A multi-source data fusion welding process tracing method

Abstract

The application relates to the technical field of welding process tracing, and discloses a welding process tracing method based on multi-source data fusion; first, welding multi-source data is collected and data cleaning is performed, entity types and relation types are set based on a rule base knowledge extraction method, and attribute types are matched to output welding process triplets; second, process parameter correlation mining is performed, linear correlation coefficients between process parameters are calculated through correlation analysis, welding process control relations are identified, quality relation weights are calculated, and a welding process knowledge graph is established; then, a welding process quality prediction model is established based on a full connection neural network, welding process quality is predicted, and welding process quality prediction results are output; finally, a multi-path fusion and sorting algorithm is used to determine a tracing priority and find a reverse tracing path; the welding multi-source data is fused through the knowledge graph, the purpose of welding process tracing is achieved, and the method is objective and accurate.

Description

Technical Field

[0001] This invention relates to the technical field of welding process traceability, specifically to a welding process traceability method based on multi-source data fusion. Background Technology

[0002] Multi-source data fusion technology significantly enhances perception and decision-making capabilities in industrial scenarios by collecting, correlating, and comprehensively analyzing heterogeneous data from different sensors, systems, and processes. In the field of welding process traceability, this technology integrates multi-dimensional data streams such as welding current and voltage waveforms, machine vision molten pool images, infrared thermal imaging temperature fields, gas flow and pressure values, and equipment status logs to construct a holographic digital mirror of the welding process. Currently, multi-source data fusion in welding traceability is still in its early stages of application: most systems only achieve basic data acquisition and archival storage (such as single parameter over-limit alarms), and have not yet developed in-depth data correlation and mining capabilities; cross-system and cross-level data barriers are prevalent, and time-series alignment and feature correlation technologies are still immature; real-time data processing and immediate quality control decisions still face challenges.

[0003] Traditional welding process traceability methods mainly rely on operators manually filling in welding voltage, current, speed and other parameters on the welding machine instrument readings. Quality inspectors independently record the location and type of welding process quality defects. They use barcode labels to link paper reports of welded parts in different processes (such as assembly, welding and flaw detection), and establish the basic process relationship by scanning the barcodes to obtain independent data.

[0004] Traditional welding process traceability methods store multi-source data such as process parameters, equipment logs, and quality inspection reports independently during the welding process. Spatiotemporal alignment requires manual matching, resulting in severe data silos. This makes it difficult to quantify and capture the core factors affecting welding quality, and makes it impossible to trace the root cause when quality problems occur. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a welding process traceability method based on multi-source data fusion, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution:

[0007] This invention relates to a welding process traceability method based on multi-source data fusion, comprising the following steps:

[0008] S1. Collect welding multi-source data and perform data cleaning. Based on the knowledge extraction method of the rule base, set entity type and relation type, introduce similarity coefficient to identify entities and relations, and then match attribute type to output welding process triplet.

[0009] S2. Based on the welding process triplet, perform process parameter correlation mining, calculate the linear correlation between process parameters through correlation analysis, identify the welding process control relationship, calculate the quality relationship weight, and establish a welding process knowledge graph.

[0010] S3. Establish a welding process quality prediction model based on a fully connected neural network, predict the welding process quality, and output the welding process quality prediction results.

[0011] S4. Set path search rules. Based on the welding process quality prediction results in the welding process knowledge graph, determine the tracing priority based on the multi-path fusion and sorting algorithm, find the reverse tracing path, and realize welding process tracing.

[0012] Preferably, the process of collecting welding multi-source data and performing data cleaning includes:

[0013] Collect welding multi-source data, and denote the text data in the welding multi-source data as the set of text data to be cleaned;

[0014] The text data set to be cleaned is subjected to unit unification processing and timestamp alignment. Then, a welding terminology database is set up for adaptive word segmentation, stop word filtering is performed, and entity type marking is performed to generate a processed welding multi-source data set.

[0015] Preferably, the introduction of similarity coefficients to identify entities and relationships includes:

[0016] The processed multi-source welding data set is used to set a relation type library based on the rule base to obtain relation types;

[0017] Based on the entity type, establish entity type combinations, obtain entity type attributes, and calculate the similarity coefficient of entity type attributes;

[0018] Set a similarity threshold. When the similarity coefficient is not greater than the similarity threshold, there is no relationship between the entity type combinations, and the corresponding entity type combination is deleted; otherwise, there is a relationship between the entity type combinations, the corresponding relationship type is marked between the entity type combinations, the entity and the relationship are identified, and a tuple is established.

[0019] Preferably, the rematching attribute type output welding process triplet includes:

[0020] Based on the entity type attributes in the entity type combination, attribute matching is performed on the entities in the binary tuple until the binary tuple is fully matched, thus forming a welding process triplet.

[0021] Preferably, the calculation of the linear correlation between process parameters through correlation analysis includes:

[0022] Based on the process parameter data corresponding to the entities in the welding process triplet, and after standardization, a standardized process parameter data set is obtained. The linear correlation coefficient between the process parameters is calculated through correlation analysis.

[0023] Preferably, the calculation of quality relationship weights to establish a welding process knowledge graph includes:

[0024] Set a correlation threshold. When the linear correlation coefficient is greater than the correlation threshold, the corresponding process parameter is recorded as positively correlated, forming a positively correlated process parameter pair. Otherwise, it is recorded as negatively correlated, forming a negatively correlated process parameter pair. Then, a positive control chain and a negative control chain are established in sequence to generate the welding process control relationship.

[0025] The entities in the welding process control relationship are used as knowledge graph nodes. The quality relationship weights are calculated. The relationships and attributes in the welding process control relationship are used as knowledge graph edges. The knowledge graph nodes are connected using knowledge graph edges, and the quality relationship weights are marked to establish a welding process knowledge graph.

[0026] Preferably, the welding process quality prediction model based on a fully connected neural network includes:

[0027] New weld samples were acquired, and multi-source welding data and environmental parameters were collected to form a multi-source data fusion tensor.

[0028] The welding process quality prediction model consists of an input layer, a hidden layer, and an output layer. The hidden layer uses a three-layer dense layer structure, denoted as the first dense layer, the second dense layer, and the third dense layer, respectively. The gradient norm is constrained. The first dense layer uses the ELU activation function and performs batch normalization. The dropout ratio is set and the data is transferred to the second dense layer. The second dense layer uses the Leaky ReLU activation function and the data is transferred to the third dense layer. The third dense layer uses the Swish activation function to establish the welding process quality prediction model.

[0029] Preferably, the prediction of welding process quality includes:

[0030] The multi-source data fusion tensor is divided into a sample training set and a sample test set, which are then input into the welding process quality prediction model for training to obtain the final welding process quality prediction model.

[0031] Based on the processed multi-source welding data set and combined with environmental parameters, the data is input into the final welding process quality prediction model, and the welding process quality prediction result is output.

[0032] Preferably, the determination of tracing priority based on the multi-path fusion and sorting algorithm includes:

[0033] Set welding process quality constraints. When the welding process quality prediction result exceeds the welding process quality constraints, the welding process traceability trigger condition is met. Set path search rules and generate first-level related knowledge graph nodes.

[0034] When the first-level related knowledge graph node is the root node, the tracing path is output; otherwise, the knowledge graph node is traced backward, the tracing path integral is calculated, and the tracing priority is obtained.

[0035] Preferably, the process of finding the reverse tracing path includes:

[0036] Select the tracing path corresponding to the highest tracing priority and denote it as the reverse tracing path. Record all knowledge graph edges and knowledge graph nodes on the reverse tracing path. Perform spatial mapping on the reverse tracing path in chronological order to obtain the position of the knowledge graph node, determine the time of occurrence of welding process quality defects, and realize welding process tracing.

[0037] The present invention has the following beneficial effects:

[0038] 1. This invention solves the problem of data silos in traditional welding by collecting multi-source welding data and cleaning the data to unify heterogeneous data into structured entities. Based on the knowledge extraction method of rule base, entity types and relationship types are set. At the same time, a similarity coefficient calculation and threshold judgment mechanism is introduced to match attribute types, which breaks through the limitations of human experience and realizes the objective diagnosis of process anomalies. The use of knowledge graph technology facilitates subsequent rapid retrieval and reasoning.

[0039] 2. This invention identifies welding process control relationships by performing process parameter correlation mining, using correlation analysis to calculate the linear correlation between process parameters, generating positive / reverse control chains to achieve adaptive parameter adjustment, and establishing a welding process knowledge graph. Through the chain reasoning capability of the knowledge graph, the traceability time of welding defects (such as porosity and lack of fusion) is greatly reduced compared with traditional manual analysis methods.

[0040] 3. This invention establishes a welding process quality prediction model based on a fully connected neural network to predict the welding process quality. It can effectively capture nonlinear process characteristics, ensure stable convergence of the model in dynamic data streams, and trigger a self-compensation mechanism, which greatly improves the accuracy of quality prediction. It achieves parallel high-precision prediction of three key indicators: penetration depth, porosity, and hardness, outputs welding process quality prediction results, and then feeds back into the welding process knowledge graph to form a continuously optimized intelligent decision-making closed loop.

[0041] 4. This invention, by setting path search rules and determining tracing priorities based on multi-path fusion and sorting algorithms, quickly identifies critical paths, reduces the time spent investigating welding process causes, automatically strengthens the weight of high-frequency fault paths, and accurately associates defects with the time / space nodes of equipment failures, thereby reducing the incidence of similar defects.

[0042] By finding a reverse tracing path, we can accurately trace the welding process and complete the closed-loop optimization of tracing.

[0043] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 This invention provides a flowchart illustrating a welding process tracing method based on multi-source data fusion.

[0046] Figure 2 This is a schematic diagram of the process for establishing a welding process quality prediction model provided by the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Traditional welding process traceability methods store multi-source data such as process parameters, equipment logs, and quality inspection reports independently during the welding process. Spatiotemporal alignment requires manual matching, resulting in severe data silos. This makes it difficult to quantify and capture the core factors affecting welding quality, and makes it impossible to trace the root cause when quality problems occur.

[0049] To address the aforementioned technical issues, a specific embodiment is an automated welding production line for LNG pipelines in an energy equipment company (annual output of 50km of welds). The welding equipment uses a Fronius TPS-500 welding machine (real-time monitoring of current / voltage / wire feed speed), environmental sensors are used to collect preheating temperature (±1℃ accuracy) and ambient humidity (±3%RH), and the quality inspection system uses an X-ray flaw detector (porosity detection accuracy of 0.1mm), providing a good foundation for the implementation of this invention.

[0050] In the specific implementation of the above embodiments, firstly, multi-source welding data is collected and cleaned. Based on the rule-based knowledge extraction method, entity types and relation types are set, and similarity coefficients are introduced to identify entities and relations. Then, attribute types are matched to output welding process triples. Timestamp alignment and adaptive word segmentation based on a terminology database are used to unify heterogeneous data into structured entities, solving the problem of isolated welding data in traditional methods. At the same time, a similarity coefficient calculation and threshold determination mechanism is introduced to overcome the limitations of manual experience and achieve objective diagnosis of process anomalies. Knowledge graph technology is used to facilitate rapid subsequent retrieval and reasoning. Secondly, process parameter association mining is performed. Through correlation analysis, the linear correlation between process parameters is calculated to identify welding process control relationships. Quality relationship weights are calculated to establish a welding process knowledge graph. This method uses linear correlation to generate positive / negative control chains to achieve adaptive parameter adjustment. Through the chain reasoning capability of the knowledge graph, compared with traditional manual analysis methods, it is significantly more efficient. This approach significantly reduces the time required to trace welding defects (such as porosity and lack of fusion). Furthermore, a welding process quality prediction model is established based on a fully connected neural network to predict welding process quality and output the prediction results. This model effectively captures nonlinear process characteristics, ensuring stable convergence in dynamic data streams and triggering a self-compensation mechanism, greatly improving the accuracy of quality prediction. It achieves parallel high-precision prediction of three key indicators: penetration depth, porosity, and hardness. The prediction results feed back into the welding process knowledge graph, forming a continuously optimized intelligent decision-making closed loop. Finally, path search rules are set, and tracing priorities are determined based on multi-path fusion and ranking algorithms. Reverse tracing paths are found, and critical paths are quickly locked through multi-path fusion and ranking algorithms, reducing the time for investigating welding process causes. The weight of high-frequency fault paths is automatically strengthened, and the temporal / spatial nodes of defects and equipment failures are accurately associated, reducing the incidence of similar defects and accurately achieving welding process tracing, thus completing the tracing closed-loop optimization.

[0051] Furthermore, to better illustrate the technical solutions of the embodiments of the present invention, such as... Figure 1 As shown, this embodiment of the invention provides a welding process traceability method based on multi-source data fusion, specifically including the following:

[0052] S1. Collect welding multi-source data and perform data cleaning. Based on the knowledge extraction method of the rule base, set entity type and relation type, introduce similarity coefficient to identify entities and relations, and then match attribute type to output welding process triplet.

[0053] S1 includes the following steps:

[0054] S11. Label different weld passes during the welding process to form weld pass IDs, and then collect welding multi-source data. The welding multi-source data includes process parameter data, process specification text, quality inspection report, equipment log, etc. Among them, the process parameter data includes welding current, arc voltage, wire feeding speed, etc. The text data in the welding multi-source data is recorded as the text data set to be cleaned.

[0055] The text data set to be cleaned is subjected to unit unification processing and timestamp alignment. Then, a welding terminology library is set up, and adaptive word segmentation is performed on the text data set to be cleaned. The text data is segmented and synonyms are merged to obtain the word-segmented text data set. Then, stop word filtering is performed and entity type marking is performed. The entity types include process parameters, equipment entities, material properties, etc., to generate a processed welding multi-source data set.

[0056] S12. For the processed welding multi-source data set, a relationship type library is set based on a rule base. The relationship type library includes causal influence, influence scope, equipment association, etc., to obtain relationship types. Then, entity type combinations are established according to the entity types, and a similarity coefficient is introduced to determine the relationship type between entity type combinations, thus identifying entities and relationships. The specific steps are as follows:

[0057] S121. Obtain the entity type attributes of the entity type combination. The entity type attributes include standard range, level, specific data, etc., and are respectively denoted as the first entity type attribute set and the second entity type attribute set. Calculate the intersection and union of the first entity type attribute set and the second entity type attribute set. Then, assign attribute weights to the intersection of the first entity type attribute set and the second entity type attribute set to calculate the similarity coefficient. ,in Represents the set of attributes of the first entity type. Represents the set of attributes of the second entity type. Represents the first entity type attribute set and the second entity type attribute set. a The attribute weights of the intersection of the elements;

[0058] S122. Set a similarity threshold. When the similarity coefficient is not greater than the similarity threshold, there is no relationship between the entity type combinations, and the corresponding entity type combination is deleted. Otherwise, there is a relationship between the entity type combinations. Mark the corresponding relationship between the entity type combinations, identify the entities and relationships, and establish a tuple.

[0059] S13. Based on the entity type attributes in the entity type combination, perform attribute matching on the entities in the binary tuples until the binary tuples are matched, forming a {entity, relation, attribute} triple, denoted as the welding process triple.

[0060] In this embodiment, welding data from multiple sources is collected and cleaned. A knowledge extraction method based on a rule base is used to set entity and relation types, and a similarity coefficient is introduced to identify entities and relations. Then, attribute types are matched to output welding process triples. Timestamp alignment and adaptive word segmentation based on a terminology database are used to unify heterogeneous data into structured entities, solving the problem of isolated welding data. Simultaneously, a similarity coefficient calculation and threshold determination mechanism is introduced to overcome the limitations of manual experience and achieve objective diagnosis of process anomalies. Knowledge graph technology is used to facilitate rapid subsequent retrieval and reasoning. Specifically, for example, a weld ID is marked for a stainless steel pipe welding task, and weld data is collected. Data from multiple sources, including process parameters: current 150A, voltage 22V, wire feed speed 12cm / min; process specification text: using ER308L welding wire, interpass temperature ≤150℃; quality inspection report: weld #W107 showed 0.8mm porosity; equipment log: abnormal vibration of the wire feeder (2025-05-10 14:22); unit standardization: 12cm / min → 120mm / min; original text: "interpass temperature needs to be strictly controlled"; after word segmentation: ["interpass temperature", "needs", "strictly", "control"]; synonym merging: "welding wire ER308L" → "filler material: E "R308L", stop word filtering: remove "need", "strict", "control", etc.; where, current 150A → process parameter, wire feeder → equipment entity, ER308L → material property; set relationship type library: causal influence (e.g., voltage fluctuation → porosity), equipment association (e.g., wire feeder → wire feeding speed), first entity type attribute set (welding current attribute set) and second entity type attribute set (penetration depth attribute set), where, welding current attribute set: {value, unit (A), stability}, penetration depth attribute set: {depth value, unit (mm), uniformity, current association}, intersection: {unit} → cardinality 1, union: {Numerical value, unit (A), stability, depth value, unit (mm), uniformity, current correlation} → base 7; set attribute weights according to domain importance, and calculate the similarity coefficient; acquire welding multi-source data generated from 500 welding operations, compare the similarity between different entity types, and find that when the similarity is 0.3, there is a correlation between entity types. Label the corresponding relationship type and perform attribute matching to form a triple {entity, relationship, attribute}, such as {voltage 22V, causal influence, abnormal fluctuation ±2V}, {wire feeder XYZ200, equipment correlation, rated wire feeding speed 130mm / min}, etc.

[0061] S2. Based on the welding process triplet, perform process parameter correlation mining, calculate the linear correlation between process parameters through correlation analysis, identify the welding process control relationship, calculate the quality relationship weight, and establish a welding process knowledge graph.

[0062] S2 includes the following steps:

[0063] S21. Based on the process parameter data corresponding to the entity in the welding process ternary set, and according to the weld ID...

[0064] Record the process parameter data sequentially to obtain an initial process parameter data set. Calculate the mean and range of the process parameter data for each weld pass. Subtract the mean of the process parameter data from the process parameter data and then divide by the range of the process parameter data to obtain a standardized process parameter data set.

[0065] Different types of process parameter data are selected from the standardized process parameter data set, and denoted as welding current data set and arc voltage data set, respectively. The mean of welding current data set and the mean of arc voltage data set are calculated. The linear correlation coefficient between the process parameters is calculated through correlation analysis. ,in This indicates the first [number] in the welding current data set. i One welding current data point, This represents the first arc voltage data set. i One arc voltage data, This represents the mean of the welding current dataset. This represents the mean of the arc voltage data set. m Indicates the number of weld IDs. ;

[0066] S22. Calculate the linear correlation coefficient for all types of process parameter data in the standardized process parameter data set sequentially to generate a correlation matrix; set a correlation threshold, and when the linear correlation coefficient is greater than the correlation threshold, the corresponding process parameter is recorded as positively correlated, forming a positively correlated process parameter pair; otherwise, it is recorded as negatively correlated, forming a negatively correlated process parameter pair; establish a positive control chain between the positively correlated process parameter pairs and an inverse control chain between the negatively correlated process parameter pairs, converting the welding process triplet into a welding process control relationship;

[0067] S23. Based on the forward and reverse control chains in the welding process control relationship, calculate the quality relationship weights and establish a welding process knowledge graph. The specific steps are as follows:

[0068] S231. Take the entities in the welding process control relationship as knowledge graph nodes, arbitrarily select knowledge graph nodes as key knowledge graph nodes, find knowledge graph nodes that have positive and negative control chains with key knowledge graph nodes, and use random forest to calculate the quality relationship weight by aggregating the feature contributions of multiple decision trees.

[0069] S232. For the welding process triplet in the welding process control relationship, the relationship and attribute in the welding process control relationship are used as knowledge graph edges, and knowledge graph nodes are connected using knowledge graph edges; all welding process triplets are extracted in sequence, knowledge graph nodes are connected and quality relationship weights are marked to establish a welding process knowledge graph.

[0070] In this embodiment, process parameter correlation mining is performed. Linear correlation between process parameters is calculated through correlation analysis to identify welding process control relationships. Quality relationship weights are calculated to establish a welding process knowledge graph. This method uses linear correlation to generate positive / negative control chains to achieve adaptive parameter adjustment. Through the chain-like reasoning capability of the knowledge graph, the tracing time for welding defects (such as porosity and lack of fusion) is significantly reduced compared to traditional manual analysis methods. Specifically, for example, weld IDs are labeled: #W101-#W110 (a total of 10 welds), where the average current is 152A, the range is 20A, and the average voltage is... The average current is 23V, with a range of 5V. The average wire feed speed is 118mm / min, with a range of 30mm / min. Taking an average current of 152A and an average voltage of 23V as an example, the calculated linear correlation coefficient is -0.85. For welding multi-source data from 500 welding operations, the linear correlation coefficients between process parameters were calculated sequentially. Analysis showed that only when the linear correlation coefficient is greater than 0.8 does a correlation exist between process parameters. Therefore, a correlation threshold of 0.8 was set, leading to the conclusion that current and voltage are positively correlated process parameter pairs. Based on the linear correlation coefficients between current, voltage, and wire feed speed, a correlation matrix was established. A control chain is generated, where the forward control chain is: current increase → forward drive → wire feed speed increase, and the reverse control chain is: voltage increase → reverse suppression → current decrease, and wire feed speed increase → reverse suppression → voltage decrease. The contribution of voltage fluctuation features is calculated to be 0.38 and the contribution of wire feed anomaly features is 0.29, etc., using random forest weight calculation. A knowledge graph is established: nodes represent entities (current, voltage, wire feed speed, porosity defects), and edges represent control relationships and quality weights.

[0071] S3. Establish a welding process quality prediction model based on a fully connected neural network, predict the welding process quality, and output the welding process quality prediction results.

[0072] S3 includes the following steps:

[0073] S31. Acquire new weld samples, collect welding multi-source data and environmental parameters, including preheating temperature, ambient humidity, working temperature, etc., acquire time-series segments of the collected welding multi-source data, form a multi-source data fusion tensor, and set welding multi-source data constraints.

[0074] S32. Establish a welding process quality prediction model based on a fully connected neural network, train the model, predict the welding process quality based on the processed multi-source welding data set, and output the welding process quality prediction result, such as... Figure 2 The specific steps for establishing a welding process quality prediction model are as follows:

[0075] S321. The welding process quality prediction model is set to include an input layer, a hidden layer, and an output layer. The hidden layer uses a three-layer dense layer structure, which are respectively denoted as the first dense layer, the second dense layer, and the third dense layer. The gradient norm is restricted. The first dense layer adopts the ELU activation function and performs batch normalization. The Dropout ratio is set and the data is transmitted to the second dense layer. The second dense layer adopts the LeakyReLU activation function and the data is transmitted to the third dense layer. The third dense layer adopts the Swish activation function to establish the welding process quality prediction model.

[0076] S322. Divide the multi-source data fusion tensor into a sample training set and a sample test set. Input the sample training set into the welding process quality prediction model for training until the welding process quality prediction model converges to obtain a trained welding process quality prediction model. Then input the sample test set into the trained welding process quality prediction model, set an accuracy threshold, and stop iterating when the accuracy of the output result is greater than the accuracy threshold to obtain the final welding process quality prediction model. Otherwise, adjust the weights until the accuracy of the output result is greater than the accuracy threshold.

[0077] S323. Based on the processed welding multi-source data set and combined with environmental parameters, input the data into the final welding process quality prediction model and output the welding process quality prediction results. The welding process quality prediction results include the predicted value of welding penetration, the predicted value of porosity, and the predicted value of the highest hardness of the welding heat-affected zone.

[0078] In this embodiment, a welding process quality prediction model is established based on a fully connected neural network to predict the welding process quality and output the prediction results. This model can effectively capture nonlinear process characteristics, ensure stable convergence of the model in dynamic data streams, and trigger a self-compensation mechanism, greatly improving the accuracy of quality prediction. It achieves parallel high-precision prediction of three key indicators: penetration depth, porosity, and hardness. The prediction results feed back into the welding process knowledge graph, forming a continuously optimized intelligent decision-making closed loop. Specifically, for example, weld IDs: #W201-#W250 are selected to obtain multi-source welding data (current, voltage, wire feed speed) and environmental parameters (preheating temperature, ambient humidity). (Degree, working temperature), generate fusion tensor dimensions: [50 weld beads] × [120 time points] × [6 parameters], current fluctuation range: ±15A, humidity safety threshold: ≤70%RH, penetration depth qualification standard: ≥5.0mm; start training the model, training set: #W201-#W230 (30 weld beads), test set: #W231-#W250 (20 weld beads), to obtain the final welding process quality prediction model; for example, input the welding multi-source data and environmental parameters corresponding to weld bead #W107, output the predicted welding penetration depth value of 5.72mm, the predicted porosity value of 1.25%, and the predicted maximum hardness value of the welding heat-affected zone of 254HV;

[0079] S4. Set path search rules. Based on the welding process quality prediction results in the welding process knowledge graph, determine the tracing priority based on the multi-path fusion and sorting algorithm, find the reverse tracing path, and realize welding process tracing.

[0080] S4 includes the following steps:

[0081] S41. Based on the welding process quality prediction results, welding process quality constraints are set. When the predicted value of welding penetration, porosity, or maximum hardness of the weld heat-affected zone exceeds the welding process quality constraints, the welding process traceability triggering condition is met. Path search rules are set, dynamic weighted path search is performed, and the traceability path integral is calculated. The specific steps are as follows:

[0082] S411. After the welding process traceability triggering condition is met, the knowledge graph node corresponding to the welding process quality defect is determined in the welding process knowledge graph as the starting point of the traceability path; the maximum traceability level of the knowledge graph node is set, the weight filtering threshold is set, and the knowledge graph nodes with quality relationship weights greater than the weight filtering threshold are retained to obtain the first-level associated knowledge graph nodes.

[0083] S412. When the first-level associated knowledge graph node is the root node, the tracing path is output. Otherwise, the knowledge graph node is traced backward to determine whether the associated knowledge graph node is the root node, until the maximum tracing level of the knowledge graph node is reached, the total number of tracing paths is determined, and several tracing paths are obtained to form a tracing path set. At this time, the product of the quality relationship weights of all associated knowledge graph nodes on the tracing path is calculated to obtain the tracing path integral.

[0084] S42. Find the traceability path integral corresponding to the traceability path in the traceability path set, calculate the data integrity rate of the traceability path in the traceability path set, and calculate the time from the starting point of the traceability path to the end point of the traceability path, which is recorded as the timeliness. Assign weights to the traceability path integral, data integrity rate and timeliness, and calculate the total traceability path integral. Sort the traceability paths in descending order according to the total traceability path integral, and then assign them to the traceability priorities. Select the traceability path corresponding to the highest traceability priority, which is recorded as the reverse traceability path. Record all knowledge graph edges and knowledge graph nodes on the reverse traceability path. Spatially map the reverse traceability path according to the time order to obtain the position of the knowledge graph node, determine the time of occurrence of welding process quality defects, and realize welding process traceability.

[0085] In this embodiment, path search rules are set, and the tracing priority is determined based on the multi-path fusion and sorting algorithm. Reverse tracing paths are found, and the critical path is quickly locked through the multi-path fusion and sorting algorithm, reducing the time for investigating welding process causes. The weight of high-frequency fault paths is automatically strengthened, and the time / space nodes of defects and equipment failures are accurately associated, reducing the incidence of similar defects and accurately achieving welding process tracing. Specifically, for example, the predicted result for weld #W300 is: porosity = 3.5% > constraint threshold (2.0%). The defect node is located as a porosity defect node in the knowledge graph. QP300 is used as the starting point of the tracing path, and the maximum tracing level is set to 3 to limit the search depth. The weight filtering threshold is 0.5 to retain key influencing paths. A set of paths is generated: Path 1: QP300 ← [weight filtering threshold 0.76] voltage fluctuation (level 1), Path 2: QP300 ← [weight filtering threshold 0.76] voltage fluctuation ← abnormal wire feed speed (level 2), Path 3: QP300 ← [weight filtering threshold 0.76] voltage fluctuation ← current fluctuation (level 2). Level), Path 4: QP300 ← [weighted filtering threshold 0.76] voltage fluctuation ← abnormal wire feeding speed ← [weighted filtering threshold 0.58] wire feeder vibration (Level 3), calculate the traceability path integral in sequence (for example, the integral of path 1: 0.76, the integral of path 4: 0.76×1.0×0.58=0.441, the default is 1.0 when there is no weighted filtering threshold); multi-path fusion sorting: path 1: data integrity rate 100%, timeliness (hours) 0.8, path 4: data integrity rate 85%, timeliness (hours) (Hours) 3.2, assigning weights to traceability path integral, data integrity rate, and timeliness proximity. Based on the multi-source welding data generated from 500 welding operations, a new knowledge graph is established. The traceability time of each path in the new knowledge graph is analyzed, and the influence of traceability path integral, data integrity rate, and timeliness proximity on path traceability time is quantified, with weights of 0.6, 0.2, and 0.2, respectively. At this time, the total integral of path 1 is 0.906 (highest priority), the root cause is located: voltage fluctuation, the spatial location is obtained, and welding process traceability is realized.

[0086] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0087] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A multi-source data fusion welding process traceability method, characterized by, The method comprises the following steps: S1, collecting welding multi-source data and performing data cleaning, performing knowledge extraction based on a rule base, setting entity types and relationship types, introducing a similarity coefficient to identify entities and relationships, and matching attribute types to output welding process triples; S2, according to the welding process triples, process parameter association mining is performed, linear correlation between process parameters is calculated through correlation analysis, welding process control relationships are identified, quality relationship weights are calculated, and a welding process knowledge graph is established; S3, a welding process quality prediction model is established based on a full connection neural network, welding process quality is predicted, and a welding process quality prediction result is output; S4, path search rules are set, a welding process quality prediction result is determined in the welding process knowledge graph, a multi-path fusion and sorting algorithm is used to determine a traceability priority, a reverse traceability path is found, and welding process traceability is realized; The welding process quality prediction result is determined in the welding process knowledge graph, a multi-path fusion and sorting algorithm is used to determine a traceability priority, a reverse traceability path is found, and welding process traceability is realized; The welding process quality prediction result is determined in the welding process knowledge graph, a multi-path fusion and sorting algorithm is used to determine a traceability priority, a reverse traceability path is found, and welding process traceability is realized; The welding process quality prediction result is determined in the welding process knowledge graph, a multi-path fusion and sorting algorithm is used to determine a traceability priority, a reverse traceability path is found, and welding process traceability is realized; The welding process quality prediction result is determined in the welding process knowledge graph, a multi-path fusion and sorting algorithm is used to determine a traceability priority, a reverse traceability path is found, and welding process traceability is realized; S41, according to the welding process quality prediction result, a welding process quality constraint is set, when a welding penetration prediction value or a porosity prediction value or a welding heat affected zone maximum hardness prediction value exceeds the welding process quality constraint, a welding process traceability trigger condition is met, path search rules are set, a dynamic weight path search is performed, and traceability path scores are calculated; the specific steps are as follows: S411, after the welding process traceability trigger condition is met, welding process quality defect corresponding knowledge graph nodes are determined in the welding process knowledge graph as traceability path starting points; a maximum traceability number of knowledge graph nodes is set, a weight filtering threshold is set, knowledge graph node edges with quality relationship weights greater than the weight filtering threshold are retained, and primary associated knowledge graph nodes are obtained; S412, when the primary associated knowledge graph nodes are root nodes, the traceability path is output, otherwise, the knowledge graph nodes are continuously traced back, it is judged whether the associated knowledge graph nodes are root nodes, until the maximum traceability number of knowledge graph nodes is reached, the traceability path total score is determined, a plurality of traceability paths are obtained, and a traceability path set is formed; at this time, the quality relationship weight products of all associated knowledge graph nodes on the traceability path are calculated, and the traceability path scores are obtained; ​ S42, find the trace path integral corresponding to the trace path in the set of trace paths, count the data integrity rate of the trace path in the set of trace paths, and calculate the time from the trace path starting point to the trace path key point, denoted as time efficiency closeness; give the trace path integral, data integrity rate and time efficiency closeness a weight, and calculate the total trace path integral; according to the descending order of the total trace path integral, the trace priority is correspondingly arranged in sequence; the trace path corresponding to the highest trace priority is selected, denoted as the reverse trace path, all knowledge graph edges and knowledge graph nodes on the reverse trace path are recorded, the reverse trace path is spatially mapped in time sequence, the position of the knowledge graph node is obtained, the welding process quality defect occurrence time is determined, and the welding process trace is realized.

2. The multi-source data fusion based welding process traceability method of claim 1, wherein, The collection of welding multi-source data and data cleaning comprises: Collecting welding multi-source data, the text data in the welding multi-source data is denoted as a set of to-be-cleaned text data; The set of to-be-cleaned text data is uniformly processed, and the time stamp is aligned, then the welding terminology library is set for adaptive word segmentation, stop word filtering is performed, and entity type marking is performed to generate a set of processed welding multi-source data.

3. The multi-source data fusion welding process traceability method of claim 2, wherein, The introduction of similarity coefficient to identify entities and relationships comprises: The set of processed welding multi-source data is used to set a relationship type library based on a rule library to obtain a relationship type; According to the entity type, an entity type combination is established, an entity type attribute is obtained, and a similarity coefficient of the entity type attribute is calculated; A similarity threshold is set, when the similarity coefficient is not greater than the similarity threshold, there is no relationship type between the entity type combinations at this time, and the corresponding entity type combination is deleted; otherwise, there is a relationship type between the entity type combinations, the corresponding relationship type is marked between the entity type combinations, the entities and relationships are identified, and a binary tuple is established.

4. The multi-source data fusion welding process traceability method of claim 3, wherein, The matching attribute type output welding process triple comprises: According to the entity type attribute in the entity type combination, the attributes of the entities in the binary tuple are matched until the binary tuple is matched, and a welding process triple is formed.

5. The multi-source data fusion welding process traceability method of claim 4, wherein, The correlation analysis to calculate the linear correlation between process parameters comprises: According to the process parameter data corresponding to the entities in the welding process triple, and standardizing processing, a set of standardized process parameter data is obtained, and the linear correlation coefficient between the process parameters is calculated through correlation analysis.

6. The multi-source data fusion welding process traceability method of claim 5, wherein, The calculation of quality relationship weight to establish the welding process knowledge graph comprises: A correlation threshold is set, when the linear correlation coefficient is greater than the correlation threshold, the corresponding process parameter is recorded as a positive correlation at this time, a positive correlation process parameter pair is formed, otherwise a negative correlation process parameter pair is formed, and a forward control chain and a reverse control chain are established in sequence to generate a welding process control relationship; The entities in the welding process control relationship are used as knowledge graph nodes, the quality relationship weight is calculated, the relationships and attributes in the welding process control relationship are used as knowledge graph edges, the knowledge graph nodes are connected using the knowledge graph edges, and the quality relationship weight is marked to establish the welding process knowledge graph.

7. The multi-source data fusion welding process traceability method of claim 6, wherein, The welding process quality prediction model based on a full connection neural network comprises: A new welding bead sample is reacquired, welding multi-source data and environmental parameters are collected, and a multi-source data fusion tensor is formed; The welding process quality prediction model comprises an input layer, a hidden layer and an output layer, the hidden layer uses a three-layer dense layer structure, and is respectively denoted as a first dense layer, a second dense layer and a third dense layer, the gradient norm is limited, the first dense layer adopts an ELU activation function, batch normalization is performed, a Dropout ratio is set, and the first dense layer is transmitted to the second dense layer, the second dense layer adopts a Leaky Relu activation function, and is transmitted to the third dense layer, the third dense layer adopts a Swish activation function, and the welding process quality prediction model is established.

8. The multi-source data fusion welding process traceability method of claim 7, wherein, The welding process quality is predicted by: The multi-source data fusion tensor is divided into a sample training set and a sample test set, is input into the welding process quality prediction model for training, and a final welding process quality prediction model is obtained; According to the processed welding multi-source data set, in combination with the environmental parameters, the final welding process quality prediction model is input, and a welding process quality prediction result is output.

Images