A root cause analysis method for stud welding defects based on knowledge graph
Through the knowledge graph, the multi-level root cause map is constructed and machine learning technology is combined, the problem of root cause analysis in the stud welding process is solved, efficient root cause positioning and analysis is achieved, the interpretability of the analysis is enhanced, and the traditional database is overcome. Data analysis difficulties without topology.
Patent Information
- Application Number
- CN202210460684.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-28
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-04-28
AI Technical Summary
In complex industrial scenarios, the root cause analysis of the root cause of stud welding process is difficult and costly, and it is difficult for the existing technology to effectively conduct root cause analysis, especially when the data has no topological structure.
Using a knowledge graph-based method, a multi-level root cause map is constructed by organizing stud welding process data, combining event extraction algorithms, machine learning and deep learning, analyzing and optimizing the root cause links is carried out to enhance the interpretability of the analysis.
The root cause analysis under the data topology is realized, which improves the accuracy and interpretability of stud welding defect analysis, reduces the burden on traditional relational databases, and helps engineers quickly locate and analyze the root cause of stud welding defects.
Smart Images

Figure CN114936755B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer software technology, and in particular to a stud welding defect root cause analysis method based on knowledge graph. Background Art
[0002] Compared with traditional databases, knowledge graphs can not only store large amounts of data, but also can well describe the image of data objects and the relationship network between data objects. In particular, they provide comprehensive and multi-dimensional data support in the fields of knowledge retrieval, question-answering systems, e-commerce systems, etc., playing a huge role.
[0003] However, the root cause analysis of faults is often concentrated in large Internet systems, which have sufficient alarm information and massive logs. The quality of the logs can largely reflect the current defect performance. However, in complex industrial scenarios, the amount of information contained in the alarm log is not enough to support root cause analysis. It is necessary to combine multi-dimensional data and conduct root cause analysis based on statistical or machine learning algorithms.
[0004] There are few studies on the root cause analysis of stud welding process quality problems. Due to the complexity of the stud welding process, the excessive number of influencing factors involved, and the lack of algorithm engineers with in-depth understanding of the business, the root cause analysis of stud welding problems is difficult and the research cost is high. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a root cause analysis method for stud welding defects based on knowledge graph, which overcomes the shortcomings of the prior art and solves the problem of performing root cause analysis when the data has no topological structure. At the same time, with the help of knowledge graph, based on the designed graph structure, the data of stud welding is reasonably organized, and combined with various algorithms related to root cause analysis, while accurately performing root cause analysis of stud welding defects, the basis for root cause analysis judgment is displayed, thereby enhancing the interpretability of root cause analysis.
[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0007] A method for analyzing root causes of stud welding defects based on knowledge graph, comprising the following steps:
[0008] Step S1: According to the data generated by the actual production of the stud welding process, the data is organized according to the structural relationship combining time and space to establish a complete data relationship;
[0009] Step S2: Based on the existing expert experience database data, stud welding process principles and relevant case data information, various defects in the stud welding process are sorted out, and a multi-level root cause map of stud welding defects is obtained;
[0010] Step S3: Combined with the problem description of the new defect entered on the mobile phone, the corresponding root cause chain and optimization solution chain are extracted from the multi-level root cause map in step S2 using the event extraction algorithm. After confirmation, they are deposited into the multi-level root cause map again.
[0011] Step S4: Initialize different probabilities on each chain propagation chain based on the statistical results of the historical occurrence frequency of the root cause link; and fine-tune all links based on the statistics of the recent occurrence frequency of stud welding defects of different studs;
[0012] Step S5: predicting the current stud welding defect according to the composite parameter layer in the multi-level root cause map, and performing statistics and scoring;
[0013] Step S6: predicting the current stud welding defect according to the basic parameter layer in the multi-level root cause map, and performing statistics and scoring;
[0014] Step S7: For the working condition equipment layer, using the deep learning method, from the data level, learn the different impacts of different working conditions on the data, and perform multi-classification to calculate the probability of different stud welding defects in multiple working conditions;
[0015] Step S8: After the root cause scoring of each layer is completed, the initial score of the root cause chain is calculated by accumulating the scores of each node in the chain, and finally the scores of each chain are normalized and used as the final score;
[0016] Step S9: Through the data and the finally confirmed root cause chain, the model and statistical results are updated regularly through continuous accumulation.
[0017] Preferably, the multi-level root cause map constructed in step S2 includes: a first parameter composite layer, a second basic parameter layer, and a third operating condition equipment layer.
[0018] Preferably, for the second basic parameter layer, a classification model for different defects is used with the help of machine learning, and based on the weight of each dimension of data in the model, the importance and relevance of the data are analyzed.
[0019] Preferably, the step S3 specifically includes the following steps:
[0020] Step S31: Record the root cause analysis of the on-site input, and extract multiple subjects, predicates, and objects from the description by designing an event extraction algorithm, and finally form multiple causal events or sequential events. The subject, predicate, and object form the event in the following way:
[0021] (Subject, Predicate, Object) 1 =Event i
[0022] The form is as follows:
[0023] event 1 -[Cause and Effect / Sequence]->Event 2 ->…->Events n ;
[0024] Step S32: Through similarity model training, the current extracted event can be calculated i All events in the root cause graph j The similarity S ij , set the threshold x = 0.80, if S ij ≥x, then judge the event i With events j It’s the same event;
[0025] Step S33: If there is a matching result for the event in step S32, then repeat step S32;
[0026] Step S34: If step S33 is completed and S is not satisfied ij Events ≥ x i , then the event i and its subsequent events, spliced into the last one in the root cause graph that satisfies S ij Events ≥ x j Later, and integrated into the atlas.
[0027] Preferably, in step S5, based on the big data threshold statistical results, the probability of different root cause propagation paths is determined by whether the actual value crosses the threshold and the degree of proximity between the actual value and the threshold; the specific threshold statistical method is as follows:
[0028] Step S51: Calculate the average heat value in the historical data of a certain soldering point
[0029] Step S52: Calculate the standard deviation δ of heat in the historical data of a certain soldering point;
[0030] Step S53: Set greater than or less than The values of are considered outliers and scored as 100%.
[0031] Preferably, in step S6, based on the statistical results of the big data parameter space, the probability of different root cause propagation paths is determined through the distribution set of the actual value in the parameter space; the specific parameter space statistical method is as follows:
[0032] Step S61: The arc striking voltage in the data is used as the evaluation value, and the data of other dimensions are reduced to 2 dimensions by the PCA method, and their distribution in space is counted and divided into different areas. The central 70% of the data is area A, the area of 70%-80% is area B, the area of 80%-90% is area C, and the area above 90% is area D;
[0033] Step S62: Area A, Area B, Area C, and Area D represent normal area, sub-abnormal area, abnormal area, and potential problem area, respectively, and the corresponding multi-dimensional data of the original data are normal parameter range, sub-abnormal parameter range, abnormal parameter range, and potential problem parameter range. At the same time, the multi-dimensional data of each stud welding defect is scored.
[0034] Preferably, in step S8, based on the trained decision tree model, the decision boundaries of the decision tree classification of each dimensional data are obtained, and finally, by characterizing the degree of deviation between the actual value and the corresponding decision boundary, the corresponding node in the corresponding root cause graph is calculated, and the weight of the node is calculated.
[0035] The present invention provides a method for analyzing the root causes of stud welding defects based on a knowledge graph. It has the following beneficial effects: The method proposed in this method describes the image of stud welding based on a specific structure based on a knowledge graph, which can better serve information retrieval while analyzing the root causes of stud welding defects. It greatly reduces the difficulty of mapping in traditional relational databases and facilitates construction and query. It can help stud welding engineers quickly analyze the root causes of stud welding defects and provide strong persuasiveness and explainability. By using machine learning models and related statistical methods, bound to the corresponding root cause nodes, when the data is stored in the graph, the root cause analysis can be automatically performed, and with the help of a comprehensive scoring strategy, the accuracy of the root cause analysis can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the present invention or the technical solutions in the prior art, the drawings required for describing the prior art are briefly introduced below.
[0037] Figure 1 A flowchart of the steps of the present invention;
[0038] Figure 2 An organizational diagram of the data in step S1 of the present invention. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention.
[0040] like Figure 1 to Figure 2 As shown, a method for root cause analysis of stud welding defects based on knowledge graph includes the following steps:
[0041] Step S1: Based on the data generated by the actual production of the stud welding process, the data is organized according to the structural relationship of time and space, and the ontology of the stud welding process is designed; the data organized according to the ontology design mainly provides four functions: stud portrait description, information retrieval, algorithm data support and enhancing the interpretability of root cause analysis results;
[0042] The data generated include: arc starting voltage, main welding voltage, main welding current, main welding time, penetration depth, lifting height, energy, stroke height, penetration time, alarm log and other information;
[0043] Image description: The data related to the stud welding process at different times and positions are expanded around the stud as the image description of the stud at the current time and current position. The image description of the stud can well describe the relevant detailed information of the stud.
[0044] Information retrieval: Organizing data in a time-space manner helps to query stud-related information. Through data stud number, workstation, robot, etc., according to different query modules, the corresponding stud location, welding plate type, plate thickness, welding controller real-time production data and alarm information, and controller design parameter layer can be queried.
[0045] Algorithm data support: The multi-dimensional data involved in the portrait description is based on a certain operator (attached Figure 2 The data required by the triangle in the figure is used to link the corresponding data of the portrait to the operator, based on the prediction support of the operator.
[0046] Enhance the interpretability of analysis results: Figure 2 For this stud welding defect, the operator predicts that the probability of "situation B" is higher than that of "situation A". At this time, we process the operator's data and the strong related information generated by the operator's prediction process, bind them to the operator node, and finally return to the root cause chain. This part of information (current trend change chart, model statistical current value distribution chart, current historical evolution data, etc.) is attached. Based on the rich actual data on site and the model statistical results, the persuasiveness and explainability of the root cause chain prediction are greatly enhanced.
[0047] Step S2: Based on the existing expert experience database data, stud welding process principles and related case data information, various defects in the stud welding process are sorted out, and a multi-level root cause map of stud welding defects is obtained; the first layer is the parameter composite layer, the second layer is the basic parameter layer, and the third layer is divided into the working condition equipment layer. At the same time, for the second basic parameter layer, a decision tree is used to classify the defect categories, sort out the different contributions of key data, and combine the stud welding process mechanism to characterize the propagation relationship between different factors. This structure and causal propagation relationship are conducive to providing stud welding engineers with powerful root cause analysis results of stud welding defects.
[0048] Therefore, based on the visualization of the decision tree, the contribution of parameters such as pressing time, energy, arc striking voltage, and pressing depth can be analyzed, and the influence relationship between different parameters can be depicted based on the process mechanism of stud welding.
[0049] Step S3: Based on the stud welding defect input situation of the root cause analysis system, a description of the input information in a specific format is designed ("input 1"->"input 2"->"input n"). After the expert confirms that there is no problem with this description, the multi-level root cause map is automatically updated without human participation. Through the event extraction algorithm in natural language, the engineer's problem description is extracted into the corresponding event causal propagation chain, and the nodes of the link are aligned with the existing nodes in the map. If an unaligned node is found, the node is fed back to the expert, and after the expert confirms it, it is integrated into the root cause map to achieve automatic update. It is specifically divided into the following steps:
[0050] Step S31: Record the root cause analysis of the on-site input, and extract multiple subjects, predicates, and objects from the description by designing an event extraction algorithm, and finally form multiple causal events or sequential events. The subject, predicate, and object form the event in the following way:
[0051] (Subject, Predicate, Object) 1 =Event i
[0052] The form is as follows:
[0053] event 1 -[Cause and Effect / Sequence]->Event 2 ->…->Events n ;
[0054] Step S32: Through similarity model training, the current extracted event can be calculated i All events in the root cause graph j The similarity S ij , set the threshold x = 0.80, if S ij ≥x, then judge the event i With events j It’s the same event;
[0055] Step S33: If there is a matching result for the event in step S32, then repeat step S32;
[0056] Step S34: If step S33 is completed and S is not satisfied ij Events ≥ x i , then the event i and its subsequent events, spliced into the last one in the root cause graph that satisfies S ij Events ≥ x j Later, and integrated into the atlas.
[0057] Step S4: After the complete data relationship and graph are established through steps S1 and S2, the cold start method of the root cause analysis system mainly relies on expert experience and initializes different probabilities on each chain propagation chain according to the statistical results of the historical occurrence frequency of the root cause link; and fine-tunes all links in combination with the statistics of the recent occurrence frequency of stud welding defects of different studs; realizes the adaptive function of cold start, which is mainly divided into the following four steps:
[0058] Step S41: The probability of the jth root cause in all studs set by original expert experience is: P j
[0059] Step S42: recent statistical results of all studs:
[0060] a. Set a time window of one month for the root cause of stud welding defects, and the total number of problems found in all studs is t;
[0061] b. Count the number of occurrences of each root cause in t problems j ;
[0062] c. Through Calculate the probability of occurrence of the jth root cause of all studs;
[0063] Step S43: recent statistical results of the i-th stud:
[0064] a. Set a time window of one month for the root cause of stud welding defects. The total number of problems found in the i-th stud is t. i ;
[0065] b. Count the number of occurrences t of the jth root cause of the i-th stud ij ;
[0066] c. Through Calculate the probability of the jth root cause of the i-th stud occurring in the near future;
[0067] Step S44: The probability P of each root cause set based on the original expert experience jand the recent statistical result P of all studs in step S42 j and the recent statistical result P of a stud in step S43 ij , respectively set the weight w 1 , w 2 , w 3 , we get the jth root cause probability P of the i-th stud ij =w 1 ×P j +w 2 ×P j +w 3 ×P ij .
[0068] Step S5: Predict the current stud welding defect based on the composite parameter layer in the multi-level root cause map. It is necessary to judge the probability of different root cause propagation paths based on the statistical results of the big data threshold by whether the actual value crosses the threshold and the degree of proximity between the actual value and the threshold.
[0069] The specific method is as follows (taking heat as an example):
[0070] Threshold statistical method:
[0071] Step S51: Calculate the average heat value in the historical data of a certain soldering point
[0072] Step S52: Calculate the standard deviation δ of heat in the historical data of a certain soldering point;
[0073] Step S53: Set greater than or less than The values of are considered outliers and scored as 100%.
[0074] Step S6: To predict the current stud welding defect according to the basic parameter layer in the multi-level root cause map, it is necessary to judge the probability of different root cause propagation paths based on the statistical results of the big data parameter space and the distribution of the actual value in the parameter space. The specific method is as follows (taking arc starting voltage as an example):
[0075] Parameter space statistical methods:
[0076] Step S61: The arc striking voltage in the data is used as the evaluation value, and the data of other dimensions are reduced to 2 dimensions by the PCA method, and their distribution in space is counted and divided into different areas. The central 70% of the data is area A, the area of 70%-80% is area B, the area of 80%-90% is area C, and the area above 90% is area D;
[0077] Step S62: Area A, Area B, Area C, and Area D represent normal area, sub-abnormal area, abnormal area, and potential problem area, respectively, and the corresponding multi-dimensional data of the original data are normal parameter range, sub-abnormal parameter range, abnormal parameter range, and potential problem parameter range. At the same time, the multi-dimensional data of each stud welding defect is scored.
[0078] Step S7: For the working condition and equipment layer, with the help of deep learning methods, from the data level, learn the different impacts of different working conditions on the data, and perform multi-classification to calculate the possibility of different stud welding defects in multiple working conditions; the input data includes factors such as current, voltage, and time. In addition, it also includes the introduction of physical prior knowledge, and the calculated equivalent resistance, power and other values. Multi-classification is performed through a multi-layer neural network, and the classification results are different working conditions and equipment problems, such as: foreign matter on the surface of parts, unreasonable plate thickness, welding gun failure to spit nails, etc.
[0079] Step S8: After the root cause scoring of each layer is completed, the initial score of the root cause chain is calculated by accumulating the scores of each node in the chain, and finally the scores of each chain are normalized and used as the final score;
[0080] Based on the trained decision tree model, the decision boundaries of the decision tree classification of each dimension data are obtained. Finally, by describing the degree of deviation between the actual value and the corresponding decision boundary, the corresponding node in the corresponding root cause graph is calculated, and the weight of the node is calculated. The specific calculation formula is as follows:
[0081] According to the decision tree model, the decision boundary of the i-th dimension is obtained The corresponding actual value is x i , the final node weight Finally, after the accumulation of each node, the total score of each link is normalized and used as the weight of the node. Finally, the scores of different root cause chains are calculated and the results are returned.
[0082] Step S9: Through the use of the system, the stud welding engineer will confirm the final root cause chain when the problem is finally closed. Through the data and the final confirmed root cause chain, and continuous accumulation, the model and statistical results will be updated regularly. In addition, the distribution of the results of model training in different periods will be compared regularly. If there is a large difference in the distribution, the stud welding engineer can choose to use one or more different models and set the contribution ratio of each model. If the root cause chain finally confirmed by the stud welding engineer is inconsistent with the results of the root cause analysis system, at this time, the data will be automatically sent to the working condition equipment layer model for fine-tuning together with the historical data.
[0083] The above method is applied to the analysis of stud welding quality problems. The analysis is conducted on the basis of the quality problems of on-site spot checks. Based on this method, the most likely root cause chain is analyzed to help on-site engineers find the root cause of the problem and improve efficiency. Compared with traditional databases, this method solves the problem of root cause analysis when the data has no topological structure. At the same time, with the help of knowledge graphs, based on the designed graph structure, the data of stud welding is reasonably organized, combined with various algorithms related to root cause analysis, while accurately performing root cause analysis of stud welding defects, the basis for root cause analysis judgment is displayed, and the interpretability of root cause analysis is enhanced. The difficulty of mapping traditional relational databases is greatly reduced, making it easier to build and query.
[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for root cause analysis of stud welding defects based on knowledge graph, characterized in that: The following steps are involved: Step S1: According to the data generated by the actual production of the stud welding process, the data is organized according to the structural relationship combining time and space to establish a complete data relationship; Step S2: Based on the existing expert experience database data, stud welding process principles and relevant case data information, various defects in the stud welding process are sorted out, and a multi-level root cause map of stud welding defects is obtained; Step S3: Combined with the problem description of the new defect entered on the mobile phone, the corresponding root cause chain and optimization solution chain are extracted from the multi-level root cause map in step S2 using the event extraction algorithm. After confirmation, they are deposited into the multi-level root cause map again. Step S4: Initialize different probabilities on each chain propagation chain based on the statistical results of the historical occurrence frequency of the root cause link; and fine-tune all links based on the statistics of the recent occurrence frequency of stud welding defects of different studs; Step S5: predicting the current stud welding defect according to the composite parameter layer in the multi-level root cause map, and performing statistics and scoring; Step S6: predicting the current stud welding defect according to the basic parameter layer in the multi-level root cause map, and performing statistics and scoring; Step S7: For the working condition equipment layer, using the deep learning method, from the data level, learn the different impacts of different working conditions on the data, and perform multi-classification to calculate the probability of different stud welding defects in multiple working conditions; Step S8: After the root cause scoring of each layer is completed, the initial score of each root cause chain is calculated by accumulating the scores of each node in the link, and finally the scores of each link are normalized and used as the final score; Step S9: Through the data and the finally confirmed root cause chain, and constantly accumulate, the model and statistical results are finally updated regularly; Wherein, the step S3 specifically includes the following steps: Step S31: Record the input of stud welding defects based on the root cause analysis system, and extract multiple subject, predicate and object information from the description of the specific format of the input information by designing an event extraction algorithm, and finally form multiple causal events or sequential events; Step S32: Through similarity model training, the current extraction event can be calculated i All events in the root cause graph j The similarity S ij , set the threshold x = 0.80, if S ij ≥x, then judge the event i With events j It’s the same event; Step S33: If there is a matching result for the event in step S32, then repeat step S32; Step S34: If step S33 is completed and S is not satisfied ij Events ≥ x i , then the event i and its subsequent events, spliced into the last one in the root cause graph that satisfies S ij Events ≥ x j Behind, and integrated into the atlas; In step S6, based on the statistical results of the big data parameter space, the probability of different root cause propagation paths is determined through the distribution set of actual values in the parameter space; the specific parameter space statistical method is as follows: Step S61: The arc striking voltage in the data is used as the evaluation value, and the data of other dimensions are reduced to 2 dimensions by the PCA method, and their distribution in space is counted and divided into different areas. The central 70% of the data is area A, the area of 70%-80% is area B, the area of 80%-90% is area C, and the area above 90% is area D; Step S62: Area A, Area B, Area C, and Area D represent the normal area, the secondary abnormal area, the abnormal area, and the potential problem area, respectively, and the corresponding multi-dimensional data of the original data are the normal parameter range, the secondary abnormal parameter range, the abnormal parameter range, and the potential problem parameter range; and at the same time, the multi-dimensional data of each stud welding defect is scored; In step S8, based on the trained decision tree model, the decision boundaries of the decision tree classification of each dimensional data are obtained, and finally, by characterizing the degree of deviation between the actual value and the corresponding decision boundary, the corresponding node in the corresponding root cause graph is calculated, and the weight of the node is calculated.
2. According to claim 1, a method for analyzing the root cause of stud welding defects based on knowledge graph is characterized in that: The multi-level root cause map constructed in step S2 includes: a first-level parameter composite layer, a second-level basic parameter layer, and a third-level operating condition equipment layer.
3. The method for root cause analysis of stud welding defects based on knowledge graph according to claim 2 is characterized in that: For the second basic parameter layer, machine learning is used to classify different defects, and based on the model's weights for each dimension of data, the importance and relevance of the data are analyzed.
4. The method for root cause analysis of stud welding defects based on knowledge graph according to claim 1, characterized in that: In step S5, based on the big data threshold statistical results, the probability of different root cause propagation paths is determined by whether the actual value crosses the threshold and the degree of proximity between the actual value and the threshold; the specific threshold statistical method is as follows: Step S51: Calculate the average heat value in the historical data of a certain soldering point Step S52: Calculate the standard deviation δ of heat in the historical data of a certain soldering point; Step S53: Set greater than or less than The values of are considered outliers and scored as 100%.
Citation Information
Patent Citations
Real-time root cause analysis method based on application knowledge graph
CN113032238A
Steel product surface longitudinal crack defect tracing and analyzing method based on knowledge graph
CN113420157A
