Pipeline hot melt welding joint construction process data recording and tracing method and system

Abstract

The present application relates to a pipeline hot melt welding joint construction process data recording and tracing method and system, first through obtaining and synchronously heating temperature, pressure application time, environmental humidity and melting zone infrared image and other multi-source sensing data, preprocessing and feature extraction are carried out, and welding process parameter sequence is constructed. Subsequently, combined with the pre-training knowledge graph model, the process parameters and quality entity relationship are mapped, the key causal path is calculated with the help of the causal rule base and dynamic graph, and the abnormal attribution results and key data are stored by using the blockchain. Finally, the blockchain storage decoding, data cross-modal alignment analysis, dominant abnormal factor automatic positioning and construction specification matching report generation are realized. The method effectively improves the data quality of welding process parameters and the accuracy of abnormal tracing, realizes the solidification of the whole process responsibility chain and the improvement of the legal acceptance force.

Description

Technical Field

[0001] This invention relates to the field of pipeline hot-melt welding data traceability and responsibility attribution technology, and in particular to a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints. Background Technology

[0002] Pipeline thermal fusion welding is a critical step in engineering construction, and the quality of the welded joints directly affects the safety and stable operation of the pipeline. Currently, mainstream welding quality recording and traceability systems in the industry collect welding parameters (such as heating temperature, pressure, time, and ambient humidity), equipment status data, and quality inspection results on-site. Through storage, comparison, and anomaly alarms, they enable post-event accountability and quality review. Some systems further incorporate blockchain technology, uploading welding parameter summaries or raw data to the blockchain to enhance data tamper-proof capabilities and legal validity. The industry's development trend is continuously evolving towards intelligence, automation, data reliability, and traceability and explainability.

[0003] In existing technologies, most pipeline hot-melt welding data traceability systems employ static data acquisition and comparison methods. For example, they use local or cloud databases to record parameter curves for each welding stage and detect anomalies using simple threshold rules. However, this only identifies non-compliance in a specific parameter and cannot provide in-depth analysis of the underlying causal relationships between multiple factors such as the process flow, equipment status, and environmental influences. Systems with blockchain data solidification capabilities typically store raw welding data, parameter summaries, or hash values ​​on the chain. While this ensures data immutability, on-chain evidence is generally "weakly semantic" recording. The lack of deep logical modeling between process technology and final quality indicators makes it difficult for the system to establish a clear causal path between abnormal parameters and weld quality defects when welding anomalies occur. This limits the interpretability and scientific support for subsequent scenarios such as responsibility allocation, identification of responsible parties, problem tracing, and risk recurrence warnings.

[0004] Common technological applications in the field of intelligent monitoring and accountability for pipeline hot-melt welding include: integrating multimodal sensors into welding robots to achieve automatic parameter acquisition throughout the entire process; using edge computing to achieve real-time trend analysis and automatic early warning; and storing parameter summaries via blockchain to support compliant traceability. These technologies lay the foundation for improving data authenticity and traceability, but their causal inference level mostly relies on empirical rules and manual interpretation, resulting in coarse-grained analysis, and data storage faces bandwidth and storage pressures.

[0005] Existing technologies have the following key shortcomings: First, in scenarios involving multiple factors related to welding process anomalies, conventional monitoring and traceability systems struggle to automatically identify causal relationships between abnormal parameters and the logical sources of quality defects, lacking an "explanatory path" for problem localization and failing to provide effective technical support for attribution of responsibility. Second, storing evidence on-chain using hashing, digests, or other methods results in the loss of detailed semantics of the process, limiting its reliable reference value for legal proof, dispute resolution, and quality risk reproduction. Third, limited by bandwidth and storage costs, existing blockchain welding systems cannot store all raw data on-chain, and using pure parameter digests for solidification prevents in-depth attribution and statistical analysis of dynamic, multi-granular, and cross-period process anomalies.

[0006] Pipeline welding construction companies urgently need an innovative technological solution that can dynamically infer causal relationships between heterogeneous data such as multi-source sensor parameters, operational behaviors, equipment status, and environmental information; automatically generate anomaly attribution paths; and solidify on-chain evidence in an efficient and low-cost manner to improve the explainability and legal validity of accountability. Especially when abnormal parameters occur during welding, the system needs to automatically identify the main influencing factors leading to the final welding defects, clarify the responsibility attribution and reproducibility risks at each stage, and provide more scientific and systematic technical support for project quality audits, legal evidence collection, risk management, and personnel training.

[0007] In summary, existing technologies have not yet achieved a welding anomaly attribution system centered on dynamic causal graphs, oriented towards deep process-quality modeling, and possessing lightweight on-chain evidence storage and highly interpretable traceability capabilities. This makes it difficult to meet the comprehensive management needs of smart construction sites, digital twins, and engineering legal affairs in the new era. Therefore, proposing new technical solutions to address the problem of how to integrate multi-source data dynamic causal analysis with low-cost blockchain evidence storage to achieve scientific attribution of welding process anomaly data and automatic generation of traceability reports has significant application value and practical urgency. Summary of the Invention

[0008] This application provides a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0009] This application provides a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints, specifically including: S1: Acquire multi-source heterogeneous raw data streams from a sensor set during the pipeline hot-melt welding process, and preprocess the multi-source heterogeneous raw data streams to generate a welding process parameter sequence; S2: Call the pre-trained knowledge graph of the pipeline hot melt welding domain to initialize the model and map the welding process parameter sequence into a set of process-quality related entities; S3: Using the typical causal rule library extracted from the historical qualified and abnormal weld sample dataset, the connection relationship between each entity in the process-quality related entity set is dynamically activated to calculate the intensity of the connection relationship, and a dynamic causal graph model representing the evolution trajectory of the confidence of each causal edge under the current welding state is generated. S4: Based on the dynamic causal graph model, select the highest weight causal edge with a confidence level higher than a preset threshold, extract the directed subgraph encoding, and generate an abnormal attribution path snapshot; S5: Perform a hash operation on the abnormal attribution path snapshot to generate a data digest, and package the data digest together with the spatiotemporal identity into a minimal transaction unit to generate a blockchain evidence record; S6: Read the blockchain evidence record of the specified weld joint and decode and restore the causal path structure, and pull the original parameter fragments corresponding to the nodes involved in the causal path structure to generate a bidirectional graph reconstruction dataset; S7: Based on the bidirectional map reconstruction dataset, calculate the deviation between the path semantics and the measured data distribution, and generate the localization result of the dominant anomaly factor.

[0010] S8: Based on the location result of the dominant anomaly factor, map it to the construction specification clause number and associate and display the clustering results of the frequency of occurrence and recurrence conditions of similar factors with the same or similar process characteristics as the dominant anomaly factor within a preset time period, and generate a welding anomaly attribution and tracing report.

[0011] The method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints provided in this application have the following beneficial effects: (1) The welding quality traceability method based on a lightweight graph inference engine provided in this application addresses the common problem of "heavy data storage and light causal analysis" in existing industrial blockchain traceability systems. It proposes a dynamic attribution mechanism that embeds process knowledge into edge computing nodes, effectively overcoming the shortcomings of traditional solutions such as redundant traceability information, weak off-chain analysis capabilities, and reliance on human experience for anomaly attribution caused by relying solely on hashing raw sensor data on the chain. By deploying a lightweight graph inference module at the welding equipment end, combined with a pre-trained knowledge graph in the field of pipeline hot melt welding, the system can capture the causal relationship evolution process between key process parameters in real time during each welding process, rather than simply recording numerical snapshots; a local graph update is triggered every 500 milliseconds, and the contribution weight of each causal path is dynamically evaluated based on the activation intensity and confidence decay model, ultimately generating an "anomaly attribution path snapshot" composed of 3 to 5 high-weight causal edges. This snapshot encodes the logical propagation path from the initial deviation to the final quality result in the form of a directed subgraph. It contains a complete node sequence and edge confidence interval identifiers, and the structure is valid after logical consistency verification. This achieves a structured compressed expression of the welding anomaly propagation path, which significantly improves the semantic density and interpretability of on-chain information.

[0012] (2) Furthermore, this solution innovatively adopts a hybrid architecture of "causal path summary + on-demand retrieval of off-chain original fragments", which significantly reduces system resource overhead while ensuring the effectiveness of legal tracing. Unlike the common practice in existing technologies of putting all data on the chain or using complex privacy protection mechanisms (such as zero-knowledge proofs and proxy re-encryption), this invention only submits the path summary after SM3 hashing as the minimum transaction unit to the lightweight consensus layer to complete the on-chain storage, avoiding the bandwidth pressure and storage expansion risk brought by high-frequency sensor streams; when responding to tracing requests in the cloud, a bidirectional graph reconstruction process is initiated: on the one hand, the on-chain summary is decoded to restore the causal path structure, and on the other hand, the original parameter fragments in the time neighborhood of the nodes related to the path are accurately pulled from the distributed database, and semantic alignment analysis is performed using graph neural networks to automatically compare the degree of deviation between the measured data distribution and the expected model, thereby identifying the key links that dominate the anomaly. This mechanism not only realizes a paradigm shift from "data tracing" to "causal logic tracing," but also improves the accuracy and response efficiency of anomaly location through a semantically driven selective data retrieval strategy. This makes quality responsibility determination no longer rely on subjective expert judgment, but is based on verifiable and reproducible dynamic knowledge reasoning.

[0013] The aforementioned technologies collectively construct an efficient, intelligent, and legally credible closed-loop traceability system for welding quality. This system, without introducing multi-level consortium blockchain governance, IPFS mapping, or complex cryptographic tools, achieves rapid attribution of process anomalies, low-overhead evidence storage, and highly reliable verification, significantly reducing the system deployment threshold and long-term operation and maintenance costs. It is particularly suitable for special equipment manufacturing scenarios with high security requirements and strong regulatory compliance, such as pressure pipelines and nuclear power plant welds. It can support the automated mapping of construction specification clauses and the clustering analysis of historical risk patterns, providing objective, structured, and auditable technical evidence for defining engineering quality responsibility, and has good engineering applicability and promising prospects for widespread adoption. Attached Figure Description

[0014] Figure 1 This is the main flowchart of a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints; Figure 2 This is a sub-flowchart of a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints; Figure 3 This is another sub-flowchart of a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints. Detailed Implementation

[0015] 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.

[0016] 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.

[0017] like Figure 1 As shown, this application provides a method and system for recording and tracing data during the construction process of pipeline hot-melt welded joints, specifically including: S1: Acquire multi-source heterogeneous raw data streams from a sensor set during the pipeline hot-melt welding process, and preprocess the multi-source heterogeneous raw data streams to generate a welding process parameter sequence with a unified timing reference.

[0018] The sensor set includes a multi-source heterogeneous raw data stream comprising a heating temperature sensor, a pressure application duration timer, an ambient humidity detector, and an infrared imaging device for the molten zone. Preprocessing of the multi-source heterogeneous raw data stream specifically involves timestamp alignment and noise filtering.

[0019] S2: Based on the welding process parameter sequence, call the pre-trained knowledge graph of pipeline hot melt welding domain to initialize the model, and map the welding process parameter sequence into a set of process-quality related entities.

[0020] The process-quality related entity set includes: heating temperature node, pressure application duration node, ambient humidity node, and infrared image texture entropy value node of the molten zone.

[0021] S3: Using the typical causal rule library extracted from the historical qualified and abnormal weld sample dataset, the connection relationship between each entity in the process-quality related entity set is dynamically activated to calculate the intensity of the connection relationship, and a dynamic causal graph model representing the evolution trajectory of the confidence of each causal edge under the current welding state is generated.

[0022] S4: Based on the dynamic causal graph model, the highest-weighted causal edges with confidence levels higher than a preset threshold are selected, and the directed subgraph encoding is extracted to generate a snapshot of the abnormal attribution path for this welding process. The directed subgraph encoding consists of a unique identifier for the starting node, a unique identifier for the ending node, a sequence of intermediate jump nodes, and confidence interval identifiers for each edge.

[0023] S5: Perform a hash operation on the snapshot of the abnormal attribution path to generate a data digest, and package the data digest together with the spatiotemporal identity into a minimal transaction unit to generate a blockchain evidence record.

[0024] Specifically, the SM3 hash algorithm (a national cryptographic standard) is applied to the snapshot of the anomaly attribution path to generate a fixed-length data digest. This data digest, along with a spatiotemporal identity identifier, is then packaged into a minimal transaction unit to complete a lightweight on-chain existence operation and generate a blockchain existence record. The spatiotemporal identity identifier includes: a timestamp, a unique identification code for the welding equipment, and a biometric binding code for the operator.

[0025] S6: Read the blockchain evidence record of the specified weld joint from the blockchain and decode and restore the causal path structure. At the same time, pull the original parameter fragments within the upstream and downstream time windows corresponding to the nodes involved in the causal path structure from the off-chain distributed database to generate a bidirectional graph reconstruction dataset.

[0026] S7: Based on the bidirectional graph reconstruction dataset, calculate the deviation between the path semantics and the distribution of measured data, identify the link where the measured value deviates most significantly from the expected model, and generate the location result of the dominant anomaly factor.

[0027] S8: Based on the location result of the dominant anomaly factor, map it to the construction specification clause number and associate it with the clustering results of the frequency of occurrence and recurrence conditions of similar factors with the same or similar process characteristics as the dominant anomaly factor within a preset time period, and generate a welding anomaly attribution and tracing report with legal credibility.

[0028] Step S1: Acquire multi-source heterogeneous raw data streams from a sensor set during the pipeline hot-melt welding process, and preprocess the multi-source heterogeneous raw data streams to generate a welding process parameter sequence with a unified timing reference. The sensor set includes: a heating temperature sensor, a pressure application duration timer, an ambient humidity detector, and a molten zone infrared imaging device, providing multi-source heterogeneous raw data streams. The preprocessing of the multi-source heterogeneous raw data streams specifically involves timestamp alignment and noise filtering. This includes: S1.1: Acquire multi-source heterogeneous raw data streams from heating temperature sensor, pressure application duration timer, ambient humidity detector and molten zone infrared imaging device, and perform protocol parsing and format standardization processing on the multi-source heterogeneous raw data streams to generate a standardized raw signal set with a unified data interface.

[0029] The system captures input signals from multi-source, heterogeneous raw data streams from heating temperature sensors, pressure application duration timers, ambient humidity detectors, and infrared imaging devices in the molten zone. It establishes independent data channels for each physical quantity and locks their sampling rate, field definitions, and byte order structure. For each data channel, the system calls a protocol parsing module to perform header stripping, payload extraction, and control character checks based on the communication protocol used by each type of sensor (including but not limited to Modbus, CAN, RS485, or TCP / IP) to remove non-data components and verify the integrity and validity of the data frames. The parsed data payload undergoes format mapping processing. Combined with a unified data interface specification, raw signals from different sources, with different precisions, and of different data types are converted into a unified numerical encoding system, including time fields, measured value fields, and status flag fields, ensuring consistent representation of data from different sensors at the interface call layer. Standardization of data types and units is applied to all measured value fields. Temperature values ​​are uniformly converted to degrees Celsius, pressure durations are uniformly represented as integer seconds, humidity values ​​are uniformly represented as percentage floating-point numbers, and infrared entropy values ​​are uniformly represented as normalized floating-point numbers. This eliminates the impact of data type differences on subsequent processing. Structured encapsulation is performed on each standardized data record to generate a standardized raw signal set, including raw measured values, timestamps, status flags, and interface call identifiers. This serves as the input benchmark for subsequent timing alignment and noise filtering. Through the above protocol parsing and format standardization processing, multi-source heterogeneous raw data streams are transformed into a standardized raw signal set with a unified data interface, enabling cross-sensor and cross-protocol data fusion preparation.

[0030] S1.2: Based on the standardized original signal set, the timestamps of each signal channel are corrected for deviation using a global clock synchronization mechanism to generate a time-aligned signal sequence with a unified timing reference.

[0031] S1.3: Perform adaptive wavelet threshold denoising algorithm processing on the time-aligned signal sequence to remove high-frequency random noise interference and generate clean process parameter waveforms with improved signal-to-noise ratio.

[0032] S1.4: Based on the waveform of the clean process parameters, linear interpolation resampling technology is used to complete the missing data points and rate normalization is performed on signals with different sampling frequencies to generate a sequence of equally spaced process parameters with a fixed sampling interval.

[0033] Based on the clean process parameter waveforms obtained through adaptive wavelet threshold denoising, time series indexes are established for each sensor signal channel to identify the locations of missing sampling points under a unified time series reference, thus forming a missing point location index table.

[0034] For each missing point in the missing point location index table, the linear interpolation algorithm is called to construct numerical pairs between two adjacent known sampling points. The interpolated values ​​are calculated sequentially along the time axis. The interpolation formula is expressed as:

[0035] Where V(t) is the interpolation result at time t corresponding to the missing point. and These are the sampling times immediately preceding the missing point. The corresponding known sampled value and the next adjacent sampling time after the missing point The corresponding known sampled values, The timestamp of the sampling point adjacent to the missing point. This is the timestamp of the next adjacent sampling point after the missing point.

[0036] The sampling frequency differences of each signal channel after completion are detected, and the ratio factor between the original sampling interval and the target unified sampling interval is calculated to form a rate normalization factor table.

[0037] Based on the rate normalization factor table, decimation is performed on signals with sampling frequencies higher than the target frequency, and interpolation is used to supplement the number of sampling points for signals with sampling frequencies lower than the target frequency, so that the sampling interval of all signals tends to a fixed value Δt.

[0038] The signal sequences that have been completed and normalized are reordered according to timestamps to form a sequence of equally spaced process parameters with a fixed sampling interval.

[0039] By using linear interpolation and rate normalization, the clean waveform from the previous step is transformed into an equidistant sequence of welding process parameters with no missing points and consistent sampling intervals, thus achieving structural consistency of multi-channel data under a unified time-series benchmark and ensuring the accuracy of subsequent feature extraction.

[0040] S1.5: Perform sliding window feature extraction and outlier truncation on the equidistant process parameter sequence to finally generate a welding process parameter sequence with a unified time series reference and meeting the map mapping requirements.

[0041] Step S2: Based on the welding process parameter sequence, initialize the model using a pre-trained knowledge graph of the pipeline hot-melt welding domain, mapping the welding process parameter sequence to a set of process-quality related entities. This set of process-quality related entities includes: heating temperature nodes, pressure application duration nodes, ambient humidity nodes, and infrared image texture entropy value nodes of the molten zone. Specifically, it includes: S2.1: Based on the welding process parameter sequence with a unified time reference, semantic type recognition is performed on multi-source heterogeneous data using the ontology definition library in the field of pipeline hot melt welding. The heating temperature numerical stream, cumulative pressure application time, environmental humidity percentage sequence, and infrared image texture entropy matrix of the molten zone are extracted to generate an original feature vector group with standard semantic labels.

[0042] Based on a welding process parameter sequence with a unified temporal reference, the multi-source data semantic parsing module in the ontology definition library for pipeline hot-melt welding is invoked to perform semantic classification processing on each signal channel in the sequence. The data stream from the heating temperature sensor is mapped to a temperature numerical domain type, the data stream from the pressure application duration timer is mapped to a pressure duration cumulative domain type, the data stream from the ambient humidity detector is mapped to a humidity percentage domain type, and the data stream from the molten zone infrared imaging device is mapped to a texture entropy matrix domain type. In the temperature numerical domain type data, a unit unification operation is performed according to the physical quantity unit specifications preset in the ontology library, converting all possible Fahrenheit and Celsius isotherms to Celsius standard values, and adding the semantic tag "heating temperature" according to the physical quantity namespace defined in the ontology. In the pressure duration cumulative domain type data, combined with the defined range of construction pressure duration in the ontology library, cumulative integration calculation is performed, merging each pressure application time into a single cumulative value, and adding the semantic tag "pressure application duration". In the humidity percentage domain type data, a percentage normalization algorithm is invoked to standardize the humidity output of different sensors into percentage form, and adding the semantic tag "ambient humidity". In texture entropy matrix domain type data, a preset gray-level distribution statistical model is used to extract texture entropy values ​​from infrared images, and the image texture entropy value at a specified time is calculated using a formula:

[0043] in, This represents the probability of gray level i appearing in the image, where n is the total number of gray levels. The image texture entropy value is used, along with the semantic label "Melted Zone Infrared Image Texture Entropy Value". The above four types of data with standard semantic labels are combined to construct an original feature vector group. A unified interface format is achieved through a multi-source feature vector merging module, ensuring input consistency for subsequent entity alignment and node instantiation processing. Through the above semantic type recognition and feature construction processing method, the unified time-series sequence result from the previous step is transformed into an original feature vector group with clear semantic labels that can be directly mapped to the knowledge graph node space, achieving a precise conversion from physical signals to semantic features.

[0044] S2.2: Based on the original feature vector group with standard semantic labels, call the entity alignment algorithm in the pre-trained knowledge graph initialization model for pipeline hot-melt welding, perform a nonlinear mapping transformation from feature space to graph node space, and generate a candidate entity set containing heating temperature node instances, pressure application duration node instances, ambient humidity node instances, and infrared image texture entropy value node instances of the molten zone.

[0045] The knowledge graph initialization model for the pipeline hot-melt welding field includes: a data normalization transformer, a hidden layer representation generator, a cosine similarity calculator, a Top-K truncation filter, a weight normalization processor, and an entity alignment algorithm engine. Specifically, the data normalization transformer performs dimensional unification on the original process feature vectors; the hidden layer representation generator extracts nonlinear combined features through a multilayer perceptron; the cosine similarity calculator performs similarity matching between the hidden layer features and the knowledge graph node library; the Top-K truncation filter retains the candidate nodes with the highest matching degree; the weight normalization processor normalizes the weights of the candidate nodes; and the entity alignment algorithm engine, as the scheduling core, drives the above components to work collaboratively, ultimately outputting a structured set of node instances, completing the accurate alignment and instantiation from process features to knowledge graph nodes.

[0046] Specifically, when the data normalization transformer performs a nonlinear mapping transformation from the feature space to the graph node space on the original feature vector group with standard semantic labels, it first calls the entity alignment algorithm in the pre-trained knowledge graph initialization model of the pipeline hot melt welding domain to normalize the input feature vector group in its original numerical domain in order to eliminate the mapping scale difference caused by different units. The normalized feature vector is input into the hidden layer representation generator of the entity alignment algorithm, and nonlinear combination features are extracted using the multilayer perceptron structure. The discriminative ability of the features in different process stages is enhanced by the activation function. The cosine similarity calculator performs cosine similarity calculations on the generated hidden layer representation, i.e., nonlinear combination features, and the domain knowledge graph node vector library to determine the matching degree between each feature vector and the candidate node vector; The Top-K truncator uses a Top-K truncating strategy to truncate the matching results, retains the candidate node with the highest matching degree, and establishes a mapping index between its node identifier and the original feature vector. The weight normalization processor performs weight normalization processing on the mapping index, so that the influence of the feature vector can be distributed proportionally when subsequent nodes are instantiated, and finally a candidate entity set is formed, including heating temperature node instances, pressure application duration node instances, ambient humidity node instances, and infrared image texture entropy value node instances of the molten zone.

[0047] Through the aforementioned nonlinear mapping and matching degree filtering processing, the entity alignment algorithm engine transforms the standard semantic feature vector from the previous step into node instance data that can be used for dynamic causal graph construction, thereby achieving accurate alignment of multi-source parameters in the welding process to semantic entities.

[0048] S2.3: For each node instance in the candidate entity set, the instantaneous state attribute and cumulative trend attribute are calculated by applying a time window sliding aggregation strategy. The discrete sensor readings are transformed into node attribute vectors that characterize the current welding stage process state, generating a process parameter node attribute set with dynamic timeliness.

[0049] S2.4: Based on the dynamic time-sensitive process parameter node attribute set, combined with predefined quality index association rules, instantiate the final joint quality assessment node, and establish an initial potential connection index between the process parameter node instance and the final joint quality assessment node to generate a complete set of process-quality associated entities.

[0050] like Figure 2 As shown, step S3: Using the typical causal rule library extracted from the historical qualified and abnormal weld sample dataset, the connection relationships between entities in the process-quality related entity set are dynamically activated to calculate the intensity, generating a dynamic causal graph model characterizing the evolution trajectory of the confidence of each causal edge under the current welding state. Specifically, this includes: S3.1: Based on the historical qualified and abnormal weld sample dataset, perform reverse engineering feature mining to extract the statistical correlation matrix between key process variables and generate a typical causal rule library containing at least seventeen typical causal rules; wherein, the key process variables include: heating temperature deviation, pressure application duration fluctuation rate, and gradient of texture entropy value change in the molten zone, etc.

[0051] The historical qualified and abnormal weld sample dataset refers to the basic data set used for feature mining and quality correlation analysis in reverse engineering of pipeline hot-melt welding processes. It includes weld samples collected from historical production records that have undergone quality inspection and are clearly marked as "qualified" or "abnormal." This dataset not only covers dynamic measured data of key process variables such as heating temperature deviation, pressure application duration fluctuation rate, and molten zone texture entropy gradient, but also transforms multi-source heterogeneous sensor numerical streams and image features into standardized multi-dimensional feature matrices through feature alignment processing under a unified time-series benchmark. Furthermore, it utilizes the composite Pearson correlation coefficient calculation method to construct a statistical correlation matrix between variables, thus forming a core data asset supporting causal rule extraction and process-quality dynamic correlation modeling. The historical qualified and abnormal weld sample dataset is constructed by collecting closed-loop data from the welding monitoring system and quality inspection system, followed by time-series alignment, feature extraction, and standardization cleaning.

[0052] Based on the node instances and their dynamic attribute data in the mapped process-quality related entity set, historical qualified weld sample data and abnormal weld sample data are feature aligned under a unified time-series benchmark to ensure consistency of data from different sources in terms of time axis and semantic labels. The aligned sample data undergoes multi-dimensional feature construction, converting the heating temperature sensor numerical stream into deviation features, the cumulative pressure application time into volatility features, and the molten zone infrared image texture entropy matrix into gradient features, forming a multi-source process variable feature matrix. Statistical correlation analysis is performed on the feature matrix using the composite Pearson correlation coefficient calculation method. Based on the linear correlation between multi-source process variables, a pairwise correlation matrix is ​​constructed, with the following formula:

[0053] Here, x and y are the sample value sequences of two types of process variables, respectively, and μ is the sample mean of the corresponding variable. Combining the statistical distribution characteristics of deviation, volatility, and gradient of change, significance tests are performed to eliminate non-significant correlations and retain highly significant variable pairs. By performing rule mining on the correlation matrix, variable combinations that frequently appear and have strong associations in historical anomaly samples are identified as antecedent conditions for causal rules. Using rule pattern induction, the 17 types of variable combinations and their triggering conditions are solidified into a typical causal rule library to support the subsequent calculation of the activation strength of dynamic causal edges.

[0054] In this embodiment, at least 17 typical causal rules characterizing the causal relationship between process parameters and welding quality are preset to drive the construction and reasoning of the causal map. These typical causal rules include, but are not limited to: insufficient heating temperature rule: if the average temperature during the heating stage is lower than the standard lower limit threshold, it is highly likely that the texture entropy of the molten zone will be lower than the acceptable range, thus leading to poor fusion defects; pressure duration overload rule: if the pressure application duration exceeds the standard upper limit threshold, and the ambient humidity is higher than the critical value at the same time, it may jointly lead to abnormal pipe deformation; temperature and pressure mismatch chain rule: if the variance of heating temperature fluctuation is higher than the threshold, and the pressure application curve is negatively correlated with it, it may trigger the risk of joint stress concentration.

[0055] Through the above statistical correlation calculation and rule extraction processing, the process-quality related entity set of the previous step is transformed into a typical causal rule library that can be used for causal relationship activation discrimination, realizing the structured storage and high-reliability definition of rule antecedent conditions.

[0056] For example, in a scenario containing 5000 historical qualified weld samples and 1500 abnormal weld samples, the heating temperature records of all samples are first mapped to temperature deviation features, with a set value of 260℃, and the deviation of each sample is calculated. The standard value for pressure application duration is set to 12 seconds, and the standard deviation of the deviation duration for each sample is calculated as the volatility feature. The standard value for the texture entropy value of the molten zone is set to 2.5, and the difference gradient between the actual value and the standard value for each sample is calculated. The above three types of features are constructed into 5000×3 and 1500×3 feature matrices, aligned with historical data under a unified time-series benchmark, with the alignment accuracy error controlled within ±0.5 seconds. The correlation between heating temperature deviation and cold weld defect level is calculated using the Pearson correlation coefficient, yielding a correlation coefficient of 0.82. After a significance test, the p-value is 0.004, indicating a significant correlation. The correlation coefficient between the volatility of pressure application duration and the probability of empty welds is 0.75, with a p-value of 0.007, also indicating a significant correlation. The correlation coefficient between the texture entropy gradient of the molten zone and the crack defect level was 0.79, with a p-value of 0.005. Process variable pairs with correlation coefficients greater than 0.7 and p-values ​​less than 0.01 were extracted and summarized into rules such as "a heating temperature deviation of ±8℃ lasting more than 12 seconds is significantly associated with the probability of cold welding." Seventeen typical causal rules were stored in the rule base for subsequent use in constructing dynamic causal graphs. Validation results show that this rule base can accurately cover the main causal relationships in abnormal samples, significantly improving the accuracy of causal edge activation determination.

[0057] S3.2: Using the rule antecedent conditions in the typical causal rule base, perform matching degree discrimination processing on the real-time monitoring values ​​of each entity in the process-quality related entity set to identify potential causal connection pairs that are active in the current welding process and generate an initial set of active causal edges.

[0058] Based on the existing typical causal rule library, the real-time monitoring values ​​of nodes in the process-quality related entity set are used as the input for discrimination. A rule antecedent condition matching mechanism is employed to perform a one-to-one comparison between node attribute value ranges and rule triggering conditions. For each causal rule, the normalized indicators involved in the antecedent conditions, such as process parameter thresholds, variation ranges, and durations, are mapped to the coefficient matrix of the discriminant function, and the condition matching degree is calculated using the node attribute vectors at the current time point. The matching degree calculation depends on the set normalized difference ratio formula. For example, the deviation ratio when comparing process parameter values ​​with rule thresholds can be obtained through the following formula:

[0059] Where x is the real-time monitoring value, As a rule threshold benchmark, and This defines the allowed range of values ​​for this parameter. From the calculated matching results, rule instances that satisfy all antecedent conditions and have a matching degree higher than the set lower limit are selected, and their corresponding node pair identifiers are extracted.

[0060] Within the set of rule instances that have been successfully matched, a causal relationship index table is uniformly invoked to generate potential causal connection pairs. Each connection pair consists of parameter nodes and quality assessment nodes from the rules, and directed edges are established according to the causal directionality defined in the rules. The generated set of connection pairs is then indexed and serialized to ensure that subsequent strength calculations can be performed by quickly accessing the unique keys of the node pairs.

[0061] By using multi-condition matching and pair extraction, the real-time monitoring data of the process-quality related entity set is transformed into a set of potential causal pair sets in an active state, thereby generating the initial set of active causal edges.

[0062] S3.3: Based on the real-time input deviation amplitude and duration window parameters of each causal edge in the initial set of activated causal edges, perform weighted cumulative activation intensity calculation within the sliding time window to quantify the instantaneous influence of each causal relationship at the current moment and generate an instantaneous activation intensity sequence with timestamps.

[0063] The real-time input deviation amplitude refers to the absolute or relative difference between the sensor monitoring value actually collected from the process-quality associated entity set at the current moment and its corresponding preset standard operating condition threshold or benchmark reference value. The real-time input deviation amplitude is obtained through standardized difference calculation or proportional conversion based on the process tolerance band, and is used to quantify the degree to which the process parameter corresponding to the causal edge instantaneously deviates from the normal allowable range. The larger the value, the further the current process state deviates from the standard, and the stronger the driving signal of this parameter anomaly for potential quality defects, thus contributing a higher instantaneous activation intensity in the weighted cumulative calculation.

[0064] The duration window parameter refers to the length of a predefined continuous time interval for performing weighted cumulative activation intensity calculation, typically measured in seconds or minutes. This parameter sets the time range for the system to backtrack historical data and statistically analyze the persistent impact of abnormal states. In the sliding time window calculation, the system checks the proportion of time or consecutive segments exceeding the preset sensitivity threshold for the real-time input deviation of the corresponding causal edge within this time window, and uses these as weighting factors. The duration window parameter ensures that only abnormal deviations lasting for a sufficiently long time are accumulated into significant activation intensity, thereby filtering out transient noise interference and accurately reflecting persistent abnormal patterns with process-damaging potential.

[0065] S3.4: Based on the historical attenuation coefficient of the instantaneous activation intensity sequence and the environmental interference factor of the current welding stage, perform exponential smoothing confidence evolution extrapolation processing on the instantaneous activation intensity sequence to eliminate random noise interference and reflect the continuous stability of causal relationships, and generate dynamic confidence evolution trajectory data of each causal edge.

[0066] The historical decay coefficient is a quantitative parameter established in the historical rule engine for each causal edge, used to characterize the decay rate and memory depth of the process anomaly pattern represented by that edge in historical data. The historical decay coefficient is pre-statistically derived based on the process variable category to which the causal edge belongs and the long-term evolution pattern in the past, reflecting the natural weakening of the historical influence of anomaly factors over time at different welding stages.

[0067] Based on the instantaneous activation intensity sequence and its corresponding timestamp data, the historical decay coefficient matrix stored in the rule engine is retrieved, and the corresponding decay factor is located in the matrix according to the process variable category to which the causal edge belongs and the past evolution mode.

[0068] The attenuation factor obtained from the location is numerically synthesized with the set of environmental interference factors in the current welding stage. The set of environmental interference factors includes parameters such as temperature fluctuation amplitude, humidity change rate, and equipment vibration energy level. The joint interference correction coefficient vector is formed by element-level multiplication.

[0069] The instantaneous activation intensity sequence is weighted by using a joint interference correction coefficient vector, so that the instantaneous intensity value under high interference conditions is reduced proportionally, and the instantaneous intensity value under low interference conditions is increased proportionally.

[0070] Based on the adjusted instantaneous activation intensity sequence, the core of the exponential smoothing algorithm is invoked to iteratively update the intensity value of each causal edge according to the time series. The exponential smoothing calculation formula is as follows:

[0071] in, This represents the smoothed confidence score at the current time t. This represents the instantaneous activation intensity value adjusted at the current time t. For smoothing coefficients, This represents the smooth confidence level value at the previous time t-1.

[0072] For the confidence evolution trajectory of each causal edge, a time series consistency test is performed to remove isolated peaks caused by sudden abnormal fluctuations, ensuring the continuity and stability of the trajectory curve.

[0073] The trajectory set after smoothing and consistency testing is output as dynamic confidence evolution trajectory data. Its structure includes edge identifiers, timestamp sequences, smoothed confidence sequences, and trajectory stability coefficients. Through exponential smoothing and interference correction, the instantaneous activation intensity sequence of the previous step is transformed into a dynamic confidence evolution trajectory that represents the continuous influence and stability, thereby achieving anti-interference optimization of the causal relationship confidence curve.

[0074] For example, in a welding task, the instantaneous activation intensity sequence has a length of 120 time slices. The corresponding attenuation factor in the historical attenuation coefficient matrix is ​​0.85. The environmental interference factor set has a temperature fluctuation amplitude parameter of 1.2, a humidity change rate parameter of 0.9, and an equipment vibration energy level parameter of 1.1. The joint interference correction coefficient vector is obtained by element-wise multiplication as {1.02, 0.765, 0.935}. ​​After applying this vector to the instantaneous intensity sequence, an exponential smoothing algorithm with a smoothing coefficient of 0.3 is called to iteratively calculate the intensity value of each time slice. During the calculation, the smoothed confidence value of the previous moment is used as input for recursive updating. The smoothed confidence curve has a reduced overall fluctuation amplitude, and the trajectory stability coefficient has increased from the original 0.65 to 0.92. The output dynamic confidence evolution trajectory data maintains high consistency in multiple sampling tests, effectively supporting the reliable screening of subsequent causal edge weights.

[0075] S3.5: Integrate the node attributes of the process-quality related entity set with the dynamic confidence evolution trajectory data, and perform dynamic graph topology reconstruction processing to form a dynamic causal graph model that can reflect the transmission path of quality risks in the welding process in real time.

[0076] Given that the dynamic and timely process parameter node attribute set and the dynamic confidence evolution trajectory data of each causal edge have been generated, the attribute vectors of all node instances in the process-quality related entity set are selected as the basic input for topology calculation, and the timestamp sequence and corresponding weight value in the confidence evolution trajectory of the causal edge are used as the edge attribute input.

[0077] Normalization is performed on the node attribute vectors to ensure consistency in graph network computation when there are large differences in multi-source heterogeneous data. The normalization factor is taken from the mean and variance of historical qualified weld samples.

[0078] The normalized node attribute and edge attribute matrices are simultaneously input into the dynamic topology reconstruction algorithm. The connection state between nodes is updated using an adjacency matrix adjustment strategy based on confidence weights. The adjacency matrix element update rule is calculated according to the following formula:

[0079] Where A is the initial adjacency matrix element value, and w is the weight coefficient of the current causal edge.

[0080] After the adjacency matrix is ​​updated, connectivity component analysis is performed on the graph structure to remove edge connections that have failed due to their weights being below a threshold, while retaining high-weight risk propagation paths.

[0081] A depth-first traversal is applied to the retained risk propagation paths to construct causal edge strips. Nodes and edges are arranged in time stamp order to achieve a directed dynamic mapping of welding process quality risks in the graph.

[0082] The dynamic causal edge pairs constructed above are applied to the graph visualization model to generate a dynamic graph view that simultaneously reflects the changes in the state of risk propagation nodes and edge confidence.

[0083] Through dynamic topology reconstruction, the node attributes and edge confidence trajectory data from the previous step are transformed into a dynamic causal graph model that can reflect the transmission path of quality risks in the welding process in real time, thereby realizing the temporal location and propagation analysis of abnormal factors in causal edge strips.

[0084] For example, in the reconstruction of a risk propagation path containing heating temperature nodes and molten zone texture entropy value nodes, the normalized value of the heating temperature node attribute vector is set to 0.72, the normalized value of the molten zone texture entropy value node attribute vector is set to 0.65, the initial adjacency matrix element value is 0.50, the current causal edge weight coefficient is 0.30, and the updated value calculated according to the formula is 0.65. This value is higher than the preset threshold of 0.60, so the connection is retained and added to the high-weight propagation path set in connected component analysis. The final constructed dynamic graph shows that when the heating temperature deviates from the set value and the entropy value decreases, the path presents a continuously rising confidence curve in the graph. The application effect is that the risk propagation range is significantly improved, the anomaly location efficiency is enhanced, and the dominant nodes and key edges of the risk chain can be intuitively displayed in graph visualization.

[0085] like Figure 3 As shown, step S4: Based on the dynamic causal graph model, the highest-weighted causal edges with confidence levels higher than a preset threshold are selected, and the directed subgraph encoding is extracted to generate a snapshot of the abnormal attribution path for this welding process. The directed subgraph encoding consists of a unique identifier for the starting node, a unique identifier for the ending node, a sequence of intermediate jump nodes, and confidence interval identifiers for each edge. Specifically, it includes: S4.1: Threshold filtering is performed on the confidence evolution trajectory of all causal edges in the dynamic causal graph model to select a set of candidate causal edges with confidence higher than a preset threshold, thereby eliminating low-correlation interference connections and establishing the foundation for constructing high-weight causal edges.

[0086] S4.2: Execute a directed path search algorithm based on the candidate causal edge set to calculate the total weight of the connected path from the process parameter starting node to the quality defect ending node, thereby generating the highest weight causal edge sequence arranged in descending order of weight as the core candidate object for anomaly attribution.

[0087] S4.3: Perform topological structure parsing on each causal edge in the highest weight causal edge sequence to extract the unique identifier of the starting node, the unique identifier of the ending node, and the sequence of intermediate jump nodes, thereby forming the original directed subgraph encoded data representing the anomaly propagation logic.

[0088] The input conditions include the highest weight causal edge sequence arranged in descending order of weight. This sequence is derived from the total weight calculation result of the path obtained by the directed path search algorithm from the candidate causal edge set. The data structure includes the connected path information from the process parameter start node to the quality defect end node.

[0089] For the highest weight causal edge sequence, node index parsing is performed. The node index array of each causal edge in the sequence is decomposed into the starting node index, the ending node index, and the intermediate node index group according to the path order. During the decomposition process, the original topological arrangement between nodes is maintained to avoid path order distortion.

[0090] Perform a unique identifier mapping operation on the decomposed node index, call the globally unique identifier (UUID) mapping table of node instances in the domain knowledge graph, and replace the digital index value with the entity identifier one by one to ensure that the subsequent path encoding has cross-system uniqueness.

[0091] Based on the mapped node identifier group, the path connectivity structure is reorganized. The unique identifier of the starting node, the identifier group of intermediate nodes, and the unique identifier of the ending node are connected in topological order to form an ordered node sequence. The directional property of the path is maintained and the direction consistency check is performed to remove abnormal connections that do not conform to the causal edge direction.

[0092] The ordered node sequence is encapsulated, and the node position relationships and edge direction identifiers are marked in the data structure using the directed graph encoding standard, providing a deterministic topological skeleton for subsequent confidence interval mapping.

[0093] By using topological structure analysis and unique identifier mapping, the results of the previous step are transformed into original directed subgraph encoded data that represents the anomaly propagation logic, achieving the expected technical effect of visualizing causal edge input and solidifying the logical structure.

[0094] For example, in a welding operation, the highest-weighted causal edge sequence contains three paths, one of which is temperature node index 12 → melt texture entropy node index 8 → cooling rate node index 15 → quality defect node index 21. During node index parsing, this chain is decomposed into a starting node index 12, an ending node index 21, and an intermediate node index group {8,15}. The mapping table is called, assigning index 12 to UUID "TEMP-AX92", index 8 to UUID "TEXTURE-IR03", index 15 to UUID "COOL-SLOPE07", and index 21 to UUID "XDEFECT-LVL5". This is then reassembled into an ordered node sequence ["TEMP-AX92", "TEXTURE-IR03", "COOL-SLOPE07", "XDEFECT-LVL5"]. After performing a direction consistency check, it is confirmed that the path is propagating in the forward direction. The data is encapsulated into a directed subgraph encoding structure, where the edge identifiers are connected in the following order: "TEMP-AX92" → "TEXTURE-IR03" → "COOL-SLOPE07" → "XDEFECT-LVL5". This encoded data structure can be directly referenced in subsequent confidence interval mapping, realizing a logically solidified representation of the welding anomaly formation process. In the verification, the positioning accuracy was significantly improved and the path parsing time was reduced.

[0095] S4.4: Based on the original directed subgraph encoding data, the confidence intervals corresponding to each causal edge are discretized and mapped to generate an edge attribute label set containing high, medium and low confidence interval identifiers, thereby quantifying the logical reliability level of each causal path.

[0096] Based on the unique identifiers of the starting and ending nodes and the sequence of intermediate jump nodes contained in the original directed subgraph encoded data, the dynamic confidence evolution trajectory values ​​corresponding to each causal edge are read as input to construct a calculation object for confidence interval mapping. In the calculation object, the confidence value normalization processing module is first called to uniformly scale confidence values ​​of different dimensions and data ranges to the [0,1] interval, eliminating numerical imbalance caused by differences in acquisition channels. Based on the normalization result, an interval partitioning algorithm is executed. According to the preset three-segment boundary values ​​α1 and α2, the normalized confidence values ​​are mapped to high, medium, and low discrete level labels respectively. The boundary values ​​are set based on the quantile calculation results of historical statistical distributions and are obtained through the following calculation formula:

[0097] Where P(q) is the q-quantile of the confidence value sequence, and max(q) and min(q) are the maximum and minimum values ​​of the sequence, respectively. Using these threshold values, the confidence values ​​of each causal edge are discretized and mapped to generate an edge attribute label set. Each element in the label set corresponds to the logical reliability level of a causal edge at the current welding stage. The edge attribute label set is bound to the node sequence encoded in the original directed subgraph, forming structured data with both node identification and edge confidence level attributes, providing direct input for subsequent path snapshot encapsulation. Through discretization and mapping, the results of the previous step are transformed into quantified logical reliability indicators, achieving a clear classification of reliability for different edges in the causal path structure, improving the interpretability of abnormal attribution paths and their subsequent legal credibility.

[0098] S4.5: Perform serialization and encapsulation processing on the unique identifier of the starting node, the unique identifier of the ending node, the sequence of intermediate jump nodes and the set of edge attribute labels to generate an abnormal attribution path snapshot of this welding process, thereby outputting a standardized logical evidence storage unit with a fixed format that can be directly used for hash digest calculation.

[0099] Based on the complete data structure generated in the previous sub-step, including a unique identifier for the starting node, a unique identifier for the ending node, a sequence of intermediate jump nodes, and a set of edge attribute labels, byte-order structured encapsulation is performed. Each node ID is bound to its corresponding edge attribute label according to a fixed field arrangement rule, forming a prototype logical evidence data that can be serialized. For this prototype data, a custom data type conversion module is called to uniformly encode the node identifier field into a fixed-length integer representation and map the edge attribute label set to an enumerated combination of sign bits, thus achieving encoding consistency during cross-platform transmission. The encoded data fields are then processed for serialization format organization. A multi-segment binary concatenation method is used to integrate the node sequence segments and edge label segments into a continuous byte stream in a preset order, ensuring that the position and content of each field are unambiguous during subsequent hash digest operations. Length labeling and error detection bit appending operations are performed on the concatenated continuous byte stream. By adding a fixed-length Cyclic Redundancy Check (CRC) code to the end of the data, structural integrity verification is achieved during transmission or storage. A binding protocol is applied to the serialized byte stream containing node sequences, edge labels, length annotations, and checksums. This byte stream is then embedded into the payload area of ​​a standardized logical evidence storage unit, and version number, format identifier, and generation timestamp are filled in. This achieves structural solidification and format finalization of the snapshot of the abnormal attribution path in this welding process. Through serialization encapsulation, the topology parsing result from the previous step is transformed into a standardized logical evidence storage unit with a fixed format that can be directly used for hash digest calculation. This ensures the verifiability and portability of the causal path snapshot in subsequent data solidification stages.

[0100] For example, in an anomaly attribution analysis of a welding process, the highest-weighted causal edge contains 5 nodes: node IDs are 101, 203, 315, 420, and 509, and edge attribute label sets are high, medium, high, and low, respectively. The node IDs are converted to 4-byte integers (big-endian) and arranged in link order. The edge attribute labels are mapped to single-byte sign bits (high = 0x02, medium = 0x01, low = 0x00) according to enumeration encoding. After concatenation, a 29-byte original logical evidence byte stream is formed. A 2-byte CRC checksum (calculated using polynomial 0x1021) is appended to the end of this byte stream. A version number field (0x01), a format identifier field (0xA5), and an 8-byte timestamp field (UNIX microseconds) are then added to the encapsulation header, ultimately forming a 41-byte standardized logical evidence unit. In the verification stage, CRC checks confirm that the data has not been lost or tampered with, and this fixed-format data unit is input into the national cryptographic SM3 hash engine to generate a unique digest fingerprint value. When storing the abstract fingerprint value on the blockchain, it can significantly improve the matching efficiency of the verification process, and the integer and enumeration encoding methods used in the serialization encapsulation ensure parsing consistency and processing stability in cross-system environments.

[0101] Step S5: Perform a hash operation on the anomaly attribution path snapshot to generate a data digest, and package the data digest along with the spatiotemporal identity identifier into a minimal transaction unit to generate a blockchain evidence record. Specifically, perform a national cryptographic SM3 hash algorithm operation on the anomaly attribution path snapshot to generate a fixed-length data digest, and package the data digest along with the spatiotemporal identity identifier into a minimal transaction unit to complete the lightweight on-chain existence operation to generate a blockchain existence record. The spatiotemporal identity identifier includes: a timestamp, a unique identification code for the welding equipment, and an operator biometric binding code. Specifically, it includes: S5.1: The unique identifiers of the starting node, the unique identifiers of the ending node, the sequence of intermediate jump nodes, and the confidence interval identifiers of each side contained in the snapshot of the abnormal attribution path are serialized and encoded to generate a raw data stream to be hashed with a unified byte order structure, so as to ensure the integrity and determinism of the causal logic structure in subsequent hash operations.

[0102] S5.2: Based on the original data stream to be hashed, the national cryptographic SM3 hash algorithm engine is called to perform a one-way hash transformation operation to generate a fixed-length data digest fingerprint, compressing the variable-length causal path semantic information into a unique cryptographic feature value, thereby eliminating the redundancy of the original data and establishing a data anti-tampering benchmark.

[0103] The input is the raw data stream to be hashed, generated in the previous sub-step, which has a unified byte order structure. This data stream contains causal logical structured information such as unique identifiers for the start and end nodes, sequences of intermediate jump nodes, and confidence interval identifiers for each edge. When the national cryptographic SM3 hash algorithm engine is called to perform a one-way hash transformation operation on this data stream, the binary data blocks with unified byte order are first loaded into the hash engine's message buffer to ensure that the input byte order is consistent with the algorithm's internal operation benchmark. The message buffer is processed into 512-bit groups as specified by SM3. For data that is less than a group, additional padding based on length information is performed on the tail data to ensure that the overall message length meets the requirements of the grouping operation. A message expansion operation is performed on each group, using the SM3 message scheduling function to generate a message word sequence arranged in a specified number of rounds, and mapping it to the internal working variable register. The initial value of the working variable register is set to eight 32-bit constant values ​​of the SM3 standard, and 64 rounds of Boolean operations and modulo 2 operations are performed in combination with the message word sequence. 32 Addition and circular left shift operations are used to iteratively update the state of the working variables. All grouping operation outputs are concatenated to form a final 256-bit digest fingerprint, which is then encapsulated as a unique cryptographic feature into fixed-length data. Through this processing method, variable-length causal path semantic information is compressed into a fixed-length identifier, achieving redundancy elimination of the original data and establishing a tamper-proof baseline.

[0104] S5.3: Perform multi-field aggregation and encapsulation processing on the data digest fingerprint, timestamp, welding equipment unique identification code and operator biometric binding code to construct a minimized transaction unit payload containing complete identity authentication and spatiotemporal context information, so as to realize the logical binding and structured organization of key evidence storage elements.

[0105] The input objects are a fixed-length data digest fingerprint, timestamp, unique identification code of welding equipment, and operator biometric binding code obtained by the national cryptographic SM3 hash algorithm.

[0106] The data digest fingerprint is preprocessed by dividing it into fixed-length blocks according to a preset byte order rule and adding field type identifiers to maintain the consistency of the data structure during subsequent aggregation.

[0107] The timestamp parameter is standardized by converting the UTC time value into a uniform millisecond-level integer representation and adding a time precision flag before the field to eliminate parsing discrepancies between different systems.

[0108] The unique identification code of the welding equipment is optimized by using the Base58 encoding scheme to compress the character length and adding a displacement check code to enhance the fault tolerance of the equipment identification field.

[0109] Perform biometric serialization encapsulation on the operator's biometric binding code, converting the fingerprint feature vector or iris image hash value into a fixed-length binary sequence and attaching a biometric type identifier to ensure code compatibility for different biometric acquisition modes.

[0110] The four fields processed above are aggregated at the field level according to the preset field order, and the data digest fingerprint, timestamp field, device identification code field and biometric binding code field are arranged in sequence. Separator bytes are inserted between the fields to facilitate parsing and segmentation.

[0111] The aggregated field set is processed by structured payload construction to generate a three-part payload structure containing a field offset table, a field length table, and a field content body, enabling location-based access to the field content.

[0112] By logically binding the field offset table and the field content body, an intrinsic relationship between complete identity authentication information and spatiotemporal context information is established, thus forming a minimized transaction unit load.

[0113] By using a multi-field aggregation and encapsulation process, the hash digest result from the previous step is transformed into a storage payload with identity binding, time positioning, and structured organization characteristics, thereby enabling the formatted input required for subsequent on-chain consensus and digital signature.

[0114] For example, the hash digest fingerprint of a snapshot of a welding anomaly attribution path is set to 64 bytes in length, divided into four segments of 16 bytes each, and identified by a field ID. The timestamp value of 1687345012345 milliseconds is obtained from the device and uniformly converted to a UTC millisecond integer. The welding equipment unique identification code, after Base58 encoding, is compressed to 22 characters in length and a 2-byte CRC checksum is added. The operator biometric binding code is obtained by taking the SHA256 hash value of the fingerprint feature, serializing it to generate a 32-byte binary sequence, and adding a biometric type identifier 0x01. The four fields are aggregated in the order of digest → timestamp → device code → biometric code, and a 0xFF separator is inserted to generate payload data with a total length of approximately 150 bytes. A field offset table is constructed to record the starting position of each field (e.g., the timestamp is at offset 64, and the device code is at offset 78), and a field length table records the length of each field. The payload content contains the actual field data. After being parsed in the test chain node, the above payload structure can locate any field within 1 millisecond. After verification by the on-chain consensus module, the pass rate is significantly improved, meeting the requirements for generating standardized evidence storage transaction packages in the future.

[0115] S5.4: Based on the minimized transaction unit payload, perform lightweight consensus protocol verification and digital signature attachment operations to generate a standardized blockchain evidence-based transaction package with network broadcast qualification, ensuring that the legality verification of the transaction content among distributed nodes passes and meets the on-chain transmission format specifications.

[0116] Based on the minimized transaction unit payload generated in the previous step, the pre-verification module of the lightweight consensus protocol is invoked to perform a format integrity scan on the field structure of the payload, verifying whether the binary encoding sequence of the data digest fingerprint, timestamp, welding equipment unique identification code and operator biometric binding code meets the byte order and length requirements in the blockchain on-chain data specification.

[0117] After the format verification is completed, the dynamic verification mechanism of the consensus node list is triggered. According to the node participation ratio preset by the lightweight Byzantine fault-tolerant consensus algorithm, the node identity legality comparison and link reachability detection are performed to ensure that the transaction unit payload has the conditions to enter the consensus process in the current network topology.

[0118] Based on the successful node verification, and combined with the timestamp information embedded in the original payload, the transaction timing sorting module is invoked to calculate the sorting weight of all transactions awaiting consensus across the network. The payload of this transaction unit is then inserted into the mixed sorting result of time priority and digest fingerprint hash value to ensure the consistency of the global execution order of transactions.

[0119] After sorting is completed, the process proceeds to the digital signature appending stage. The ECC elliptic curve signature algorithm certified by the State Cryptography Administration is called to sign the hash digest of the transaction unit payload with a private key, generate a digital signature field, and append it as a new field to the end of the payload to cover all evidence elements.

[0120] The payload after the signature is appended undergoes standardized transaction package encapsulation processing, adding the chain protocol identifier header, version number, protocol length field, and network broadcast control parameters. After field assembly and CRC cyclic redundancy check, a standardized blockchain evidence storage transaction package with network broadcast qualification is formed.

[0121] By using a lightweight consensus protocol for verification and digital signature appending, the minimal transaction unit from the previous step is transformed into a standardized blockchain-based notarized transaction package with the ability to transmit across the entire network and verify its legitimacy, thus achieving the expected technical effect of meeting the on-chain transmission format specifications and ensuring tamper-proof security.

[0122] For example, in the process of attributing anomalies in pipeline hot-melt welding, the minimum transaction unit payload includes a set of data digest fingerprints, timestamps accurate to milliseconds, a 16-bit unique identification code for the welding equipment, and a 256-bit operator biometric binding code. When performing a format integrity scan on this payload, the digest fingerprint length is set to 256 bits, the timestamp length to 8 bytes, the equipment code length to 2 bytes, and the biometric code length to 32 bytes. If any field does not match, the consensus process is rejected. In the consensus node verification stage, 2 / 3 of the total number of currently online nodes is selected as the minimum verification ratio. Nodes that can participate are screened through node identity public key matching and Ping latency detection to ensure network health. In the time-series sorting calculation, the timestamp weight coefficient is set to 0.7, and the digest fingerprint hash value weight coefficient is set to 0.3. The sorting function is then used to determine the weight of the nodes. The transaction insertion position is calculated, where t is the timestamp sequence number and h is the hash value sequence number. During the digital signature appending process, the ECC curve secp256k1 is selected, and the digest fingerprint is signed using the system private key, generating 64 bytes of signature data which is appended to the payload. After assembling the protocol header, version number, and length fields, a CRC32 check is performed to generate a 4-byte checksum, ultimately forming a standardized blockchain-based evidence storage transaction package. After broadcast verification, network consensus can be achieved with an average network latency of 30ms, ensuring the evidence storage transaction package is legally uploaded to the blockchain and meets tamper-proof requirements.

[0123] S5.5: Perform on-chain submission and block confirmation listening processing on the standardized blockchain evidence storage transaction package to generate a blockchain evidence storage record with a globally unique transaction hash index, and complete the irreversible solidified storage of welding construction anomaly attribution logic from the local edge to the distributed ledger.

[0124] Step S6: Read the blockchain evidence record of the specified weld joint from the blockchain and decode it to restore the causal path structure. Simultaneously, pull the original parameter fragments within the upstream and downstream time windows corresponding to the nodes involved in the causal path structure from the off-chain distributed database to generate a bidirectional graph reconstruction dataset. Specifically, this includes: S6.1: Obtain the blockchain evidence record of the specified weld joint, and use the national cryptographic SM3 hash reverse mapping verification mechanism to perform integrity verification on the minimum transaction unit containing timestamp, unique identification code of welding equipment and operator biometric binding code, so as to generate a cryptographically confirmed legitimate causal path digest code.

[0125] Using blockchain-stored records as input, a globally unique transaction hash index of the specified weld joint is used to locate the target transaction package data.

[0126] An on-chain data extraction operation is performed on the transaction package data to obtain a minimized transaction unit payload containing a timestamp, a unique identification code for the welding equipment, and a biometric binding code for the operator, while maintaining the byte order consistency with the original evidence to prevent structural misjudgment.

[0127] Using the national cryptographic SM3 hash reverse mapping verification mechanism, the payload of the minimized transaction unit is recalculated one-way with the data digest fingerprint generated during on-chain notarization, and the hash values ​​are compared to determine the data integrity.

[0128] For the hash recalculation process, let the original evidence data digest fingerprint be H, the minimized transaction unit serialization of the original data stream be D, and the hash function be SM3. Then the verification process satisfies... When the comparison results match, it is confirmed that the data has not been tampered with.

[0129] After hash consistency is confirmed, a high-precision format check is performed on the timestamp field to ensure that it is within the preset global time series reference window, avoiding path index offsets caused by cross time zones or system delays.

[0130] The unique identifier field of the welding equipment is matched with the public key certificate chain of the registered equipment to verify the consistency between the on-chain identity authentication and the off-chain device registration information, thereby eliminating the risk of illegal device data injection.

[0131] Perform a similarity comparison between the operator's biometric binding code and the original biometric template library. Let the binding code vector be... Given F and template vector T, the similarity calculation formula is: When the similarity exceeds a preset threshold, the operator's identity is confirmed to be valid.

[0132] Through the above chain verification process, the on-chain evidence record of the previous step is transformed into a cryptographically confirmed legitimate causal path digest code, achieving the effect of trusted data transformation from evidence storage to a usable causal topological skeleton.

[0133] For example, in a pipeline hot-melt welding construction responsibility traceability task, after receiving a report generation request, the system extracts the target transaction package through the transaction hash index. The package contains a timestamp with nanosecond precision, a unique welding equipment identifier (a 16-bit hexadecimal sequence), and an operator binding code (a 128-dimensional biometric vector). The result of the SM3 recalculation is completely consistent with the evidence digest fingerprint. The timestamp falls within the global time series benchmark ±500 microsecond window. The unique welding equipment identifier matches successfully in the registered equipment certificate chain. The similarity calculation between the binding code and the template library yields S=0.923, which is higher than the preset threshold of 0.85. The verification results indicate that the transaction unit is legal and the data is complete. The legal causal path digest encoding output in this scenario is directly used for subsequent directed subgraph decoding, achieving high-precision, low-latency responsibility traceability link initiation.

[0134] S6.2: Based on the legal causal path digest encoding, execute the directed subgraph decoding algorithm to parse the unique identifiers of the starting node, the unique identifiers of the ending node, the sequence of intermediate jump nodes, and the confidence interval identifiers of each edge, so as to restore the structured causal path skeleton representing the logical topology of this welding process.

[0135] S6.3: Based on the sequence of intermediate jump nodes in the structured causal path skeleton, calculate the time offset of each process-quality related entity under a unified time reference, and combine it with the preset time window expansion coefficient to generate an accurate data retrieval range index covering the upstream and downstream time windows of the key causal edge.

[0136] The intermediate jump node sequence in the structured causal path skeleton is parsed using node time attribute analysis to extract the start and end times of the process parameter status records corresponding to each node under a unified time series benchmark. Based on the node time attributes, the time offset of each process-quality related entity relative to the zero point of the unified time series benchmark is calculated to ensure the time series comparability of different data sources. A preset time window expansion coefficient is introduced to expand the time offset forward and backward respectively to cover potential interference factors and preceding influencing steps that may exist upstream and downstream of the causal edge. Boundary constraint calculation methods are used to correct the upper and lower bounds of the expanded time range to ensure that the time retrieval range does not exceed the original data storage cycle and avoids crossing irrelevant process segments. The expanded time range of each node is aggregated to generate a merged range index covering the two time windows upstream and downstream of all key causal edge nodes, forming a data time range structure that can be used for precise retrieval in the off-chain database. Through the above time offset calculation and window expansion processing, the structured causal path skeleton of the previous step is transformed into a data retrieval index with precise time period positioning capabilities, realizing the efficient extraction and preparation of original parameter fragments from the off-chain database.

[0137] S6.4: Using the precise data retrieval range index, access the off-chain distributed database and extract the heating temperature sensor data stream, pressure application duration timer data stream, ambient humidity detector data stream, and molten zone infrared image texture entropy value data stream corresponding to the nodes involved in the structured causal path skeleton, so as to form a multi-source heterogeneous original parameter fragment set.

[0138] S6.5: The structured causal path skeleton and the multi-source heterogeneous original parameter fragment set are semantically aligned and fused to construct a dual-attribute data object containing logical topology and measured numerical distribution, so as to generate a bidirectional graph reconstruction dataset.

[0139] The causal path skeleton obtained by the directed subgraph decoding and the set of multi-source heterogeneous original parameter fragments are loaded into the two input buffers of the semantic matching module to ensure that the topological structure information and the measured numerical information are in a unified processing framework.

[0140] The semantic mapping table of domain process parameters is invoked to perform semantic label parsing on the node identifiers involved in the skeleton, mapping the structure definition to standardized process entity labels, and relabeling the data streams of the original parameter fragment set within the same label space to establish a precise correspondence between logical nodes and physical data sources.

[0141] Based on the correspondence, a time-series alignment algorithm is executed to calculate the index position of each process entity label under a unified time-series benchmark. The original parameter fragments are then limited to the time window range corresponding to the nodes in the causal skeleton, forming a localized numerical set containing the internal and external context of the process.

[0142] The feature vector construction process is performed on the local numerical set. The multi-scale statistical extraction method is used to generate the measured attribute vector of the node. The attribute vector is then bound to the topological location information of the skeleton node to form a fused node object.

[0143] By using the edge attribute fusion algorithm, the confidence interval identifier of the edges in the skeleton and the measured attributes of the associated node objects are synthesized through a weighted correlation matrix to form a dual attribute data object containing both logical topological structure and measured numerical distribution.

[0144] Through the semantic alignment and attribute fusion processing methods described above, the causal path skeleton and multi-source heterogeneous original parameter fragments from the previous step are transformed into a bidirectional graph reconstruction dataset.

[0145] For example, in the execution of responsibility tracing for a pipeline hot-melt welding, the causal path skeleton includes the nodes "heating temperature," "melting zone texture entropy value," and "final joint quality level." The corresponding original parameter fragments are a temperature data stream with a sampling frequency of 100Hz, an infrared image texture entropy matrix sequence with a resolution of 640×480, and a quality level record value, respectively. A semantic matching mapping table maps the "heating temperature" label to temperature channel ID=TMP01, and "melting zone texture entropy value" to infrared channel ID=IRTX05. The temporal alignment algorithm uses a global clock as a reference, setting the temperature data window length to 2.5 seconds and the IR image entropy value window length to 3 seconds, with an offset between the two not exceeding 50 milliseconds. A statistical extraction method calculates the mean and standard deviation within the temperature window, and the average entropy and coefficient of variation within the entropy value window, constructing node attribute vectors from these values. For example, if the temperature mean is 245.6, the standard deviation is 0.8, and the average entropy value is 1.32, the confidence interval of the corresponding edge is high. The edge attribute fusion algorithm maps confidence labels to numerical weights of 0.9 and synthesizes the fusion matrix according to the following formula:

[0146] in This is a matrix of node attribute vectors. The edge weight vector matrix is ​​multiplied, and the product is used as a dual-attribute fusion matrix input into the graph neural network. After this processing, the generated bidirectional graph reconstruction dataset contains high-confidence distinctions between logical topological three nodes and two edges, as well as the corresponding measured value distributions, effectively improving the interpretability and accuracy of subsequent bias analysis.

[0147] Step S7: Based on the bidirectional graph reconstruction dataset, calculate the deviation between the path semantics and the measured data distribution, identify the link where the measured value deviates most significantly from the expected model, and generate the dominant anomaly factor localization result. Here, the path semantics specifically refers to the initial graph neural network input tensor generated in step S7.1. The measured data distribution specifically refers to the physical measurement data distribution embedding vector generated in step S7.3. In subsequent calculations, the system achieves accurate localization of the anomaly link by comparing the deviation between the causal logic semantic embedding vector and the physical measurement data distribution embedding vector. Specifically, this includes: S7.1: Perform node feature vectorization mapping on the directed subgraph encoding structure in the bidirectional graph reconstruction dataset. Utilize a pre-trained knowledge graph embedding algorithm for pipeline hot-melt welding to convert the unique identifiers of the starting node, the unique identifiers of the ending node, and the sequence of intermediate jump nodes into a high-dimensional set of process entity feature vectors, so as to construct an initial graph neural network input tensor containing semantic information.

[0148] The directed subgraph encoding structure in the bidirectional graph reconstruction dataset is parsed to determine the unique identifiers of the starting and ending nodes, as well as the unique identifiers and sequential positions of intermediate jump nodes in the topology, serving as the basis for subsequent feature mapping. The identifier information of each node is matched against a predefined mapping table of pipeline hot-melt welding process entity types to extract the corresponding semantic tags and associated attributes, including process stage, equipment number, and environmental parameter category. For the matched semantic tags, a pre-trained knowledge graph embedding algorithm for pipeline hot-melt welding is invoked, loading the trained node vector encoding matrix and extracting the initial embedding vector using the node identifier as an index. Feature expansion processing is performed on the initial embedding vector, mapping the dynamic attribute values ​​of the nodes (such as current temperature, pressure duration, etc.) to numerical features according to normalization rules, and concatenating them with the embedding vector at the vector level to form a high-dimensional process entity feature vector containing both static semantics and dynamic attributes. Position encoding is added to the node feature vectors of the entire sequence, injecting the relative positions of the nodes in the directed subgraph into the feature vectors in a sine and cosine encoding form to preserve path order information. After the above expansion and encoding injection are completed, an initial graph neural network input tensor containing feature vectors of all nodes in the sequence is generated, and the path semantics and node attributes are fully expressed in the form of a feature matrix. Through node identifier parsing, semantic label matching, embedding vector extraction, dynamic attribute expansion, and position encoding injection, the bidirectional graph reconstruction dataset from the previous step is transformed into a high-dimensional set of process entity feature vectors, thus realizing the construction of the graph neural network input tensor.

[0149] For example, in a pipeline hot-melt welding anomaly tracing scenario, the directed subgraph encoding structure includes a starting node with a unique identifier of T001 (representing the heating temperature stage), an ending node with a unique identifier of Q005 (representing the joint quality defect level), and intermediate jump node sequences of P002 (pressure application stage) and E003 (ambient humidity stage). In the node identifier mapping table, T001 corresponds to the semantic label "heating temperature node", P002 corresponds to "pressure application node", E003 corresponds to "ambient humidity node", and Q005 corresponds to "quality assessment node". In the pre-training matrix of the knowledge graph embedding algorithm, the static embedding vector of T001 is 128-dimensional, and its dynamic attribute, the current temperature value of 235℃, is mapped to 0.82 according to the normalization rule; the static embedding vector of P002 is 128-dimensional, and its dynamic attribute, the pressure duration of 12.5 seconds, is mapped to 0.76; the static embedding vector of E003 is 128-dimensional, and its dynamic attribute, the humidity of 68%, is mapped to 0.68; the static embedding vector of Q005 is 128-dimensional, and its dynamic attribute, the defect level 2, is mapped to 0.2. During vector concatenation, the 128-dimensional static embedding vector of each node is combined with the 1-dimensional dynamic attribute value to form a 129-dimensional feature vector. Position encoding uses a sine and cosine method, with the node positions being 0, 1, 2, and 3 respectively. The encoded components are calculated using the following formula:

[0150] in, Here, pos is the positional encoding value, i is the feature dimension index, and 129 is the concatenated feature length. Positional encodings are calculated for each of the four nodes and appended to the 129-dimensional feature vector to form a complete feature matrix. The final input tensor shape is [4, 129 + encoding dimension], which significantly improves the model's ability to comprehensively discriminate path sequence information and dynamic process attributes in subsequent graph convolution operations.

[0151] S7.2: Based on the initial graph neural network input tensor, perform multi-layer graph convolution aggregation operation, fuse the logical weight information represented by the confidence interval identifiers of each edge, and generate a causal logical semantic embedding vector representing the topological structure of the dynamic causal graph model, so as to extract the process rule evolution features hidden in the causal path.

[0152] Based on the initial graph neural network input tensor, the first layer of graph convolution operation is performed on the feature vectors of each node. The features of adjacent nodes are weighted and converged using the numerical weights corresponding to the edge confidence intervals in the adjacency matrix, thereby generating the first layer of feature representation that integrates local topological information.

[0153] Based on the first layer of feature representation, a second layer graph convolution operator is constructed. A multi-scale aggregation mechanism is used to simultaneously aggregate the feature vectors of two-order neighboring nodes, and the edge weights are normalized to eliminate the influence introduced by different path lengths, thereby obtaining a second layer of feature representation that includes a wider range of topological dependencies.

[0154] The output of the second layer is used as input to perform the graph convolution aggregation operation of the third layer. In this process, an edge confidence smoothing factor is introduced. By updating the weight matrix, the high-confidence edge is enhanced in the aggregation process, while the weight of the low-confidence edge is weakened under the action of the smoothing factor, so as to ensure that the semantic embedding vector more accurately reflects the stability of causal relationship.

[0155] After the third layer aggregates the output, feature normalization and nonlinear activation functions are applied to compress the range of each node's embedding vector and enhance the nonlinear discrimination capability, thereby improving the separation degree of causal logic features under different path semantics.

[0156] Multi-layer graph convolution aggregation is represented by the following mathematical form:

[0157] in, The node feature matrix, This is the normalized adjacency matrix adjusted by confidence interval mapping. This is the weight matrix. It is a non-linear activation function.

[0158] By combining multi-layer graph convolutional aggregation with edge confidence weight fusion, the initial graph neural network input tensor from the previous step is transformed into a causal logical semantic embedding vector that represents the topological structure of the dynamic causal graph and its logical weight distribution, thereby achieving a high-dimensional representation of the evolutionary features of implicit process rules in the causal path.

[0159] S7.3: Perform time-series feature extraction processing on the original parameter segments within the upstream and downstream time windows corresponding to the nodes involved in the causal path structure, and use sliding window statistics calculation and frequency domain transformation algorithms to generate physical measurement data distribution embedding vectors that characterize the distribution characteristics of measured values, so as to quantify the multi-source heterogeneous data fluctuation patterns in the actual construction process.

[0160] S7.4: Based on the causal logic semantic embedding vector and the physical measurement data distribution embedding vector, perform cross-modal attention alignment mechanism calculation, construct semantic-measured deviation measurement matrix, and generate a multi-dimensional deviation value sequence between path semantics and measured data distribution to identify the degree of inconsistency between theoretical expected model and actual operating state in key process links.

[0161] Based on the generated causal logic semantic embedding vector and physical measurement data distribution embedding vector, the two types of embedding vectors are input into the cross-modal attention alignment computation unit. The semantic embedding is specified as a set of query vectors and the physical distribution embedding is specified as a set of key-value pairs. This ensures that the query vector dimension and the key-value matrix dimension are linearly normalized so that they can be accepted by the attention mechanism.

[0162] Within the alignment computation unit, dot product similarity calculation is first performed to obtain the matching confidence matrix between each causal logical semantic embedding and all physical distribution embeddings. The magnitude of the dot product result is then normalized using a scaling factor to prevent gradient vanishing or exploding.

[0163] The standardized matching confidence matrix is ​​normalized by Softmax to generate the semantic-to-measure attention weight distribution. This weight distribution is used to weight the physical distribution embedding vector, thereby obtaining the cross-modal fusion representation.

[0164] Based on the difference operation between the fused representation and the original semantic embedding vector, a semantic-measured deviation metric matrix is ​​constructed. The deviation value can be calculated using the following formula:

[0165] in, Let represent the semantic embedding vector of the i-th node. Let represent the physical measurement embedding vector of the j-th node, and D be the elements of the deviation matrix, which are calculated by taking the square root of the square difference to obtain the deviation value in Euclidean distance form.

[0166] The elements in the deviation matrix are sorted according to the topological position of the nodes and the context of causal edges to form a multidimensional deviation value sequence between path semantics and measured data distribution, providing a numerical basis for subsequent selection of the maximum deviation contribution input.

[0167] Through the aforementioned cross-modal attention alignment mechanism and deviation matrix construction, the bidirectional map reconstruction dataset from the previous step is transformed into a quantifiable indicator of the difference between the theoretical expected model and the actual operating state in key process links, thereby enabling the identification and preparation of links with significant abnormal deviations.

[0168] S7.5: Perform extreme value search and attribution mapping on the multidimensional deviation value sequence, use the maximum deviation contribution screening algorithm to identify the link node where the measured value deviates most significantly from the expected model, and generate the dominant anomaly factor location result containing the specific process parameter name and deviation magnitude, so as to complete the final logical closed loop from data deviation to responsibility factor.

[0169] A row-level extreme value search operation is performed on each deviation vector in the multidimensional deviation value sequence. Based on the preset deviation threshold judgment criteria, the deviation amplitude corresponding to each node in the deviation matrix is ​​compared one by one to extract the set of candidate nodes whose deviation amplitude exceeds the threshold.

[0170] The deviation magnitude of each node in the candidate node set is calculated by multiplying it by the weight of its causal edge in the graph, forming a sequence of deviation contribution indicators, which is used to comprehensively consider the combined effect of the absolute value of the deviation and the strength of the causal relationship.

[0171] Perform maximum value search processing on the deviation contribution index sequence to locate the target node with the highest contribution index value, and extract the process parameter name and absolute value of deviation associated with the node to form the positioning object data unit.

[0172] The data unit of the location object is mapped to a structured record of the dominant anomaly factor containing the specific process parameter name and deviation level. The factor identification field and parameter quantification field are filled to ensure that the record meets the format specifications of the subsequent report generation.

[0173] By using extreme value search and attribution mapping, the path deviation analysis results from the previous step are transformed into dominant anomaly factor location data that can be directly used for responsibility tracing, thereby achieving the expected technical effects of accurate identification of abnormal links and construction of causal closed loops.

[0174] For example, during a pipeline hot-melt welding operation, the multidimensional deviation value sequence generated by cross-modal attention alignment computation contains deviation indices for 16 nodes, each ranging from 0.05 to 1.37, with a preset deviation threshold of 0.80. When performing an extreme value search on the deviation matrix, five candidate nodes are selected, each with a deviation exceeding 0.80 and a corresponding causal edge weight between 0.4 and 0.9. The deviation contribution index is calculated for each candidate node using the following formula:

[0175] Where Δ is the absolute value of the node deviation, and w is the weight coefficient of the node in the causal edge. For example, if a node has Δ of 1.12 and w of 0.85, the contribution calculation result is 0.952. By searching for the maximum contribution, a link node is located, whose process parameter name is heating temperature, and the deviation magnitude is 1.12℃. This result is used as the dominant anomaly factor location output, mapped to the heating temperature control requirement clause in the construction specification, and used for subsequent report generation. In this embodiment, the dominant factor location result enables the traceability report to significantly improve legal credibility and location accuracy. The system demonstrates stable anomaly factor identification capability in the verification of multiple batches of construction data from multiple devices.

[0176] Step S8: Based on the location result of the dominant anomaly factor, map it to the construction specification clause number and display the clustering results of the frequency of occurrence and recurrence conditions of similar factors with the same or similar process characteristics as the dominant anomaly factor within a preset time period, generating a welding anomaly attribution and tracing report with legal credibility. Specifically, this includes: S8.1: Obtain the anomaly entity identifier and deviation quantification value from the dominant anomaly factor localization result, and perform semantic matching processing using the pre-built pipeline hot-melt welding construction specification knowledge mapping table to generate a construction specification violation mapping set containing specific violation clause numbers and corresponding technical standard descriptions.

[0177] The input conditions are the anomaly entity identifiers and deviation quantification values ​​contained in the dominant anomaly factor localization results. The anomaly entity identifiers originate from key process parameter nodes identified by the graph neural network model, and the deviation quantification value is the numerical difference between the measured value and the expected model value of that node. Both have a unique associated field for matching with the construction specification mapping table. For the anomaly entity identifier field, field standardization processing is performed, unifying the encoding format and byte order structure to ensure consistency with the index keys of the construction specification knowledge mapping table. After identifier standardization, a semantic parsing engine is invoked to perform semantic word segmentation and ontology concept mapping on the anomaly entity identifiers, converting them into a standardized semantic tag set. Each tag in this set represents a process element or quality control requirement defined in the construction specification. Based on the deviation quantification value, a threshold grading algorithm is applied to determine the level range of the deviation, mapping the numerical distribution to level tags such as "minor deviation," "moderate deviation," or "serious deviation" to enhance the matching degree with the violation level clauses in the construction specification. A standardized set of semantic tags and violation level tags are combined into a joint matching key. A multi-dimensional matching operation is performed between the constructed key-value matching vector and the clause number field in the construction specification knowledge mapping table to obtain a candidate set of violation clause numbers that satisfy both semantic consistency and level correspondence. Using this candidate set of clause numbers, the technical standard description field in the construction specification knowledge mapping table is retrieved, and the clause number and technical standard description are combined into a structured record unit, outputting a construction specification violation mapping set. Through a combined processing method of semantic parsing, level mapping, and multi-dimensional matching, the anomaly factor localization results from the previous step are transformed into violation clause numbers and standard description data that can be directly used for report generation, achieving accurate matching of violation clauses and establishment of legal citation basis.

[0178] For example, the abnormal entity identifier "Heating Temperature Deviation" displayed in the dominant anomaly factor localization results has a deviation quantification value of 12.6℃. During the input phase, field standardization is performed, unifying the original identifier encoding to UTF-8 and removing redundant symbols to ensure consistency with the "Heating Temperature" related entry index in the construction specification mapping table. The semantic parsing engine maps "Heating Temperature Deviation" to the ontology concept tag "Heating Process Temperature Control" in the construction specification knowledge base. The deviation quantification value is graded using a preset threshold grading algorithm. The threshold grading rules are: slight deviation range 5℃, moderate deviation range 5℃ ≤ deviation < 10℃, and severe deviation range ≥ 10℃. In this case, the deviation is 12.6℃, so the grade tag is "Severe Deviation". The semantic tag "Heating Process Temperature Control" and the grade tag "Severe Deviation" are combined to form a joint matching key. Multi-dimensional matching is performed using a hash index and the construction specification mapping table to obtain the violation clause number "HT-TEMP-003", corresponding to the technical standard description "The temperature during the welding heating stage shall not deviate from the set value by more than ±8℃". Finally, a violation mapping set record unit was constructed, including fields: violation clause number, technical standard description, deviation level, and actual deviation value. It was verified that this record can be directly used for legal credibility traceability reports, with accurate matching and citation capability.

[0179] S8.2: Based on the abnormal entity identifiers in the construction specification violation mapping set, retrieve the full sequence of welding process parameters within a preset time period from the off-chain distributed database, and use a multi-dimensional feature clustering algorithm to calculate the working condition similarity of the retrieved historical data to generate clusters of similar factor recurrence conditions.

[0180] S8.3: For each data cluster in the clustering results of the same factor recurrence conditions, the time frequency distribution of its occurrence, the identification of the operation team involved and the equipment operation status label are statistically analyzed, and multi-dimensional aggregation analysis is performed to generate a statistical profile of the historical occurrence frequency of the same factor.

[0181] S8.4: Integrate the aforementioned set of construction specification violation mappings, clustering results of similar factor recurrence conditions, and statistical profiles of the historical frequency of similar factors, and call the legal credibility report template engine to perform structured document assembly and causal logic chain visualization rendering to generate a welding anomaly attribution and tracing report with legal credibility.

[0182] In one embodiment, a pipeline hot-melt welded joint construction process data recording and traceability system is also included, which is used to execute the pipeline hot-melt welded joint construction process data recording and traceability method described in any of the above embodiments.

[0183] 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.

[0184] 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 element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its 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.

[0185] 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 recording and tracing data during the construction process of pipe hot-melt welded joints, characterized in that, include: S1: Acquire multi-source heterogeneous raw data streams from a sensor set during the pipeline hot-melt welding process, and preprocess the multi-source heterogeneous raw data streams to generate a welding process parameter sequence; S2: Call the pre-trained knowledge graph of the pipeline hot melt welding domain to initialize the model and map the welding process parameter sequence into a set of process-quality related entities; S3: Using the typical causal rule library extracted from the historical qualified and abnormal weld sample dataset, the connection relationship between each entity in the process-quality related entity set is dynamically activated to calculate the intensity of the connection relationship, and a dynamic causal graph model representing the evolution trajectory of the confidence of each causal edge under the current welding state is generated. S4: Based on the dynamic causal graph model, select the highest weight causal edge with a confidence level higher than a preset threshold, extract the directed subgraph encoding, and generate an abnormal attribution path snapshot; S5: Perform a hash operation on the abnormal attribution path snapshot to generate a data digest, and package the data digest together with the spatiotemporal identity into a minimal transaction unit to generate a blockchain evidence record; S6: Read the blockchain evidence record of the specified weld joint and decode and restore the causal path structure, and pull the original parameter fragments corresponding to the nodes involved in the causal path structure to generate a bidirectional graph reconstruction dataset; S7: Based on the bidirectional map reconstruction dataset, calculate the deviation between the path semantics and the measured data distribution, and generate the localization result of the dominant anomaly factor.

2. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 1, characterized in that, Step S3 specifically includes: Based on historical qualified and abnormal weld sample datasets, reverse engineering feature mining is performed to extract statistical correlation matrices between several key process variables and generate a typical causal rule library. Using the rule antecedent conditions in the typical causal rule base, the real-time monitoring values ​​of each entity in the process-quality related entity set are processed to determine the matching degree, and an initial set of activated causal edges is generated. Based on the real-time input deviation amplitude and duration window parameters of each causal edge in the initial set of activated causal edges, a weighted cumulative activation intensity calculation process is performed within the sliding time window to generate an instantaneous activation intensity sequence. Based on the historical attenuation coefficient of the instantaneous activation intensity sequence and the environmental interference factor of the current welding stage, the instantaneous activation intensity sequence is subjected to exponential smoothing confidence evolution extrapolation processing to generate dynamic confidence evolution trajectory data for each causal edge. By integrating the node attributes of the process-quality related entity set with the dynamic confidence evolution trajectory data, dynamic graph topology reconstruction processing is performed to form the dynamic causal graph model.

3. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 2, characterized in that, The key process variables include: heating temperature deviation, pressure application duration fluctuation rate, and gradient of texture entropy change in the molten zone.

4. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 1, characterized in that, The directed subgraph encoding consists of a unique identifier for the starting node, a unique identifier for the ending node, a sequence of intermediate jump nodes, and confidence interval identifiers for each edge.

5. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 4, characterized in that, Step S4 specifically includes: Threshold filtering is performed on the confidence evolution trajectory of all causal edges in the dynamic causal graph model to select a set of candidate causal edges with confidence scores higher than a preset threshold. Based on the candidate causal edge set, a directed path search algorithm is executed to calculate the total weight of the connected path from the process parameter starting node to the quality defect ending node, and to generate the highest weight causal edge sequence arranged in descending order of weight. The topological structure parsing process is performed on each causal edge in the highest weight causal edge sequence to extract the unique identifier of the starting node, the unique identifier of the ending node, and the sequence of intermediate jump nodes, forming the original directed subgraph encoded data. Based on the original directed subgraph encoding data, the confidence intervals corresponding to each causal edge are discretized and mapped to generate an edge attribute label set containing high, medium and low confidence interval identifiers; The unique identifiers of the starting node, the unique identifiers of the ending node, the sequence of intermediate jump nodes, and the set of edge attribute labels are serialized and encapsulated to generate a snapshot of the abnormal attribution path of this welding process.

6. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 4, characterized in that, The spatiotemporal identity identifier includes: a timestamp, a unique identification code for the welding equipment, and a biometric binding code for the operator; Step S5 specifically includes: The unique identifiers of the starting node, the unique identifiers of the ending node, the sequence of intermediate jump nodes, and the confidence interval identifiers of each edge are serialized and encoded to generate the original data stream to be hashed. Based on the original data stream to be hashed, a one-way hash transformation operation is performed to generate a fixed-length data digest fingerprint. The data digest fingerprint, timestamp, welding equipment unique identification code, and operator biometric binding code are aggregated and encapsulated into a multi-field process to construct a minimal transaction unit load; Based on the minimized transaction unit payload, consensus protocol verification and digital signature appending operations are performed to generate a standardized blockchain evidence storage transaction package. The standardized blockchain evidence storage transaction package is processed with on-chain submission and block confirmation listening to generate the blockchain evidence storage record.

7. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 1, characterized in that, After step S7, the following is included: S8: Based on the location result of the dominant anomaly factor, map it to the construction specification clause number and associate and display the clustering results of the frequency of occurrence and recurrence conditions of similar factors with the same or similar process characteristics as the dominant anomaly factor within a preset time period, and generate a welding anomaly attribution and tracing report.

8. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 1, characterized in that, The sensor set includes: a heating temperature sensor, a pressure application duration timer, an ambient humidity detector, and a multi-source heterogeneous raw data stream from an infrared imaging device for the molten zone; the process-quality associated entity set includes: a heating temperature node, a pressure application duration node, an ambient humidity node, and a texture entropy value node for the infrared image of the molten zone.

9. The method for recording and tracing data during the construction process of a pipeline hot-melt welded joint according to claim 1, characterized in that, Preprocessing the multi-source heterogeneous raw data stream includes: performing timestamp alignment and noise filtering on the multi-source heterogeneous raw data stream.

10. A data recording and traceability system for the construction process of pipeline hot-melt welded joints, characterized in that, This method is used to record and trace data during the construction process of a pipeline hot-melt welded joint, as described in any one of claims 1-9.

Images