images collected at different stages with the nearest neighbor of several temporal semantic anchor points to generate a preliminary screening causal chain. S3: Using the temporal semantic anchor points in the initial screening causal chain as nodes and the physical action direction and process logic flow direction as edge types, construct a directed heterogeneous graph to generate a lightweight temporal semantic graph. S4: For the new input defect sample, perform a restricted depth-first traversal along the lightweight temporal semantic graph, retain only the path sequence that simultaneously satisfies the triple checks of temporal sequence constraints, physical operability, and process reachability, and generate a candidate attribution path set that conforms to the manufacturing process ontology logic; S5: Call the pre-trained process semantic generator to process each path sequence in the candidate attribution path set, transform the abstract process relationship into a natural language description, generate attribution result text with spatiotemporal annotations, and complete the source tracing analysis of connector terminal defects. S6: Synchronously highlight the attribution result text and the original temporal semantic anchor point set, establish corresponding jump links, and generate a multimodal visualization interface that supports interactive semantic verification by engineers. S7: Receive feedback instructions submitted by engineers on the multimodal visualization interface, including acceptance, questioning, or supplementary information. Based on the type of feedback instruction, dynamically adjust the definition rules of the time-series semantic anchor points and the parameter configuration of the semantic generation template library in the process semantic generator, and generate an optimized anchor point definition rule set and semantic generation template library. S8: Based on the optimized anchor point definition rule set and semantic generation template library, update the node attributes and edge type descriptions of the lightweight temporal semantic graph to complete the reconstruction of the engineering credibility foundation of the attribution logic.

[0019] Step S1: Extract key process events with clear physical meaning and millisecond-level time accuracy from the connector equipment's programmable logic controller logs, automatic optical inspection timestamps, and manufacturing execution system work order records, and generate a set of time-series semantic anchor points. Specifically, this includes: S1.1: Perform multi-protocol parsing and format standardization on the log stream of the equipment programmable logic controller, the timestamp sequence of the automatic optical inspection system, and the work order record data of the manufacturing execution system to eliminate time baseline deviations and field definition differences of heterogeneous data sources and generate a multi-source original process dataset with a unified time base.

[0020] The communication protocols of the three types of data sources—the log stream of the programmable logic controller, the timestamp sequence of the automatic optical inspection system, and the work order record data of the manufacturing execution system—are parsed one by one using a multi-protocol parsing engine. The original fields and their corresponding timestamps, equipment codes, and process numbers are extracted to form a readable temporary data structure.

[0021] For different data sources after parsing, field format standardization is performed, unifying numeric fields to the same unit of measurement and using consistent decimal precision constraints, and using unified character encoding and field naming rules for string fields to eliminate differences in field definitions caused by different system designs.

[0022] To perform time baseline correction on the timestamp field, first read the current system clock deviation parameters of each data source, then use the synchronization reference clock signal or a known reference time event to calculate the correction offset value and apply it to all records to achieve millisecond-level time accuracy consistency.

[0023] To ensure the stability of time alignment across data sources, an interpolation-based time drift compensation method is adopted to homogenize log time series that exhibit nonlinear drift over short periods, thereby reducing time sequence misalignment during cross-source event comparison.

[0024] Perform consistency checks on the standardized and time-corrected datasets to detect missing fields, duplicate records, invalid timestamps, and equipment number mapping errors, and merge the data that passes the checks into a multi-source original process dataset with a unified time base.

[0025] Through multi-protocol parsing and format standardization, the raw heterogeneous data from the previous step is transformed into a multi-source raw process dataset with a unified time base and consistent field definitions, thus providing time-consistent and field-compatible data input for subsequent physical semantic rule extraction.

[0026] For example, in a connector terminal production line, the PLC log stream is output using the Modbus TCP protocol, the AOI system timestamp sequence uses the NTP synchronization format, and the MES work order records are stored in JSON file format. A multi-protocol parsing engine loads the Modbus parsing module, the NTP parsing module, and the JSON parsing module respectively, representing the crimping force value in the PLC record using a unified unit N·mm with a decimal precision of 0.001, converting the timestamps in the AOI inspection to UTC milliseconds, and mapping the process number field of the MES work order to the number in the standard process code table. During the time baseline correction process, by comparing the three types of data with the reference GPS time event, the PLC log deviation is calculated to be +12ms, and the AOI event deviation is... With an 8ms error and a MES recording deviation of +3ms, the timestamp accuracy of each data point was improved to the millisecond level after applying correction values. To address time drift in AOI records, linear interpolation was used to compensate for drift values ​​not exceeding 2ms within a 5-second sliding window. Finally, after consistency verification, a unified time-based multi-source raw process dataset was merged, including fields such as pressure force value, tin plating current, hot zone image index, and work order execution order. This provides high-quality time-series data input for subsequent sliding window feature extraction and threshold-triggered detection in S1.2.

[0027] S1.2: Based on the multi-source original process dataset with the unified time benchmark, a sliding window feature extraction and threshold trigger detection are performed on the continuous data stream using a predefined physical semantic rule library to identify key process state change points with clear physical meaning and time accuracy down to the millisecond level, and to generate a list of candidate key process events.

[0028] Based on a unified time base, multi-source raw process datasets are continuously read into and synchronously buffered for high-frequency sampling signals from the equipment's programmable logic controller, imaging time series from the automated optical inspection system, and work order execution records from the manufacturing execution system, to ensure that all signals are precisely aligned on the millisecond-level time axis.

[0029] The synchronized multi-source signals are divided into windows according to a preset sliding window length and step size. The window length is set based on the shortest observable period of process state changes, and the step size depends on the equipment sampling rate and process dynamic response characteristics, so as to capture the complete state evolution process within each window.

[0030] In each segmented window, a predefined physical semantic rule base is invoked for feature extraction. The rule base includes physical quantity detection modes for different processes, such as transient crimping force, coating current fluctuation, and abnormal springback of insertion and extraction force, which map the original signal into a set of feature parameters with physical meaning.

[0031] Threshold-triggered detection is performed on the extracted feature parameter set. The threshold is dynamically set based on the process tolerance range and equipment calibration data. Change points in the state are identified by comparing the relationship between the feature parameters and the threshold. The change determination formula is as follows: in, Expressed as the mean of mutations, This represents the average value of physical quantities within the current window. This is the reference mean of the previous time window. When the threshold is exceeded, it is determined to be a state change.

[0032] The state mutation points obtained through the above detection are recorded with time index under a unified time base to form candidate key process event entries including mutation time, process category, equipment identifier and mutation parameters, which serve as inputs for subsequent unique identifier encoding and process context association mapping.

[0033] By using sliding window feature extraction and threshold-triggered detection, the multi-source raw process dataset from the previous step is transformed into a list of candidate key process events that can characterize changes in physical state, enabling accurate identification of process state abrupt changes with clear physical meaning and millisecond-level time precision.

[0034] For example, in a connector terminal production line, the PLC sampling frequency is set to 500Hz, the AOI detection image frame time interval is 8ms, and the MES work order recording accuracy is 1ms. After unifying the time base, the crimping force signal is processed using a sliding window with a window length of 20ms and a step size of 10ms. The physical semantic rules identify the average crimping force within the window as P=153.2N, and the reference average value of the previous window is... =150.0N, calculate for Exceeding the set threshold The peak current change ΔI was detected in the coating current signal. A, exceeding the threshold A; The change in rebound force ΔF detected in the insertion and extraction force test record is... N exceeds the threshold N. These mutations are recorded as candidate critical process events, with attached process category and equipment identifier, providing highly reliable data for the subsequent generation of structured process event objects.

[0035] S1.3: Perform unique identifier encoding and process context association mapping processing on each event entry in the candidate critical process event list to bind the process name, equipment asset number and standard operating procedure version information, and generate a structured process event object set carrying complete metadata description.

[0036] S1.4: Based on the structured process event object set, use high-precision time slicing technology to capture snapshots of the original sensor parameters within a preset time window before and after the event occurs, so as to solidify the instantaneous operating condition data at the time of the event and generate an enhanced process event record including static attributes and dynamic parameter snapshots.

[0037] Based on the static metadata already included in the structured process event object set, such as unique identifiers, processes, equipment asset numbers, and standard operating procedure version information, the high-precision time slicing control module is invoked to perform millisecond-level positioning of the event timestamp to determine the center point time of the slicing range.

[0038] The center point time is input to the time window generator, and the time window lengths L1 and L2 before and after are set according to the physical duration characteristics of the event type and the sensor response delay parameters, forming a bidirectional symmetrical or asymmetrical time interception interval.

[0039] Within the time truncation interval, the raw parameter stream matching the event-bound device identifier is retrieved from the multi-source sensor data bus, including the instantaneous stress curve of the mechanical sensor, the transient current and voltage sampling sequence of the electrical sensor, the surface light intensity reflectivity change record of the optical sensor, etc., and arranged in the order of timestamps to form a parameter snapshot matrix.

[0040] The parameter snapshot matrix is ​​subjected to instantaneous condition solidification processing, and the sampling gap is filled by numerical interpolation. Anti-aliasing filtering is performed when necessary to ensure that the state of the event and its preceding and following boundaries can be completely recorded.

[0041] By fusing static metadata of events with a solidified parameter snapshot matrix, an enhanced process event record object is constructed. This object includes both static attributes (process name, equipment number, version information) and dynamic attributes (multivariable sensor snapshots with millisecond-level time index), which meets the input requirements for subsequent automatic labeling and efficient retrieval of status tags.

[0042] By using high-precision time slicing technology and instantaneous condition solidification processing, the structured process event objects of the previous step are transformed into enhanced process event records that retain both physical background and fine-grained data, thereby achieving full reproducibility of event states and cross-source consistency.

[0043] For example, in the 5G communication precision connector terminal crimping process, an event timestamp is 14:22:11.083, the equipment number is CRIMP-03, and the process name is crimping. The time window generator sets the previous time window length L1 to 500 milliseconds and the subsequent time window length L2 to 300 milliseconds, forming a cutoff interval of [14:22:10.583, 14:22:11.383]. Within this interval, the mechanical sensor provides a crimping force curve with a sampling frequency of 5 kHz, the electrical sensor provides a drive motor current curve with a sampling frequency of 10 kHz, and the optical sensor provides terminal positioning image data at 25 frames per second. After aligning the three types of data by timestamp, a parameter snapshot matrix is ​​formed, in which the crimping force curve has a peak value of 6.82 kN at the event center point, the corresponding current peak value on the motor current curve is 3.51 A and lasts for 210 ms, and a terminal position deviation of 0.08 mm is detected in the positioning image. Interpolation filled in the missing sampling points in the force curve, and low-pass filtering was applied to the current curve to remove high-frequency noise. The final enhanced process event record object carries static attributes CRIMP-03, crimping, standard operating procedure V1.2, and dynamic attributes in millisecond indexed multivariate snapshots of force, electricity, and light, which can be used for subsequent automatic status labeling and causal path generation.

[0044] S1.5: Perform automatic status labeling and time-series indexing on the enhanced process event records to give each record a clear status classification identifier and establish a fast retrieval mechanism based on timestamps, and finally generate a set of time-series semantic anchor points including a unique identifier, the process to which it belongs, the equipment identifier, the timestamp, the status label, and the original parameter snapshot.

[0045] Step S2: Based on the timestamp information in the set of temporal semantic anchor points, multi-source time alignment technology is used to accurately associate the coordinates of the defect thermal area, the three-dimensional point cloud deformation region, and the electrical performance abnormality interval of the same workpiece acquired at different stages with several nearest-neighbor temporal semantic anchor points, generating an initial screening causal chain. Specifically, this includes: S2.1: Obtain the image defect hot zone coordinate sequence, three-dimensional point cloud deformation region dataset and electrical performance abnormality interval curve under the same workpiece identification. Use high-precision timestamp interpolation to perform time baseline unification processing on the above multi-source heterogeneous detection data to eliminate the timing misalignment caused by the clock source deviation of different detection equipment and generate a multimodal defect feature time series with a unified time reference.

[0046] S2.2: Based on the multimodal defect feature time series with a unified time reference, extract the occurrence center point of each defect feature segment, and use the sliding time window matching method to retrieve several candidate temporal semantic anchors with the smallest time distance from the occurrence center point in the temporal semantic anchor set, and generate a list of spatiotemporal proximity pairs including defect features and potential associated anchors.

[0047] Based on the multimodal defect feature time series with a unified time base, the time axis data of each defect feature segment is loaded into the feature parsing engine, and the time series segmentation module is called to perform center point calculation within the feature segment. The center point time value is determined with millisecond-level precision to ensure the spatiotemporal consistency of the subsequent matching process.

[0048] Using the determined time center point as the query key, the sliding time window matching method is called to set the preset time window length parameter. The time window is then used as the center point to sequentially slide and search the time semantic anchor point set. The anchor point entries with the smallest time difference from the center point and within the window range are identified as candidate matching points.

[0049] During the matching process, a time distance value is calculated for each candidate anchor point. The time distance value is calculated using the following formula: in, Indicates the anchor timestamp. Indicates the timestamp of the center point of the defect feature segment. This is the absolute value of the time difference between the two values; the result is used for sorting.

[0050] All satisfied Candidate anchor points smaller than the time window radius are sorted in ascending order of time distance value. Several anchor points with the smallest time distance are retained to form a subset of candidate anchor points. A spatiotemporal proximity mapping relationship is established between each pair of defect features and candidate anchor points.

[0051] For a subset of candidate anchor points, a list of spatiotemporal neighbor pairs is constructed, including defect feature ID, center point time value, anchor point unique identifier and its timestamp. This list serves as input for subsequent physical causality verification.

[0052] By using sliding time window matching and sorting, the multimodal defect feature time series from the previous step is transformed into a list of spatiotemporal proximity pairs of defect features and potential associated anchor points, thereby achieving precise temporal localization and structured representation of candidate associations.

[0053] S2.3: For each set of data in the spatiotemporal proximity pair list, perform time causality verification based on the physical conduction delay constant of the connector manufacturing process, remove invalid candidate temporal semantic anchors whose timestamps are later than the time of defect occurrence or exceed the maximum hysteresis threshold of physical action, and generate a subset of valid anchors that conform to physical causal logic.

[0054] S2.4: Based on the effective anchor point association subset, the associated temporal semantic anchor points are classified and labeled according to the process stage using the process topology sorting rules, distinguishing between trigger anchor points as the origin of defects, transmission anchor points as intermediate transmission links, and final state anchor points as the final manifestation position, and generating an anchor point classification result set with causal role labels.

[0055] Based on the effective anchor point association subset, the process identifier and process execution sequence number of each anchor point are obtained as classification input conditions. Process topology sorting and parsing are performed on the input anchor point data, and an ordered set of processes is constructed using a pre-stored Standard Operating Procedure (SOP) process sequence template to ensure logical consistency in subsequent classification labeling. The process execution sequence number of each anchor point is compared with the starting process number in the SOP. If a match is found and the status label indicates a defect source triggering condition, a causal role label is assigned to the triggering anchor point. The remaining anchor points are analyzed progressively according to the process topology order, retrieving intermediate process nodes that have a physical transmission relationship with the defect phenomenon, and marking them as transmission anchor points under the condition of meeting the delay window parameters. Anchor points whose process numbers match the ending process numbers in the SOP, and whose status labels are consistent with the defect final state manifestation conditions, are marked as final state anchor points. Causal role label encoding is performed on all classified anchor points, mapping the trigger, transmission, and final state labels to standardized role identifiers, and outputting a set of anchor point classification results with causal role labels. By using process topology sorting rules and role labeling processing methods, the effective anchor point association subset of the previous step is transformed into structured anchor point classification data with causal role labels, thereby realizing the basic role positioning of causal chain construction.

[0056] For example, in the connector terminal production process, using the crimping process number 01, plating process number 04, and insertion / extraction test process number 06 recorded in the PLC log as the standard process sequence template, three anchor points are selected from the effective anchor point association subset: anchor point A (crimping machine positioning deviation), anchor point B (plating current abnormality), and anchor point C (insertion / extraction force abnormality). The process number to which anchor point A belongs is matched with the template's starting process number 01, and the status label matches the trigger condition, thus marking it as a trigger anchor point. The process number 04 to which anchor point B belongs is matched with the process in the template located between the trigger and final states, and based on the physical conduction delay parameter of 5s, it is determined that it has a valid transmission relationship with anchor point A, thus marking it as a conduction anchor point. The process number 06 to which anchor point C belongs is matched with the template's ending process number, and the status label matches the final state performance condition, thus marking it as a final state anchor point. Using role coding rules, the trigger anchor point is coded as RF01, the conduction anchor point as RD04, and the final state anchor point as RE06, forming a structured anchor point classification result set. In this embodiment, the classification results accurately establish the causal role basis of the defect phenomenon from the triggering source through the transmission link to the final state manifestation, improving the accuracy of the causal chain construction and the verifiability for engineers.

[0057] S2.5: Integrate the anchor point classification result set with causal role labels with the original multimodal defect feature time series, and perform chain structure assembly according to the logical order of 'defect phenomenon points to trigger anchor point, trigger anchor point connects to conduction anchor point, conduction anchor point points to final state anchor point' to generate a preliminary causal chain with a complete structure of 'defect phenomenon → trigger anchor point → conduction anchor point → final state anchor point'.

[0058] like Figure 2 As shown, step S3 involves constructing a directed heterogeneous graph using the temporal semantic anchor points in the initial causal chain as nodes and the physical action direction and process logic flow direction as edge types, thereby generating a lightweight temporal semantic graph. Specifically, this includes: S3.1: Obtain the set of temporal semantic anchor points in the initial causal chain, and instantiate the temporal semantic anchor points based on the unique identifier and the process attribute to which they belong, to generate a graph node object including equipment identifier, timestamp and original parameter snapshot attributes, so as to establish the basic entity unit of the lightweight temporal semantic graph.

[0059] S3.2: Based on the temporal sequence and process logic flow of the graph node objects, the causal correlation strength between adjacent temporal semantic anchors is calculated using the physical action mechanism analysis method. A set of candidate directed edges representing heat conduction, stress transfer, or ion migration relationships is generated to initially establish the topological connection structure between nodes.

[0060] S3.3: For each directed edge in the candidate directed edge set, perform directional labeling processing according to the predefined process influence rule base, map the abstract physical association into positive reinforcement, negative inhibition and threshold trigger type attribute labels, and generate process influence directional description data with clear semantic definition to replace the traditional probability weight storage mode.

[0061] The predefined process influence rule base includes: 1. Process-Equipment Rule Index Unit A classification index is established based on process type and equipment category, storing the physical action benchmark paradigm corresponding to different process steps. This allows for rapid matching of process succession patterns between adjacent time nodes, generating combined description vectors, and providing a classification retrieval entry point for rule matching.

[0062] 2. Physical parameter threshold calibration unit Preset the positive reinforcement threshold range, negative inhibition threshold range, and critical trigger threshold for various process physical quantities (crushing force, current, temperature, stress, insertion and extraction force, etc.); solidify the historical extreme values ​​and baseline reference values ​​of each physical quantity as the benchmark for influence calculation and category determination.

[0063] 3. Mechanism of Action Glossary Unit Standardized physical mechanism description entries such as heat conduction, stress transfer, and ion migration are pre-stored and bound to three types of edge attributes. After the rule matching and judgment are completed, the corresponding mechanism entries are automatically associated to form interpretable process causal semantics.

[0064] 4. Normalized Influence Calculation Unit It has a built-in unified dimensional influence calculation formula, which normalizes and calculates the dimensionless influence index based on the measured value of the physical quantity, the baseline value, and the historical maximum / minimum value; it also helps to verify whether the direction and magnitude of parameter changes match the corresponding rule category, thus improving the accuracy of calibration.

[0065] 5. Directional classification and labeling unit Using a combined descriptive vector and influence metric as input, the system automatically determines the appropriate criteria based on rule-based thresholds. Falling into the positive range → calibrating positive reinforcement Falling into the negative range → Calibrate negative suppression Triggered only near critical values ​​→ Calibration threshold triggered output includes standardized process influence directionality description data with attribute labels, mechanism entries, and critical conditions.

[0066] 6. Map adaptation constraint unit Built-in process flow logic constraint rules are used to verify the legality of directed edges across processes and eliminate invalid connections that do not conform to the production process topology; output compliant node and edge datasets to directly support subsequent directed heterogeneous graph assembly, topology indexing and restricted depth-first traversal.

[0067] For each directed edge in the candidate directed edge set, the system receives the physical action mechanism analysis results and node attribute data generated by S3.2, and calls the predefined process influence rule base as the judgment basis. Between the two temporal semantic anchors corresponding to each directed edge, the system reads the process type, equipment category, and original parameter snapshot to form a combined description vector including the rate of change of physical quantities and the process succession mode. Using this combined description vector as input, the system performs rule matching operations, compares it with the action mode thresholds marked in the rule base, and determines the physical association category it conforms to. When the parameter change direction in the combined description vector is consistent with the threshold range of the "positive reinforcement" class in the rule base, the directed edge is marked as a positive reinforcement attribute, and the corresponding physical action mechanism term is recorded. When the parameter change direction in the combined description vector is consistent with the threshold of the "negative suppression" class in the rule base, the directed edge is marked as a negative suppression attribute, and the corresponding influence mechanism description is stored. When the trigger value in the combined description vector crosses the critical value set by the rule base and only activates near the critical value, the directed edge is marked as a threshold trigger type attribute, and the critical value and trigger condition are recorded. S3.4: Utilize graph node objects and directed edges carrying directional descriptions of process influence to perform directed heterogeneous graph assembly operations, constructing a lightweight temporal semantic graph that does not include probability weights but fully preserves the directional information of physical effects, in order to form a structured causal reasoning model that supports restricted depth-first traversal.

[0068] Using the temporal semantic anchor nodes in the initial causal chain and their corresponding process influence directional description data as input, the graph assembly engine is called to load the full set of node attribute objects and the directed edge dataset after directional calibration in memory. The node set is indexed, and a unique key-value mapping table is established for each node to ensure that the node is quickly located and duplicate instances are eliminated during the assembly process.

[0069] The directed edge dataset after directionality calibration is processed by edge type grouping. The three attributes of positive reinforcement, negative inhibition and threshold triggering are stored in independent logical partitions. Each directed edge is appended with a starting node index, an ending node index and a physical action direction symbol to form an edge attribute structure that can be directly called by the traversal method.

[0070] The node-edge connection operation of the directed heterogeneous graph is performed by using the node index mapping table and the grouped directed edge attribute structure. The outgoing edge set of each node is arranged in order according to the physical action direction and the process logic flow direction. At the same time, the original parameter snapshot reference pointer is retained in the graph structure so that the working status of the node can be read in the subsequent reasoning process.

[0071] In the above connection operation, the process logic continuity verification is performed on the edges including the cross-segment connection of the process. Edges that do not conform to the process flow constraints are marked as invalid and removed from the structural assembly process to ensure that all connections in the diagram can be legally traced in the manufacturing process topology.

[0072] After the graph structure is connected, a topology index is constructed, generating a hierarchical index for the node set based on the chronological order and a categorized index based on the physical direction of action. This, combined with the process grouping index, enables multi-dimensional path retrieval, providing an efficient data access path for limited depth-first traversal.

[0073] By using the above-mentioned node instantiation, edge attribute grouping, connection verification, and index construction processing methods, the directional calibration results of the previous step are transformed into a lightweight temporal semantic graph with complete physical action direction information, thereby realizing a structured causal reasoning model that supports restricted depth-first traversal.

[0074] For example, the initial screening causal chain in the connector terminal production line includes 120 temporal semantic anchor nodes, each carrying an equipment number, a millisecond-level timestamp, and a status label. After being calibrated by the process influence rule base, 240 directed edges are formed, including 80 positive reinforcement attribute edges, 100 negative suppression attribute edges, and 60 threshold-triggered edges. During graph assembly, the node index range is set to 1–120, and the edge attribute partitions are assigned directional codes 1, -1, and 0 to represent the three types of physical effects. After the physical effect direction symbols are established, the edge start index and end index are arranged in ascending order, and a snapshot pointer of the original parameters is attached to ensure that working condition data such as crimping force value or plating current value can be read during traversal. During cross-segment connection verification, 15 edges lacking conduction anchors between the start and end processes are detected and removed, ultimately retaining 225 valid directed edges. The constructed topology index includes a time-layered index (e.g., divided into 8 time stages) and a process-grouped index (e.g., crimping, plating, and insertion / extraction force testing are each grouped independently). In a limited depth-first traversal, the starting node can be quickly located and the path can be extended according to the direction of physical action. The output candidate attribution path set significantly improves traversal efficiency and path retrieval accuracy in subsequent calculations.

[0075] S3.5: Based on the topology of the lightweight temporal semantic graph, a consistency check method is performed to detect and correct redundant edges and abnormal nodes that violate the temporal order constraint or physical operability principle, and generate an optimized final lightweight temporal semantic graph to ensure the logical rigor and engineering accuracy of the subsequent candidate attribution path set generation.

[0076] Based on the constructed lightweight temporal semantic graph topology, a consistency check method is performed on all node and edge relationship data for input initialization, and the constraints to be checked are defined, including time sequence constraints, physical operability principles, and process logic flow rules.

[0077] The node timestamp sequence is sorted, and the time difference calculation module is used to detect directed edges that violate the time sequence constraint. If the start timestamp of an edge is later than the end timestamp, the edge is marked as a redundant edge and added to the set of edges to be corrected.

[0078] For edges marked as redundant, perform physical availability verification. Call the physical action mechanism matcher to compare the directional description of the edge with the predefined physical action mechanism library one by one to determine whether its physical relationship is consistent with the actual process transmission path. If they are inconsistent, the edge is confirmed as an abnormal edge and enters the deletion queue.

[0079] Perform process logic flow consistency checks on the entire node set. Use the process topology verifier to compare the position index of the process to which the node belongs in the standard operating procedure, identify nodes that cross non-adjacent processes and are not supported by intermediate transmission anchors, mark such nodes as abnormal nodes and add them to the set to be removed.

[0080] Perform topology reconstruction on the set of edges to be corrected and the set of nodes to be removed. Use graph reconstruction to remove invalid nodes and edges, and insert complete nodes at necessary positions to restore the connectivity of valid paths and update the graph topology.

[0081] By using consistency verification and topology reconstruction, the lightweight temporal semantic graph from the previous step is optimized into a final lightweight temporal semantic graph that does not contain content that violates temporal order constraints or physical operability principles, thereby achieving logical rigor and engineering accuracy in the candidate attribution path set generation process.

[0082] like Figure 3 As shown, step S4 involves performing a restricted depth-first traversal along the lightweight temporal semantic graph for each new input defect sample, retaining only path sequences that simultaneously satisfy the triple checks of temporal sequence constraints, physical operability, and process reachability, thereby generating a candidate attribution path set that conforms to the manufacturing process ontology logic. Specifically, this includes: S4.1: Obtain the spatiotemporal feature vector of the new defect sample and the full node data of the lightweight temporal semantic graph. Perform time window slicing on the graph nodes based on the defect occurrence timestamp to remove all invalid nodes whose timestamps are later than the defect occurrence time or earlier than the start time of the first process, and generate a temporal subgraph structure with time sequence constraint attributes.

[0083] S4.2: Receive the starting defect node and the ending process node in the time sequence subgraph structure, and use the predefined physical action direction edge type matching method to perform physical mechanism consistency verification on the directed edges connecting adjacent nodes, so as to eliminate virtual connected edges that do not have clear physical associations such as heat conduction, stress transmission or ion migration, and generate a connected subgraph including physical actionability markers.

[0084] It receives the starting defect node and the ending process node in the time sequence subgraph structure with time sequence constraints, and calls the node data solving module to extract the node attribute vector including equipment identifier, process code and original parameter snapshot.

[0085] Based on the node attribute vector, a predefined set of rules for matching edge types in physical action direction is loaded. All candidate directed edges connecting the starting defect node and the ending process node are retrieved according to edge type, and the physical action mechanism parameter field of each edge is extracted.

[0086] For each candidate directed edge, a physical mechanism consistency check is performed. Parameters from three types of physical models—heat conduction, stress transfer, and ion migration—are used to quantitatively verify the edge's action mechanism. The matching degree of the action mechanism is calculated using the following formula: Where M is the physical action matching degree, q is the physical action mechanism parameter value, and w is the edge action weight. If M is lower than the preset threshold, it is determined to be an invalid edge and is removed.

[0087] For the set of invalid edges after removal, record the reason for removal and the index information of the relevant node pairs so as to avoid invalid physical connections when generating subsequent attribution paths.

[0088] Mark the physical availability field on the retained set of valid edges, bind the direction type of the edges to the physical model verification results, and generate a connected subgraph object including positive valid connection identifiers.

[0089] Through the above processing method, the temporal subgraph structure of the previous step is transformed into a connected subgraph with physical mechanism consistency verification results and actionability markers, thereby achieving the filtering and enhancement of physical constraints in attribution path generation.

[0090] S4.3: Based on the node process attribute labels in the connected subgraph, perform manufacturing process topology sorting verification, compare the logical order of the process to which the node belongs in the standard operating procedure, eliminate jump paths that cross non-adjacent processes and have no intermediate transmission anchor points, and generate a set of legal path candidates that meet the process reachability constraints.

[0091] S4.4: For each path sequence in the set of legal path candidates, start the restricted depth-first traversal engine, trace back from the final state anchor point to the trigger anchor point, and dynamically accumulate the process influence directional description attributes on the path during the traversal process to synthesize a complete causal transmission chain and generate the original attribution path sequence with triple verification pass mark.

[0092] For the path sequence in the set of legal path candidates, the system receives the node attribute data of the final anchor point and the trigger anchor point, as well as the directed edge attributes marked with directionality. It then calls a restricted depth-first traversal engine to build a traversal stack in reverse path search mode, setting the final anchor point as the top node and initializing the path accumulator to an empty state. For the current node in the traversal stack, a backtracking instruction is executed. Predecessor nodes that satisfy the conditions of physical action direction consistency and timestamp decrementation are selected from the current node's incoming edges. These predecessor nodes are pushed onto the traversal stack according to the process logic order, and the process influence directionality description attribute of the incoming edge is appended to the accumulator. A dynamic accumulation operation is performed on the appended process influence directionality description attributes, merging the directionality descriptions between adjacent nodes into a continuous causal relationship segment in the form of a structured chain. During the accumulation process, a time sequence constraint verification flag is maintained to ensure the temporal consistency of the path. The traversal stack is iteratively backtracked until the trigger anchor point node is pushed onto the stack and passes the triple verification flag check, generating the original attribution path sequence including a complete causal direction description. By using the above reverse traversal and dynamic accumulation processing method, the set of legal path candidates from the previous step is transformed into the original attribution path sequence with the directional attribute of process influence, thereby realizing the complete construction of the causal transmission chain and the verification of triple constraints.

[0093] For example, for a path sequence in a set of legal path candidates, with the final anchor node timestamp being 14:23:05.821, the trigger anchor node timestamp being 14:22:11.000, and the node status labels being "insertion / extraction force attenuation" and "pressing offset" respectively, the constrained depth-first traversal engine calls the reverse search mode to establish a traversal stack. The traversal engine searches for predecessor nodes according to the incoming edge direction, finding that the coating microcrack node timestamp is 14:22:45.000 and its directional description is negative suppression. A push operation is performed, and this attribute is appended to the accumulator. The accumulator merges the negative suppression attribute with the directional description chain from the final anchor node to the coating microcrack node, and then appends the positive reinforcement attribute from the coating microcrack node to the pressing offset node, resulting in two continuous causal relationship chains. During the accumulation process, the timestamps of the three nodes are checked by the time sequence constraint verification flag to ensure that they sequentially satisfy the decreasing condition, and that the physical action directionality and process reachability both conform to the rules. The final generated original attribution path sequence is "press-fit offset → positive reinforcement → coating microcrack → negative suppression → insertion / extraction force attenuation", with a triple verification pass marker added to the path sequence. In the subsequent natural language generation step, this path sequence is parsed into a causal transmission chain including equipment number, positioning deviation value, stress concentration description, and performance degradation data, realizing the engineering verifiability of the attribution path.

[0094] S4.5: Perform redundancy pruning and confidence ranking on the original attribution path sequence. Based on the principle of shortest path length and the abnormal significance weight of node state labels, remove suboptimal paths including loops or low-relevance nodes to output a concise and logically rigorous candidate attribution path set as the direct input object for subsequent natural language narrative generation.

[0095] For the original attribution path sequence generated by triple verification, the path structure data and node status label weight matrix are loaded, and a path redundancy detection index is established for subsequent structure optimization processing.

[0096] The cyclic topology detection method is performed on the path structure data to calculate the adjacency matrix of the node sequence in each path, detect paths with closed loop structures, and mark closed loop paths as redundant and add them to the deletion list.

[0097] After redundancy detection is completed, path relevance assessment is performed based on the abnormal significance weight of node status labels. The cumulative weight scoring method is used to calculate the overall relevance score of the path. Paths with relevance scores lower than a preset threshold are considered suboptimal paths and are placed into the elimination set.

[0098] The length of the retained path set is optimized. Based on the shortest path principle, the restricted depth-first traversal engine is called to calculate the number of nodes and edges of each path. Paths whose length exceeds the shortest path allowable difference threshold are judged as redundant structures and removed.

[0099] After redundancy removal and length optimization, the path set is sorted by confidence. Through the above redundancy pruning and confidence sorting, the original attribution path sequence in step S4.4 is transformed into a candidate attribution path set with a concise structure and rigorous logic, thus achieving the expected technical effects of path structure optimization and confidence enhancement.

[0100] Step S5: The pre-trained process semantic generator is invoked to process each path sequence in the candidate attribution path set, transforming abstract process relationships into natural language descriptions, generating attribution result text with spatiotemporal annotations, and completing the source tracing analysis of connector terminal defects. Specifically, this includes: S5.1: Obtain the original attribution path sequence from the candidate attribution path set, and use the path node attribute parsing engine to perform structured field extraction processing on the temporal semantic anchor nodes in the sequence to separate the device asset number, millisecond-level timestamp, status label and original parameter snapshot data, and generate a standardized path element set including complete physical entity information.

[0101] S5.2: Based on the standardized path element set, the directed edge attributes between adjacent nodes are semantically mapped and queried using the process influence directionality description data matching method, so as to convert the positive reinforcement, negative inhibition and threshold trigger type attribute labels into corresponding physical action mechanism description phrases, and generate a semantic connection word library carrying clear causal logical relationships.

[0102] Based on the node attribute fields in the standardized path element set, the process influence directionality description data matching method is called to perform a one-by-one search on the directed edge attributes of adjacent nodes in each candidate attribution path, and the edge attribute labels are precisely matched and verified with the preset physical action mechanism description phrase mapping table to obtain the corresponding physical mechanism candidate set.

[0103] For the matched positive reinforcement attribute tags, the fixed template entries in the physical action mechanism description phrase library are used to parse the tag into a specific positive action mechanism description phrase. Combined with the time sequence constraint information, the tag is bound to the process logic context of the parsed node to ensure the continuity and direction consistency of the description.

[0104] For the matched negative inhibition attribute tags, the reverse effect adjustment mechanism is executed. The corresponding physical description of the effect weakening, blocking or attenuation type is retrieved from the phrase library, and the numerical consistency is verified with the original parameter snapshot of the adjacent nodes to ensure that the generated semantic description is consistent with the actual sensor monitoring results.

[0105] For the matched threshold-triggered attribute tags, load the predefined process parameter threshold formula, calculate the precise physical action mechanism determination result under the triggering condition, such as the current, temperature or stress value at the triggering time, and convert the determination result into a semantic phrase describing the state of the threshold condition being met.

[0106] The three types of physical action mechanisms described above are processed into phrases using a unified coding rule to generate a semantic linking lexicon with clear causal logic. Each phrase is labeled with the type of process influence directionality attribute, applicable process stage, and matching accuracy score. In this way, the structured path elements of the previous step are converted into intermediate semantic resources that can be used for natural language generation, thus bridging the attribution logic to semantic narrative.

[0107] For example, for an attribution path consisting of three directed edges, the first edge has a positive reinforcement attribute, the physical transmission mechanism between nodes is stress transfer, and the matched stress concentration descriptive phrase is "stress concentration leads to narrowing of structural gaps". The second edge has a negative suppression attribute, corresponding to a heat conduction weakening mechanism, and the matched descriptive phrase is "temperature instability inhibits subsequent coating curing". The third edge has a threshold-triggered attribute, the trigger condition being that the coating current exceeds the limit, which is determined by the formula... in This is the actual current value. The reference current value, For the detected overcurrent, the calculated result I=12.3A is higher than the benchmark value of 10A and meets the threshold triggering condition. The matched descriptive phrase is "overcurrent triggers coating crack propagation". The above three phrases have all been verified with a matching accuracy score higher than 0.9, and their respective process stages and directional attributes are marked. The final semantic connection lexicon includes three phrase entries: "stress concentration leads to structural gap reduction", "temperature instability inhibits subsequent coating curing", and "overcurrent triggers coating crack propagation". This meets the input requirements for subsequent natural language template filling and significantly improves the transparency and verifiability of the physical mechanism of the attribution description.

[0108] S5.3: For each logical link in the semantic connection lexicon, the pre-trained process semantic generator template filling module is invoked to inject the positioning deviation value and the original parameter snapshot data in the standardized path element set into the preset natural language sentence slots to construct a preliminary causal narrative fragment including specific quantitative indicators and generate a semi-structured semantic description stream with data support.

[0109] The process semantic generator includes: 1. Path Node Structure Parsing Module The built-in path node attribute parsing engine receives candidate attribution path sequences, performs structured field decomposition on temporal semantic anchor nodes, accurately extracts equipment asset numbers, millisecond-level timestamps, process status tags, and original parameter snapshots, and integrates them to form a standardized set of path elements containing complete physical entity and operating condition information, providing a structured data source for subsequent semantic mapping and text generation.

[0110] 2. Semantic mapping module for process mechanism Equipped with a process influence directionality matching method and a pre-defined physical mechanism terminology library, this system identifies the directed edge attributes of adjacent nodes in the attribution path, distinguishing between three types of mechanisms: positive reinforcement, negative inhibition, and threshold triggering. By combining process threshold formulas and parameter snapshot verification, abstract edge attribute labels are converted into standardized physical causal description terms, constructing a semantic connectivity terminology library with type annotations, accuracy scores, and process adaptation attributes, thus completing the semantic transformation of process mechanisms.

[0111] 3. Template Fill Module As the core template-driven unit of the generator, it incorporates natural language sentence templates for various process scenarios. It performs accuracy correction, unit unification, and noise filtering on positioning deviation, peak parameters, and status snapshot data, accurately injects quantitative process indicators into sentence slots, and performs slot data consistency and dimensional matching verification to generate a semi-structured causal semantic description stream supported by real quantitative data.

[0112] 4. Spatiotemporal context fusion module An integrated time series alignment method is used to bind millisecond-level timestamps with process flow logic. According to the process sequence of defect triggering, propagation, and final state evolution, discrete semi-structured description fragments are concatenated and integrated, and time sequence adverbs and process context are supplemented to generate a process description text with complete spatiotemporal coordinates and continuous causal logic.

[0113] 5. Natural Language Post-Optimization Module It includes a built-in domain-specific terminology rule library, timing logic verification formulas, and a number formatter. It unifies domain-wide professional terminology, verifies and corrects process timing logic anomalies, standardizes numerical precision, physical units, and text format, optimizes sentence structure and grammatical fluency, and ultimately outputs standardized attribution results text with complete equipment numbers, deviation values, stress mechanism descriptions, performance degradation data, and spatiotemporal annotations.

[0114] For each logical link in the semantic connection lexicon, a parameter injection data stream is constructed based on the extracted standardized path element set. The positioning deviation value and the original parameter snapshot data are loaded into the input buffer of the template filling module in the order of the preset physical attribute fields to ensure that the quantitative indicators correspond one-to-one with the logical link nodes.

[0115] The positioning deviation values ​​in the input buffer are converted to units and corrected for accuracy. Based on the measurement accuracy level of the process testing equipment and the tolerance settings in the standard operating procedure, the physical deviation is quantized to a preset decimal place, forming a set of numerical parameters that can be directly used for natural language slot assignment.

[0116] Feature filtering and value range pruning are performed on the original parameter snapshot data to remove noisy data segments and extract core indicators that reflect the process status, such as instantaneous force value, temperature peak value or current change amplitude, and sort them by time series to form a dynamic process parameter stream.

[0117] The pre-trained process semantic generator template filling module is invoked. Based on the causal direction and physical action mechanism of the logical link, the numerical parameter set and dynamic process parameter flow are synchronously injected into the corresponding natural language sentence slots to ensure that the positioning deviation in the causal description is consistent with the time and physical relationship of the process state change.

[0118] During the template filling process, the values ​​and phrases in each slot are checked for consistency. The system checks whether they match the preset units, dimensions, and terminology specifications of the template. If they do not match, the system performs automatic formatting and unit conversion to generate a semi-structured semantic description stream with complete data support.

[0119] By using this parameter injection and template filling method, the semantic connection lexicon and standardized path element set from the previous step are transformed into preliminary causal narrative fragments including specific quantitative indicators, realizing the transformation of semi-structured descriptions into text forms that can be further spatiotemporally integrated.

[0120] For example, in a logical link including crimping offset → coating microcracks → insertion / extraction force attenuation, the positioning deviation value in the standardized path element set is 0.082, the instantaneous crimping force peak value in the original parameter snapshot is 185.6, and the temperature peak value is 231.4. The positioning deviation value is corrected to 0.08 according to the process standard tolerance of ±0.01, and the instantaneous crimping force and temperature peak value are bound in time sequence. In the template filling module, these values ​​are injected into the natural language sentence slots "Crimping positioning deviation X-axis reaches [m]", "Crimping force peak value reaches [N]", and "Coating temperature peak value reaches [°C]". The initial causal narrative generated through template invocation is as follows: "Due to the X-axis positioning deviation of 0.08 mm and the peak crimping force of 185.6 N during the crimping process of the No. 3 crimping machine, the peak temperature reached 231.4 ℃ during the subsequent tin plating process, inducing micro-cracks in the plating layer, which ultimately led to the attenuation of the force value during the insertion and extraction stages." In this process, the injection of positioning deviation m, force value, and temperature value were all verified by slot unit matching and accuracy consistency to ensure the physical accuracy and data verifiability of the narrative.

[0121] S5.4: Perform spatiotemporal context fusion processing on the semi-structured semantic description stream with data support, and use the time series alignment method to bind the millisecond-level timestamps and the logical flow of the process into a continuous time adverbial clause to form a complete dynamic process of describing the evolution of the defect from the trigger anchor point through the transmission anchor point to the final state anchor point, and generate a coherent causal narrative text with precise spatiotemporal coordinates.

[0122] S5.5: Based on the coherent causal narrative text with precise spatiotemporal coordinates, the consistency of technical terms and the fluency of temporal logic in the text are checked and corrected using a natural language post-processing optimizer to output a final natural language narrative including specific equipment number, positioning deviation value, stress concentration description and final performance degradation data, generating an attribution result text with spatiotemporal annotations.

[0123] Based on the coherent causal narrative text with precise spatiotemporal coordinates, a natural language post-processing optimizer is connected, loading a set of professional terminology consistency rules and a set of temporal logic verification rules for the connector manufacturing field. Terminology consistency matching is performed on the narrative text, retrieving and comparing equipment type names, process step names, and defect phenomenon descriptions in the text with the standard terminology list in the rule set. Terms with synonyms or abbreviation differences are replaced with unified standard terms to solidify the uniqueness of terminology expression. Temporal logic consistency analysis is performed on the terminology-standardized text to analyze the sequential relationships between various time adverbial clauses. Clauses with sequence errors or abnormal delays are rearranged or corrected, rebinding time annotations with causal relationship descriptions to restore logical coherence. The positions of quantitative data in the text are calibrated for consistency, ensuring that equipment numbers, positioning deviation values, stress concentration descriptions, and performance degradation data appear in consistent formats across multiple sentences. A number formatter is used to retain three decimal places for positioning deviation values ​​and to bind and normalize the units of performance degradation values. The text processed above is optimized for sentence fluency by invoking a rule-based grammar reconstruction engine. Sections with grammatical ambiguity or verbose sentences are rewritten into concise expressions suitable for engineering scenarios, while maintaining semantic integrity consistent with cause-effect relationships. Through terminology standardization, temporal logic verification, and data format consistency, the coherent causal narrative text from the previous step is transformed into a standardized natural language narrative that includes specific equipment numbers, positioning deviation values, stress concentration descriptions, and final performance degradation data. This achieves engineering verifiability and domain-specific standardization in the attribution result text.

[0124] Step S6: Synchronously highlight the attribution result text and the original temporal semantic anchor point set, establish corresponding jump links, and generate a multimodal visualization interface that supports interactive semantic verification by engineers. Specifically, this includes: S6.1: Perform fine-grained semantic unit segmentation on the attribution result text, and use natural language processing technology to identify equipment numbers, process parameter values ​​and physical phenomenon description keywords in the text, and generate a set of semantic anchor tags with unique location indexes, providing a structured input basis for subsequent data association.

[0125] S6.2: Based on the device identifier and timestamp information in the semantic anchor tag set, perform a multi-source heterogeneous data retrieval operation, extract the corresponding millisecond-level process event snapshot from the original time-series semantic anchor database, retrieve the defect hot zone coordinate matrix from the image storage area, load the electrical performance abnormal interval curve from the waveform database, and generate an original data source object set including spatiotemporal context.

[0126] Based on the device identifier and timestamp information in the semantic anchor tag set, a query condition set for multi-source heterogeneous data retrieval is constructed to ensure that both the device number field and the millisecond-level timestamp field are used as core matching parameters, and index constraints are performed according to a preset time tolerance range to ensure the consistency of cross-system data retrieval.

[0127] During the retrieval and execution phase, the original time-series semantic anchor database interface is called to locate the corresponding process event node based on the query condition set, and the original sensor data snapshot at the time of occurrence is extracted. The fields cover high-precision parameters such as current, voltage, temperature, and pressure, among which the timestamp and status label must be completely preserved.

[0128] For defect image information, the image storage area is accessed, and the defect hot zone coordinate matrix of the corresponding batch of workpieces is selected by combining the device identifier and timestamp. The matrix includes XY pixel coordinates, region bounding box parameters, and feature values ​​bound to process events to ensure synchronization of spatial positioning and temporal semantics.

[0129] For electrical performance test data, connect to the waveform database, retrieve abnormal interval curves according to workpiece identification and time information, load waveform objects including sampling frequency, amplitude sequence and phase offset, and perform time window truncation on the waveform to locate abnormal points.

[0130] The three types of data mentioned above are encapsulated in a spatiotemporal context to create a set of original data source objects, including process event snapshot objects, defect hot zone matrix objects, and electrical performance waveform objects. The associated index under a unified time base is recorded within the set of objects to support subsequent multimodal mapping and highlighting.

[0131] By retrieving and encapsulating multi-source heterogeneous data, the semantic anchor tags from the previous step are transformed into a set of original data source objects with complete spatiotemporal semantics and original evidence chains, thus enabling the preparation of multimodal data that supports visual tracing.

[0132] For example, in a batch of connector terminal production inspection scenarios, the device identifier in the semantic anchor tag set is "CRIMP-M03", with a timestamp of 15:42:17.384. The search criteria are set with a time tolerance range of ±5ms to match cross-device data. The temporal semantic anchor database returns a snapshot of the process event recorded by the device at that time point, showing a crimping force of 248.6N, a crimping position deviation of 0.07mm, and a temperature of 36.4℃. The image storage area returns a defect hotspot coordinate matrix with a resolution of 1920×1080. The hotspot is located at pixel coordinates (832, 614)-(876, 652), with a bounding box width of 44 pixels and a height of 38 pixels. The feature value is a local reflective intensity peak of 112.3. The waveform database returns an abnormal range where the current amplitude sequence, with a sampling frequency of 20kHz, shows a drop in amplitude to 1.2A between sampling points 1845 and 1890. The above three types of data are encapsulated into a set of original data source objects and the time base is unified to 15:42:17.384, so that process event snapshots, defect hot zone matrices and electrical performance waveforms can be highlighted synchronously in the subsequent visualization interface, which significantly improves the efficiency of engineers in locating the cause and the completeness of evidence in the verification of attribution results.

[0133] S6.3: Using a two-way mapping binding method, perform logical association calculations on the semantic anchor tag set and the original data source object set, establish pointer link relationships between semantic fragment indexes and specific data record addresses, generate a semantic data mapping graph with bidirectional jump capabilities, and realize direct tracing from textual narrative to original evidence.

[0134] The system receives a semantic anchor tag set output from S6.2 and a set of original data source objects as input. The tag set includes a unique location index, device identifier, and timestamp fields. The original data source object set includes millisecond-level process event snapshots, defect hotspot coordinate matrices, and abnormal electrical performance waveform curves. Based on the device identifier field, the system calls the device identifier matching engine to locate process event snapshots in the original data source object set that match the device identifier in the tag set and generates a preliminary matching pair list. Time alignment calculations are performed on the timestamp fields in the preliminary matching pair list. A bidirectional mapping binding method is used to establish bidirectional association pointers between the location index in the tag set and the data record addresses in the original data source object set. A bidirectional traversal rule ensures that queries in either direction return the corresponding data record within the minimum time deviation. When the deviation ratio meets a threshold condition, the matching pair is written to the semantic data mapping graph node connection table. A consistency structure check is performed on the node connection table, removing invalid links with inconsistent device identifiers or out-of-bounds timestamps to solidify the traceable mapping relationship. Through the above bidirectional mapping binding and consistency verification process, the tag set and the original data source object set from the previous step are transformed into a semantic data mapping graph with bidirectional jump capability, realizing direct tracing from textual narrative to original evidence.

[0135] S6.4: Based on the semantic data mapping map, perform multimodal synchronous highlight rendering operation, and display the selected semantic fragment, the corresponding process parameter snapshot, defect hot zone image and electrical performance waveform curve in a time-space alignment on the visualization canvas, generating a multimodal linkage view that intuitively reflects the causal chain, and enhancing the engineer's visual perception of attribution logic.

[0136] S6.5: Integrate interactive feedback controls into the multimodal linked view, configure three types of instruction listening modules for acceptance, questioning, and supplementary basis, capture the engineer's operation behavior on the multimodal linked view and encapsulate it into a standardized verification feedback instruction package, generate an interactive semantic verification interface that supports human-machine collaborative optimization, and complete the closed-loop verification of the credibility of the attribution results.

[0137] The system receives a cached rendering frame from the multimodal interactive view as the initial interface object. Based on the interaction-level metadata of the cached frame, it calls the control injection engine to write three types of instruction key-value pairs—accept, challenge, and supplement—into the interface control registry. A unique listening module ID and trigger condition combination index are bound to each type of instruction control. The trigger condition combination index is mapped to the spatial positioning matrix of the interface elements to generate a control position index table that can be used for event capture and parsing. The system periodically reads the state change flags of the interface controls through an event polling service. For event instances where the state change flag is true and matches the listening module ID, interaction log serialization is performed, encapsulating the event trigger timestamp, operator ID, interface element position index, and original instruction type into a structured event record. An instruction semantic parser is introduced into the structured event record set, converting it into a standardized verification feedback instruction package according to a predefined instruction type code table, and attaching a snapshot of the operating environment parameters for subsequent optimization engine calls. Through control injection, event capture, semantic parsing, and instruction package generation, the multimodal interactive view from the previous step is transformed into a verification instruction stream with real-time human-computer interaction capabilities, achieving closed-loop verification of the attribution result's credibility.

[0138] For example, for a batch of engineering verification data including 120 attribution result text multimodal views, the "Accept" button in the interactive feedback control is configured with a listening module ID of 101, and the trigger condition is set to the "Click" event with an operation duration exceeding 0.3 seconds; the "Question" button is configured with a listening module ID of 102, and the trigger condition is the "DoubleClick" event; the "Supplementary Evidence" control is configured with a listening module ID of 103, and the trigger condition is the "RightClick" event. The position matrix is ​​generated by the interface rendering engine, and the resolution of each control position index is 1 pixel. The event polling service frequency is set to 60 times per second, and the average size of the captured raw interactive events after serialization is 256 bytes. The semantic parser maps 101 to the Accept class, 102 to the Question class, and 103 to the Supplementary Evidence class according to the instruction type code table. The environmental parameters attached to each event include the current defect type number, device number, and graph node index. In 10 iterations, the system can stably generate verification feedback instruction packages that conform to a standardized format and successfully use them in the optimization engine to adjust the attribution model. The verification results show that the adjusted model has significantly improved the consistency of the causal chain structure in multiple backtracking tests, and the interface interaction response latency remains within 20 milliseconds.

[0139] Step S7: Receive feedback instructions submitted by engineers on the multimodal visualization interface, including acceptance, questioning, or supplementary information. Dynamically adjust the definition rules of timing semantic anchors and the parameter configuration of the semantic generation template library in the process semantic generator according to the type of feedback instruction, generating an optimized anchor definition rule set and semantic generation template library. Specifically, this includes: S7.1: Parse and process the feedback instructions submitted by engineers on the multimodal visualization interface, including acceptance, questioning, or supplementary feedback instructions, to extract a structured feedback data stream that includes feedback type identifiers, related semantic fragment indexes, and correction suggestion text.

[0140] S7.2: Based on the questioning feedback instructions in the structured feedback data stream, the existing temporal semantic anchor definition rule set is traversed and matched using the rule conflict detection method to identify the abnormal anchor trigger threshold or erroneous process logic constraint that leads to attribution bias.

[0141] Based on the questioning feedback instructions in the structured feedback data stream, the system receives input objects including a questioning identifier, an index of associated semantic fragments, and suggested correction text. It establishes a mapping retrieval condition between the questioning identifier and the existing set of temporal semantic anchor definition rules. For each anchor trigger condition in the rule set, a traversal matching operation is performed. The field comparison module in the rule conflict detection method is invoked to locate the corresponding rule entry based on the semantic fragment index associated with the questioning identifier, extracting the rule triggering threshold and process logic constraint parameters. Numerical difference calculations are performed on the triggering threshold parameters, and the difference scale is calculated using the following formula. in, This refers to the difference between the parameters suggested for adjustment in the feedback and their original values ​​in the rule set. The original values ​​are set in the rule set. Topological consistency verification is performed on the process logic constraint parameters. The sequence of questioned processes is compared with the established standard operating procedure sequence for node consistency. The process graph topology traversal module is called to mark any non-adjacent process connections or abnormal edges lacking necessary transmission links. Anomaly markers are added to all rule entries whose trigger threshold differences exceed the preset tolerance or whose process logic topology verification fails, and an abnormal rule list is output. Anomaly anchor point trigger thresholds and erroneous process logic constraints are identified using a rule conflict detection method. The detection results are used as direct input for the parameter gradient update of the next sub-step, achieving quantitative location and structured output of rule deviations.

[0142] S7.3: Based on the abnormal anchor point trigger threshold and the erroneous process logic constraint, and combined with the correction suggestion text in the supplementary feedback instruction, perform a parameter gradient update operation to obtain a candidate anchor point definition rule set including the corrected time accuracy tolerance range and the description of the new physical action direction.

[0143] Based on the identification results of abnormal anchor point trigger threshold and erroneous process logic constraints, the correction suggestion text in the supplementary feedback instruction is loaded to form a parameter update input set including time accuracy tolerance adjustment suggestions and physical action direction revision suggestions.

[0144] For the updated results of time tolerance range and process directionality labels, perform consistency verification, compare the reachability of time constraints and physical action directions on the process logic chain, eliminate combinations that do not conform to standard operating procedures, and retain candidate anchor point logic that conforms to the physical rules of the production site.

[0145] By correcting the time precision tolerance range and updating the physical action direction description, the previous abnormal rule detection results are transformed into a set of candidate anchor point definition rules with the latest physical meaning and time constraint definition, thereby realizing the dynamic repair and optimization of attribution logic.

[0146] S7.4: Using the candidate anchor point definition rule set as training corpus constraints, fine-tune the template mapping weights in the pre-trained process semantic generator to obtain an optimized semantic generation template library that adapts to the new attribution logic and has high natural language fluency.

[0147] For the input conditions of the rule set defined for candidate anchor points, the full template mapping weight matrix of the pre-trained process semantic generator is loaded as the initial calculation object. The weight fine-tuning calculation engine is called to establish the corresponding mapping relationship between the rule set parameters and the template slot placeholders, so as to ensure that the physical action direction and time precision tolerance in the rule set can be accurately reflected in the template filling process.

[0148] The various condition thresholds, device number matching patterns, and causal connection phrases included in the candidate anchor point definition rule set are parsed into two types of structured features: numerical constraints and semantic constraints. These are then input into the numerical weight channel and semantic weight channel of the weight optimization module, respectively, so that the template mapping weights can be adjusted simultaneously in the backpropagation iteration.

[0149] The update results of the numerical weight channel and the semantic weight channel are matrix-weighted and merged in the weight synthesis module to form an intermediate optimized weight matrix that includes multi-dimensional constraint information. The matrix is ​​then used to automatically score the natural language fluency on the validation set samples, and template weight update schemes with fluency below a preset threshold are removed.

[0150] The scoring results are iteratively optimized again through a dynamic weight redistribution mechanism to ensure that the generator maintains consistency of technical terminology and sentence coherence while adapting to the new attribution logic.

[0151] By linking matrix updates with fluency scoring, the candidate anchor point definition rule set from the previous step is transformed into an optimized semantic generation template library that includes multi-dimensional constraints, high semantic matching, and fluent natural language expression, thereby significantly improving the physical usability and engineering credibility of the template generation results.

[0152] S7.5: Based on the optimized semantic generation template library and candidate anchor point definition rule set, perform version merging and consistency verification to output the final anchor point definition rule set and final semantic generation template library with self-evolution capability, thereby completing the reconstruction of the engineering credibility foundation of the attribution logic.

[0153] Step S8: Based on the optimized anchor point definition rule set and semantic generation template library, update the node attributes and edge type descriptions of the lightweight temporal semantic graph to complete the reconstruction of the engineering credibility foundation of the attribution logic. Specifically, this includes: S8.1: Obtain the optimized anchor point definition rule set and semantic generation template library as input data, and use the rule parsing engine to perform syntax tree decomposition on the condition judgment logic in the rule set and the natural language sentence structure in the template library to extract a set of structured update instructions, including device number matching patterns, parameter threshold ranges and causal connection phrases.

[0154] S8.2: Based on the structured update instruction set, the graph traversal method is used to locate the time-series semantic anchor nodes to be updated and the process influence directionality description edges to be updated in the lightweight time-series semantic graph that have a mapping relationship with the instruction set, so as to obtain the graph subgraph object to be reconstructed, including the original node attribute snapshot and the original edge type description.

[0155] S8.3: For the time-series semantic anchor nodes to be updated in the subgraph object to be reconstructed, perform an attribute injection operation, write the device number matching pattern and parameter threshold range extracted from the structured update instruction set into the status label field and the original parameter snapshot field of the node, and obtain the updated time-series semantic anchor node set carrying the latest physical meaning definition.

[0156] S8.4: Based on the updated set of temporal semantic anchor nodes, perform semantic recalibration on the process influence directional description edges to be updated in the subgraph object to be reconstructed. Replace the original positive reinforcement, negative inhibition and threshold triggering attribute definitions with causal connection phrases in the template library to obtain the updated set of process influence directional description edges that reflect the latest process logic relationship.

[0157] S8.5: Integrate the updated set of temporal semantic anchor nodes with the updated set of process influence directional description edges, perform graph topology consistency verification and version merging operations, and persistently store the updated node attributes and edge type descriptions to the lightweight temporal semantic graph database.

[0158] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0159] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0160] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for tracing and analyzing defects in connector terminals, specifically including: S1: Extract key process events with clear physical meaning and millisecond-level time accuracy from the programmable logic controller logs of connector equipment, automatic optical inspection timestamps, and manufacturing execution system work order records, and generate a set of time sequence semantic anchor points; S2: Based on the timestamp information in the set of temporal semantic anchor points, multi-source time alignment technology is used to accurately associate the coordinates of the thermal defects, the three-dimensional point cloud deformation area, and the electrical performance abnormality interval of the same workpiece in images collected at different stages with the nearest neighbor of several temporal semantic anchor points to generate a preliminary screening causal chain. S3: Using the temporal semantic anchor points in the initial screening causal chain as nodes and the physical action direction and process logic flow direction as edge types, construct a directed heterogeneous graph to generate a lightweight temporal semantic graph. S4: For the new input defect sample, perform a restricted depth-first traversal along the lightweight temporal semantic graph, retain only the path sequence that simultaneously satisfies the triple checks of temporal sequence constraints, physical operability, and process reachability, and generate a candidate attribution path set that conforms to the manufacturing process ontology logic; S5: Call the pre-trained process semantic generator to process each path sequence in the candidate attribution path set, transform the abstract process relationship into a natural language description, generate attribution result text with spatiotemporal annotations, and complete the source tracing analysis of connector terminal defects.

2. The method for tracing and analyzing defects in connector terminals according to claim 1, characterized in that, The step S5 is followed by: S6: Synchronously highlight the attribution result text and the original temporal semantic anchor point set, establish corresponding jump links, and generate a multimodal visualization interface that supports interactive semantic verification by engineers. S7: Receive feedback instructions submitted by engineers on the multimodal visualization interface, including acceptance, questioning, or supplementary information. Based on the type of feedback instruction, dynamically adjust the definition rules of the time-series semantic anchor points and the parameter configuration of the semantic generation template library in the process semantic generator, and generate an optimized anchor point definition rule set and semantic generation template library. S8: Based on the optimized anchor point definition rule set and semantic generation template library, update the node attributes and edge type descriptions of the lightweight temporal semantic graph to complete the reconstruction of the engineering credibility foundation of the attribution logic.

3. The method for tracing and analyzing defects in connector terminals according to claim 1, characterized in that, The set of temporal semantic anchors includes a unique identifier, the process to which it belongs, the equipment identifier, the timestamp, the status label, and the original parameter snapshot.

4. The method for tracing and analyzing defects in connector terminals according to claim 1, characterized in that, The lightweight temporal semantic graph is specifically a lightweight temporal semantic graph that does not include probability weights but includes positive reinforcement, negative inhibition, and threshold-triggered attribute definitions.

5. The method for tracing and analyzing defects in connector terminals according to claim 1, characterized in that, The natural language description includes the specific device number, positioning deviation value, stress concentration description, and final performance degradation data.

6. The method for tracing and analyzing defects in connector terminals according to claim 3, characterized in that, Step S3 specifically includes: Obtain the set of temporal semantic anchor points in the initial causal chain, and instantiate the temporal semantic anchor points based on the unique identifier and the process attribute to which they belong, to generate a graph node object including equipment identifier, timestamp and original parameter snapshot attributes; Based on the temporal order and process logic flow of graph node objects, the causal correlation strength between adjacent temporal semantic anchors is calculated using the physical action mechanism analysis method, generating a set of candidate directed edges that characterize heat conduction, stress transfer, or ion migration relationships. Generate directed edges with clearly defined semantic descriptions of the directional nature of process influence from the candidate set of directed edges; By using graph node objects and directed edges carrying directional descriptions of process influence, a directed heterogeneous graph assembly operation is performed to construct a lightweight temporal semantic graph that does not include probability weights but fully preserves the directional information of physical effects. A topological consistency check method based on a lightweight temporal semantic graph is applied to detect and correct redundant edges and abnormal nodes that violate temporal order constraints or physical operability principles, generating an optimized final lightweight temporal semantic graph.

7. The method for tracing and analyzing defects in connector terminals according to claim 6, characterized in that, The process of generating directed edges with clearly defined semantic descriptions of process influence directional data for the candidate directed edge set includes performing directional labeling processing on each directed edge in the candidate directed edge set according to a predefined process influence rule base, mapping abstract physical associations to positive reinforcement, negative inhibition, and threshold-triggered attribute labels, and generating directed edges with clearly defined semantic descriptions of process influence directional data.

8. The method for tracing and analyzing defects in connector terminals according to claim 1, characterized in that, The extraction of key process events with clear physical meaning and millisecond-level time accuracy includes using an interpolation-based time drift compensation method for nonlinear drift equalization to ensure millisecond-level accuracy when aligning key process events.

9. The method for tracing and analyzing defects in connector terminals according to claim 1, characterized in that, The process of transforming abstract technological relationships into natural language descriptions and generating attribution result text with spatiotemporal annotations, including the transformation and generation process, has undergone professional terminology consistency standardization and temporal logic verification, and all quantitative indicators have been uniformly formatted.

10. The method for tracing and analyzing defects in connector terminals according to claim 2, characterized in that, The dynamic adjustment process semantic generator defines the rules for temporal semantic anchors and configures the parameters of the semantic generation template library, including automatically merging and verifying consistency with existing temporal semantic anchor definition rules and semantic generation template libraries, eliminating similarities and retaining differences.

Images