Multi-layer causal knowledge network for product quality problem analysis using third data

Abstract

The realization of digital intelligence is inseparable from the collection and effective application of knowledge. However, in the production process of enterprises, there is a type of “Third Data” which is rich in empirical knowledge but has not been paid attention to. This kind of data is generated from the records of the process in which enterprise employees solve quality problems and is the externalization of empirical knowledge. Because causal analysis is the key to solving quality problems, and there are many factors affecting the cause analysis and the analysis process is complex, this paper makes full use of the causal knowledge and problem contexts contained in the “Third Data”, and systematically considered the duality of cause analysis, the situation dependence of quality problems, the diversity of situations, the similarity among actual quality problems, the differences between short-term solutions and long-term solutions, then proposed the concept of Multi-layer Causal Knowledge Network(MCKN). Next, this paper analyzes the composition and expression of MCKN, and the corresponding construction process and steps are given. Finally, through the case study, the scientificity, rationality, and effectiveness of MCKN are verified. MCKN proposed in this paper provides new ideas and methods for comprehensive and in-depth analysis of the causality of quality problems and helps to improve the level of knowledge and intelligence of quality management.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-49281-w.

Introduction

With the development of digital technology and the emergence of the digitalization-intellectualization concept, the transformation toward knowledge-based and intelligent quality management has become an essential component of Quality 4.0^1–4.The importance of data, information and knowledge in quality management is increasing and has attracted the attention of relevant scholars^5–7. For example, Rajpathak et al.⁸ developed a system that can be used for fault detection in the automotive field using unstructured maintenance text data. Jiang et al.⁹ introduced fuzzy analytic hierarchy process and chaotic fuzzy regression method when using Quality Function Deployment (QFD) for new product planning. Xu et al.^10,11 proposed a component-failure mode matrix construction method applicable to failure mode and impact analysis using data from the production process, and also proposed an intelligent quality problem solving system that can be used in the automotive industry, and there are also scholars who use neural networks, regression modeling, and deep learning to intelligently diagnose faults in rolling bearings^12–15.

The solution of quality problem is an important part of quality management, and seeking the cause and effect is an effective way to solve the quality problem. Only by conducting a comprehensive and in-depth analysis of quality problems and identifying the underlying causes can we “prescribe the right medicine” to effectively resolve the problems¹⁶. At present, the research on causality of quality problems is relatively rare. One of the hotspots of causality research is to find variables that may have causality from a large amount of data^17–19, and to infer the size of the causal effect between variables by means of do operation (that is, unilaterally implementing intervention actions to judge the change of probability), counterfactual inference (what if) and other methods^20,21, so that the data has the ability to answer “why”²². Some scholars have also focused on how to use text data to construct causal networks^23–26. The basic idea of this research is to first use language pattern matching method²⁷ or machine learning-based method²⁸ to extract “causal pairs” from text, and then use precise matching or similarity-based calculation method to fuse similar nodes in “causal pairs” to form a causal network^29–31.

In the aforementioned process of constructing from data to networks, knowledge discovery methods represented by Data Mining (DM) provide powerful tools and theoretical foundations for causal analysis and knowledge extraction. DM have been widely applied to extract potentially valuable information from massive amounts of data, in the fields of quality management and knowledge discovery³². For example, Wang W et al.³³ have used association rule mining methods to analyze production data and extract the relationships among key quality factors, thereby achieving quality control of products in intelligent manufacturing processes. And some researchers have utilized text mining and natural language processing techniques to perform pattern recognition and effective information extraction from large amounts of unstructured quality reports and textual data, with the aim of improving quality, flexibility, and productivity in the production process^34–36. These outstanding achievements demonstrate the powerful capability of data mining in handling complex industrial production data. More importantly, they provide technically feasible support for the effective extraction of causal knowledge from frontline quality problem-solving records.

The above research provides methods and ideas for the causal analysis of quality problems. However, the cause analysis of quality problems has its uniqueness and complexity, which is mainly manifested in the following aspects^37–40: Firstly, the “duality” of cause analysis, that is, it is not only necessary to combine the specific situation of the problem to make a specific analysis of the specific problem, so as to formulate temporary containment measures, it is also necessary to strip the problem contexts so that the problem is abstract and general, so as to facilitate in-depth analysis of the “real cause” of the problem and formulate a long-term plan. Secondly, the causal relationship of quality problems is usually complex and changeable, mostly showing a network structure. At the same time, in the current quality management of enterprises, most of the front-line employees only rely on personal experience to conduct a single linear analysis of the cause of the problem by means of fishbone diagram or 5-WHY, and it is difficult to systematically consider many factors that affect the causal analysis of quality problems⁴¹. Moreover, personal experience is usually affected by personal knowledge and experience, which has great limitations and will be lost with the flow of employees.

With the digitalization and informatization of manufacturing enterprises, enterprise employees can now utilize information systems to objectively document the process of solving quality problems, forming a unique type of data called “Third Data”. Further analysis reveals that this data essentially represents the externalization of employees’ experiential knowledge in solving quality problems. However, personal experiential knowledge inherently carries subjectivity, varying from person to person, across different times and locations, and depending on specific problem context. Furthermore, the process of recording is actually the process of making empirical knowledge explicit, and this process will also inevitably be affected by subjectivity.

Therefore, this type of data is a knowledge-based data category characterized by both objectivity and subjectivity. It is different from the objective data generated by automatic recording by machine equipment, that is “First Data”. And it is also different from the subjective “Second Data”, which generated by consumers’ mutual exchange of product experience based on the Internet. If the “First Data” is objective, machine-generated data automatically recording production processes, states, and results, such as production logs; the “Second Data” is subjective, user-generated data produced by consumers interacting via internet platforms, such as reviews on consumption experiences, expectations, products, services, etc.; then the “Third Data” is knowledge-based, producer-generated data formed by objective records of the occurrence, causes, solutions, and effectiveness of quality problems by enterprise employees.

Thus, the “Third Data” is the data foundation for knowledge mining and knowledge management. The knowledge contained in it is not only the externalization of personal experience, but also the fusion of numerous knowledge sources, and the result of collective knowledge externalization and sharing. If this empirical knowledge can be effectively managed and utilized, it will greatly improve the effectiveness of root cause analysis, not only surpassing the limitations of personal experience in causal analysis of quality problems, but also reducing the impact of organizational knowledge loss caused by employee turnover. Nevertheless, this data category remains substantially neglected and underutilized in current practice.

In summary, based on the difficulty of quality problem cause analysis and its strong practical needs, this paper intends to use the causal knowledge and problem contexts contained in “Third Data” to propose a knowledge model suitable for quality problem causal analysis for the duality and complexity of quality problem cause analysis. It has certain practical value and theoretical significance. The main contents are as follows: Firstly, help employees conduct quality problem causal analysis from multiple perspectives, multiple levels and all directions, and provide new ideas and methods for comprehensive and in-depth analysis of quality problems. Secondly, it can help the integration of traditional quality management methods and information technology, and the improvement of quality management knowledge and intelligence level. It can make a useful contribution to the transformation and upgrading of manufacturing industry and the development of quality 4.0, and has strong application value and practical significance. Thirdly, Complementing and improving the relevant theories and methods of quality management from the perspective of knowledge is conducive to the mutual integration and complementation of quality management and knowledge management, and also makes a useful contribution to the supplement and improvement of the relevant theories and methods of knowledge management, which has certain theoretical value.

Analysis on the action mechanism of causal knowledge

When a quality problem occurs, enterprise employees usually conjecture the causes of the problem based on previous experience, and at the same time, test the conjectured causes. If the conjectured causes are wrong or inaccurate, they will refute them and put forward new possible causes. Once the causes are determined, tentative solutions to the problem can be proposed. Therefore, the process of analyzing the causal relationship of quality problems is actually a continuous process of “conjecture-refutation”. And the knowledge necessary for cause analysis will be generated and accumulated in this repeated process of “conjecture-refutation”^42–44.

From an individual perspective, in the repeated practice of “conjecture” and “refutation” when enterprise employees analyze quality problems, they will develop an understanding of them and accumulate corresponding experience and knowledge. When encountering similar problems again, the experience and knowledge will come into play, which can reduce the repeated process of “conjecture-refutation” to a certain extent, and the efficiency and effectiveness of analyzing the causes of quality problems will be improved accordingly^45,46. This is in line with the view of the theory of practice, that is: practice, knowledge, re-practice, and re-knowledge. This form repeats itself in endless cycles, and the content of each cycle of practice and knowledge advances to a higher level relatively. Meanwhile, this also explains why “veteran employees” with rich experience and knowledge can conduct causal analysis “both quickly and well”.

From the perspective of an enterprise, the “Third Data” is overlooked, and the knowledge necessary for analyzing quality problems has not been effectively collected and reused, failing to form “organizational memory”. Staff turnover and attrition (such as job transfers, retirements, resignations, etc.) will all lead to the occurrence of organizational amnesia. When similar problems arise again, the organizational memory fails to play its due role, resulting in the efficiency and effectiveness of analyzing the causes of quality problems both needing to be improved, as shown in Fig. 1.

Fig. 1.

Comparison of quality problem-solving between individual and enterprise.

If an enterprise can effectively accumulate and apply the causal knowledge generated during the analysis of quality problems, then the enterprise can also become a “veteran employee”, and the efficiency and effectiveness of analyzing quality problems will also be improved accordingly. There are many factors affecting the cause analysis of quality problems, such as the duality of cause analysis, the context-dependence of quality problems, the diversification of contexts, the similarity between actual quality problems, and the difference between containment measures and long-term solutions. This paper needs to systematically and comprehensively consider the above-mentioned many factors affecting the cause analysis of quality problems, and combine the characteristics of the causal knowledge contained in the “Third Data” to propose a knowledge model with practical application value, realize the closed loop of knowledge growth of “practice - recognition - re-practice - re-recognition”, help enterprises form “organizational memory”, and effectively improve the efficiency and effectiveness of cause analysis.

Proposal and expression of multi-layer causal knowledge network

Proposal of multi-layer causal knowledge network

Due to the duality of the cause analysis of quality problems, the knowledge model proposed in this paper needs to include two parts: The Concrete Causal Knowledge Network (hereinafter referred to as CCKN) that contains the problem context, which is used for “concrete analysis” of “specific problems”; the Abstract Causal Knowledge Network (hereinafter referred to as ACKN) that strips away the problem context. This part only retains the inherent and essential features of the problem, making the problem have a certain degree of abstraction and generality, and is applicable to the “in-depth analysis” of “class problems”.

The quality problem contexts has diversified characteristics, such as the conditions of the problem, the time of the problem, and the subject related to the problem. The choice of different situations has different meanings for the analysis of specific problems. For example, the production time is selected as a situational factor to analyze the causal relationship between quality problems that occur in different production periods. Therefore, according to the choice of problem contexts, CCKN shows multi-level characteristics. For example, CCKN can include time-layer concrete causal knowledge network, component-layer concrete causal knowledge network, and material-layer concrete causal knowledge network. In CCKN, the difference between different levels of concrete causal knowledge network is that the selected problem contexts are different, and the quality problems are the same. Therefore, there is an correlation relationship of “same quality problem but different problem contexts” between nodes representing the same quality problem in concrete causal knowledge network at different levels. For example, there is an “correlation relationship” between the node representing “cracking during the morning shift” in the time-level concrete causal knowledge network and the node representing “cracking when producing the inner panel of the car door” in the component-level concrete causal knowledge network.

There are some similarities between the actual quality problems, such as tearing, cracking, flanging cracks, etc., which can be classified as crack problems, but the specific situation is different. Whether to treat such problems differently determines the “node granularity” of “class problems”. Different node granularity affects the effect of in-depth analysis of the problem. Therefore, when constructing ACKN, it is necessary to choose whether to treat similar problems differently according to the specific situation, that is, the node granularity of screening class problems. The selection of different granularity makes ACKN present multi-level characteristics. In ACKN, abstract causal knowledge networks with different granularities can perform causal analysis with different emphases. For example, coarse-grained abstract causal knowledge networks are more conducive to analyzing causal associations between problems, while fine-grained abstract causal knowledge networks can analyze the real causes of quality problems more accurately. Due to the different choice of node granularity, there is a subordination relationship between the “class problem” nodes represented by different levels of abstract causal knowledge networks. For example, there is a “subordination relationship” between the nodes representing “split” in the fine-grained abstract causal knowledge network and the nodes representing “crack” in the coarse-grained abstract causal knowledge network, where “split” is the species concept and “crack” is the genus concept.

Since CCKN is “specific problems” that contains problem contexts, and ACKN is “similar problems” that strips off situations. Therefore, there is a subordination relationship between the nodes in CCKN and ACKN. For example, there is a subordination relationship between the node representing “cracking problem occurring when producing the inner panel of the car door” in CCKN and the node representing “crack” in ACKN. Among them, “cracking occurring when producing the inner panel of the car door” is a species concept and “crack” is a genus concept. The characteristics of each level in MCKN are both the same and interdependent. There are “correlation relationship” and “subordination relationship” between nodes in different levels. For the convenience of description, this paper collectively refers to them as “correlation-subordination relationship”. Each level is linked to each other due to the correlation-subordination relationship between nodes. In the actual production process, when it is necessary to analyze quality problems from different angles, it can jump to the most suitable level through this link for causal analysis.

Through the above analysis, the new causal knowledge model proposed in this paper is a composite knowledge model that includes two parts and multiple levels. Each level is different but interdependent. They are linked by the “correlation-subordination relationship” between nodes. They are coordinated with each other and unified with each other. They jointly act on the causal analysis of quality problems from different angles. This paper calls it “Multi-layer Causal Knowledge Network (hereinafter referred to as MCKN)”. MCKN systematically considers the factors that affect the causal analysis of quality problems, such as the duality of cause analysis, the situational dependence of quality problems, the diversification of situations, the similarity between actual quality problems, and the differences between containment measures and long-term programs. It helps employees to conduct causal analysis of quality problems from multiple perspectives, multiple levels, and three-dimensionally, and improves the efficiency and effectiveness of causal analysis.

Composition and expression of multi-layer causal knowledge network

The causal relationships among quality problems are intricate. This paper intends to use “node” to represent “problems/causes” and “edges” to represent the “causal relationships” between nodes. The starting point of an edge is the “cause”, and the pointed node represents the “result”. That is to say, the form of “node-edge-node” is adopted to represent the causal knowledge contained in the “Third Data”, and this paper refers to it as a causal knowledge group. Since when analyzing quality problems, it is necessary to combine the specific contexts in which the problems occur in order to achieve a concrete analysis of specific problems, it is required to add problem contexts to the causal knowledge group in this paper, as shown in Fig. 2.

Fig. 2.

Causal knowledge group with problem contexts.

Since the problem contexts have the characteristics of diversification, in Fig. 2, the problem context is uniformly represented by Ps. According to this article, by screening the problem context and node granularity of the causal knowledge group, and overlapping and accumulating the same nodes and edges, and expressing the cumulative frequency of nodes by the size of nodes, and expressing the cumulative frequency of edges by the thickness of edges, a number of causal knowledge networks with different levels and different effects will be formed, as shown in Fig. 3.

Fig. 3.

Schematic diagram of causal knowledge network.

In Fig. 3, N2 and N4 are the causes that lead to N1. N2 is a problem caused by N3, while N3 is caused by N4. A large number of causal relationships form a complex network structure, that is, the causal knowledge network. The different frequencies of node occurrences result in differences in node sizes. Meanwhile, the thicknesses of the edges between nodes also represent the differences in the accumulated frequencies of the edges. Multiple causal knowledge networks that are different from each other but closely related are linked to each other based on the “correlation-subordination relationship” between nodes, and then MCKN can be constructed, as shown in Fig. 4.

Fig. 4.

Schematic diagram of MCKN.

MCKN links multiple causal knowledge networks that are different but closely related to each other into a whole. When it is necessary to analyze quality problems from different angles, it can jump to the most suitable level through this link for causal analysis, helping enterprise employees conduct causal analysis of quality problems from multiple angles, at multiple levels, and in a three-dimensional way. Due to the different frequencies of accumulation of nodes and edges, parameters appear in MCKN, such as: node frequency, edge frequency, probability of inferring causes from effects, probability of predicting effects from causes, etc. (see Appendix for details). Whether it is a causal knowledge network at different levels or the correlation-subordination relationship between nodes at different levels, they are all represented by nodes and the relationships between nodes. Therefore, MCKN can be expressed by a unified mathematical formula, as shown in Eq. (1).

1

In Eq. (1), W represents MCKN; N represents the set of nodes in MCKN; E represents the set of edges in MCKN; R represents the set of parameters in MCKN.

1. Nodes in MCKN

The nodes in MCKN represent quality problems or causes, which can be recorded as .

• (2)Edges in MCKN

The edges in MCKN can also be divided into two types: edges within the causal knowledge network layer and edges between layers.

The edges within the layer represent the causal relationship between nodes and are represented by directed edges between nodes. It is recorded as . Among them, represents the directed edge from to ; is a cause of ; is a result caused by .

The edges between layers represent the correlation-subordination relationship between nodes at different levels. The correlation-subordination relationship is a general term for “correlation relationship” and “subordination relationship”. The “subordination relationship” can be represented in the form of directed edges, pointing from “specific problems” in CCKN to “abstract problems” in ACKN, and from “species concepts” in abstract causal knowledge networks with fine granularity to “genus concepts” in abstract causal knowledge networks with coarse granularity. The “correlation relationship” is the same problem under different context selections. For the uniformity of expression, directed edges can also be used for representation. For example, a node representing “cracking occurs in the production of inner door panels during the morning shift” points to a node representing “cracking occurs during the morning shift”.

Therefore, the correlation-subordination relationship is also represented by directed edges and can be recorded as . Among them, represents the directed edge from to , indicating that the problem represented by node belongs to the species concept, and the problem represented by node belongs to the genus concept, or is the direction between nodes with the correlation relationship.

• (3)Parameters in MCKN

There are numerous parameters included in MCKN, such as node frequency, edge frequency, probability of seeking causes from effects, probability of predicting effects from causes, etc. Different parameters have different functions and meanings (see the Appendix for details).

The theoretical basis and self-perfection of multi-layer causal knowledge network

MCKN can make full use of the “causal knowledge” generated and applied in the process of solving quality problems, helping enterprises to realize the closed loop of knowledge growth of “practice - recognition - re-practice - re-recognition”, form “organizational memory”, and thus improve the efficiency and effectiveness of cause analysis. Given that all aspects of quality problem solving are interrelated and mutually restrictive, and there are numerous factors affecting cause analysis, this paper needs to be guided by systematic thinking, conduct systematic analysis and consideration from an overall perspective to ensure that MCKN is scientific, reasonable and effective. Meanwhile, cause analysis, as a key link in the process of solving quality problems, is an integral part of quality management. How to manage and apply the causal knowledge contained in the “Third Data” falls within the scope of knowledge management research. The difficulty and key point in constructing MCKN lie in how to mine and represent causal knowledge from the “Third Data”. Relevant theoretical methods such as text analysis techniques, knowledge mining methods, artificial intelligence algorithms, and various related algorithms provide technical support for it. Tools like Python, MATLAB, Gephi, and SPSS guarantee its implementation.

To sum up, the MCKN proposed in this paper is based on the practical view that “the content of each cycle of practice and cognition has advanced to a higher level relatively”. Guided by systematic thinking, supported by quality management theory and knowledge management theory, with text analysis, knowledge mining, artificial intelligence algorithms and other methods as the main technical means, and with Python, MATLAB, Gephi, SPSS and other tools for implementation, it makes full use of the causal knowledge contained in the “Third Data” accumulated in the practical activities of enterprises in solving quality problems, which can help improve the efficiency and effectiveness of cause analysis, as shown in Fig. 5.

Fig. 5.

Technical & theoretical system of the causal knowledge model.

The construction of MCKN is based on the “Third Data”, and this type of data will gradually increase as the production activities of manufacturing enterprises continue. Through Data Mining methods, the causal knowledge embedded in Third Data is extracted, and this knowledge further refines MCKN. In other words, MCKN continuously self-improves as it assists enterprise employees in analyzing quality problems. Specifically, when enterprise employees use MCKN for causal analysis, if MCKN can effectively help them analyze the causes of problems, the corresponding parameters in the model will undergo corresponding changes, which helps improve the accuracy of using the model. If it fails to help enterprise employees, in accordance with the procedures, the enterprise employees will still record the process of solving this problem, generating a new piece of “Third Data”. Subsequently, the causal knowledge embedded within this data will be extracted form the piece of “Third Data” by Data Mining methods and incorporated into the MCKN to enrich relevant causal relationships, thereby refining the model.

In other words, the MCKN proposed in this paper is not a fixed and unchanging model. Instead, it will continuously undergo dynamic adjustments along with the progress of production activities. It can better adapt to the production activities of enterprises and help enterprise employees conduct causal analysis of quality problems more accurately. It has the characteristic of continuous self-improvement, as shown in Fig. 6.

Fig. 6.

Self- improvement of MCKN.

As shown in Fig. 6, with the continuous progress of enterprise production activities, a cyclic model usage pattern of “recording data - mining knowledge - improving the model - facilitating production” has gradually taken shape. MCKN dynamically changes in such a cycle and continuously improves itself, adapting better to the production activities of enterprises and helping enterprise employees analyze quality problems in a more comprehensive and effective manner. All in all, as long as enterprises keep carrying out production activities continuously and the “Third Data” keeps accumulating, MCKN will become more and more refined. It will not only contain more and more comprehensive causal knowledge, but also continuously enhance the accuracy and reliability of causal analysis, and can continuously help enterprises improve the efficiency of analyzing and solving quality problems.

The construction process of multi-layer causal knowledge network

MCKN is a composite knowledge network with two parts and multiple levels proposed according to actual needs based on the causal knowledge and problem contexts contained in the “Third Data”. Combined with the foregoing, building MCKN generally requires: Firstly, mining the causal knowledge and problem contexts contained in the “Third Data” and constructing a causal knowledge group with problem contexts. Secondly, completing the construction of CCKN by selecting an appropriate quality problem context and combining the construction method of causal knowledge network. Thirdly, completing the construction of ACKN by screening the granularity of quality problem nodes and combining the construction method of causal knowledge network. Fourthly, taking the overlapping and accumulation of the same nodes and edges in the causal knowledge group as the core idea for constructing causal knowledge network and proposing corresponding methods. Fifthly, linking the mutually distinct and interdependent causal knowledge network through the correlation-subordination relationship between nodes to construct MCKN.

The construction process is shown in Fig. 7.

Fig. 7.

The construction process of MCKN.

The specific construction steps of MCKN are as follows:

Step 1: Construct a causal knowledge group with a problem context

This step mainly uses data mining and text analysis methods according to the specific characteristics of the “Third Data” to mine causal knowledge from the data, and combines the problem contexts contained in the data to construct a causal knowledge group with problem contexts. Generally speaking, the “Third Data” is a record of the quality problem-solving process, and there are relatively many words describing quality problems, problem contexts, causes, etc. Therefore, methods such as word segmentation, word frequency statistics, text vectorization, feature extraction, part-of-speech analysis, syntactic dependency relationship analysis, semantic clustering, causal relationship identification and splitting can be used to process the “Third Data” to realize the mining of causal knowledge and construct a causal knowledge group with problem contexts.

Step 2: Select the problem context

The quality problem context has the characteristics of diversification, and the choice of different situations has different significance for the analysis of specific problems. Therefore, this step is mainly to select the appropriate problem context to help employees more effectively analyze specific problems. The choice of the problem context can be selected by the employees according to their own experience and the actual production needs, and can also be selected by the differences in the causes of quality problems in different situations in the statistical data.

Step 3: Filter node granularity

There are some similarities between the actual quality problems, such as tearing, cracking, flanging cracks, etc., which can be classified as crack problems, but the specific conditions are different. If it is not distinguished, the node granularity is larger, which can better analyze the causal relationship between quality problems. If it is distinguished and the node granularity is small, the underlying causes of quality problems can be analyzed more accurately. The screening of node granularity can be done by domain experts according to the needs of actual cause analysis and combined with their own experience, or it can be done by clustering analysis by calculating the similarity of quality problems represented by nodes in the causal knowledge group.

Step 4: Propose a causal knowledge network construction method

The core idea of the causal knowledge network construction method is to identify the same nodes and edges in the causal knowledge group and accumulate the same nodes and edges. Based on this idea, according to the selected problem context and the screened node granularity, combined with the specific situation of the data, CCKN and ACKN can be constructed.

Step 5: Establish a correlation-subordination relationship

The correlation-subordination relationship is a general term for the “correlation relationship” where “the quality problems are the same but the problem contexts are different” and the “subordination relationship” between nodes. Therefore, the correlation-subordination relationship in CCKN can be established by judging whether the quality problems are the same; the correlation-subordination relationship between ACKN can be established by the subordination relationship resulting from different node granularities; the correlation-subordination relationship between CCKN and ACKN can be established by the subordination relationship between specific quality problems containing problem contexts and “generic problems” without problem contexts.

Case study

Case background

This paper takes the quality problems encountered in the stamping production process of the body-in-white of an automobile manufacturing enterprise as an example for case study to verify the scientificity, rationality and effectiveness of MCKN. In the stamping manufacturing process of the enterprise in the example, quality problems such as cracks, wrinkles and galling often occur. In order to solve quality problems and improve production efficiency and product quality, front-line employees in the stamping workshop usually record the quality problem-solving process to form the “Stamping Workshop Product Quality Problem Record Form”. This form contains fields such as “problem description”, “cause analysis”, “product part”, “production model”, “production team”, etc., and contains a large amount of causal knowledge and problem contexts, meeting the basic requirements for constructing MCKN.

The obtained “Stamping Workshop Product Quality Problem Record Form” contains a total of 3997 pieces of data. Among them, some data are not completely recorded. For example, 483 pieces of data only fill in the “problem description” field but not the “cause analysis” field; 329 pieces of data only fill in the “cause analysis” field but not the “problem description” field; there are 42 pieces of data where neither the “problem description” field nor the “cause analysis” field is filled; 27 pieces of data do not fill in the “production team” field; 2 pieces of data do not fill in the “product part” field; 1 piece of data does not fill in the “production model” field. Among the above incomplete data, there are 27 pieces of intersecting data, and there are 3143 pieces of complete data.

The main purpose of the case study is to verify the method proposed in this paper. Eliminate the data with incomplete causes and effects and incomplete problem contexts in the “Stamping Workshop Product Quality Problem Solving Record Form”, and select 3143 complete pieces of data for case study to verify whether the MCKN proposed in this paper is scientific, reasonable and effective.

Construction of causal knowledge group

The construction of causal knowledge group is realized by mining causal knowledge from the “Third Data” and combining quality problem context. Therefore, it is necessary to mine relevant descriptors such as quality problems, problem contexts, and causes from the data. Then, process these words, and further identify the causal relationships among these descriptors from the data, and construct a causal knowledge group with the problem context.

Combined with the characteristics of the selected data, the construction of the causal knowledge group in this paper mainly includes the following steps:

Step 1: Mine relevant description words for representing quality problems, problem contexts, causes, etc

The source of the “Third Data” determines that the word frequency of words describing quality problems and their causes will be relatively high⁴⁷. Word frequency statistics can be used for screening. At the same time, syntactic structure can be used for assistance. For example, quality problem descriptions are expressions of production problems or phenomena that occur. They are mostly described using the subject-predicate structure of “subject + predicate”, that is, the pattern of “noun + adjective”⁴⁸. Words describing problem contexts such as product parts, production models, and production teams are mostly nouns.

After processing and analyzing the data, a total of 1343 relevant description words for representing quality problems, problem contexts, causes, etc. were mined. Some are shown in Table 1.

Table 1.

Relevant description words for quality problems, problem contexts, causes, etc. (part).

Relevant description words	Frequency
Crack	638
Wrinkle	481
Necking	376
Galling	342

Step 2: Construction of causal knowledge groups

In the data selected in this paper, there is a causal relationship between the two fields of “problem description” and “cause analysis”. Combined with fields that describe the problem context such as “product part”, “production model”, and “production team”, a causal knowledge group with a problem context can be constructed. Some are shown in Fig. 8. Due to the diversification of problem contexts, it is not convenient to display. Therefore, PS is used to represent the problem context set.

Fig. 8.

Causal knowledge groups (part).

Construction and application example of concrete causal knowledge network

In the data selected in this paper, the problem context mainly includes product parts, production models, production teams, etc. Combined with the actual production needs of the enterprise, two types of situational contexts, product parts and production models, are selected to construct corresponding concrete causal knowledge network respectively, as shown in Figs. 9 and 10 respectively.

Fig. 9.

Component-layer concrete causal knowledge network. The figure was generated using Gephi 0.10.1(URL link: https://gephi.org/).

Fig. 10.

Vehicle model-layer concrete causal knowledge network. The figure was generated using Gephi 0.10.1(URL link: https://gephi.org/).

Due to the addition of problem contexts, the nodes in CCKN appear rather discrete, which can be used to analyze quality problems in different contexts, as shown in Figs. 9 and 10. Among them, Fig. 9 is the concrete causal knowledge network at the component level. Through this network, the causal correlation of various problems occurring when producing different product components can be discovered. Figure 10 is the concrete causal knowledge network at the vehicle model level. Through this network, the causal correlation of various problems occurring when producing different vehicle models can be discovered. There are many nodes in Figs. 9 and 10, and the causal relationships are complex, so it is not convenient to display them one by one. Therefore, this paper focuses on selecting some nodes for display and weakening the others to a certain extent. In Fig. 9, this paper selects the necking problem that occurs when producing the inner door panel as an example. It is not difficult to find through this figure that when producing the inner door panel, there are a total of 23 reasons that may lead to the occurrence of the necking problem. Among them, wrinkles are the most likely cause of this problem, and the probability of seeking causes from effects is 0.394 (the calculation formula can be seen in Appendix). In Fig. 10, this paper selects the necking problem that occurs when producing the U5 vehicle model as an example. It is not difficult to find through this figure that when producing the U5 vehicle model, there are a total of 24 reasons that may lead to the occurrence of the necking problem. Among them, flange cracks are the most likely cause of this problem, and its probability of seeking causes from effects is 0.441(the calculation formula can be seen in Appendix). Although the dirty pressing surface is a relatively large node, indicating that this cause appears relatively frequently, its probability of seeking causes from effects is not high, only 0.009(the calculation formula can be seen in Appendix). Through the analysis of the whole figure, it is found that the dirty pressing surface can lead to many problems, and the most likely one is the crack problem when producing the U5 vehicle model.

Construction and application examples of abstract causal knowledge network

Combined with the suggestions of domain experts and the actual production situation, this paper only constructs a one-layer abstract causal knowledge network. In ACKN, the contextual factors of quality problems are stripped, and the node granularity is summarized, making the quality problems represented by it have a certain degree of abstraction and generality. This makes the causal knowledge network in this part be able to not only help analyze the deep-seated causes of problems, that is, “seeking causes from effects”, but also help predict the possible consequences that the problem may further cause, that is, “predicting effects from causes”.

Since the quality problems represented by the nodes of the abstract causal knowledge network have a certain degree of abstraction and generality, and the number of nodes is reduced compared to that of the concrete causal knowledge network, this paper has retained their labels to some extent. For the convenience of display, this paper focuses on presenting the necking problem and weakens the other nodes to a certain extent, as shown in Fig. 11.

Fig. 11.

Abstract causal knowledge network. The figure was generated using Gephi 0.10.1(URL link: https://gephi.org/).

Through the constructed abstract causal knowledge network, it can be found that there are 29 reasons that can lead to the occurrence of the necking problem. By calculating the probability of seeking the cause from the result, it can be found that wrinkles and the pressing surface damage are the main reasons for this problem. The probabilities of seeking the cause from the result for the two are 0.213 and 0.187 respectively(the calculation formula can be seen in Appendix). When this problem is found, such reasons can be considered first. While non-compliance of pressing surface gaps and insufficient hardness of castings are secondary reasons for this problem. The probabilities of seeking the cause from the result for the two are both 0.007(the calculation formula can be seen in Appendix). Through the analysis of the whole figure, further analyze the causes of problems such as wrinkles and the pressing surface damage, so as to find the deeper causes of necking. For example, there are 39 reasons for wrinkles. Among them, uneven material flow and dirty pressing surface are the main reasons for this problem. The probabilities of seeking the cause from the result are 0.177 and 0.176 respectively(the calculation formula can be seen in Appendix). There are 8 reasons for the pressing surface damage. Among them, local damage to the mold surface is the main reason, and the probability of seeking the cause from the result is 0.351(the calculation formula can be seen in Appendix). In addition, ACKN can also infer that the necking problem may further cause the occurrence of problems such as cracks and galling. The probabilities of predicting the result from the cause for the two are 0.694 and 0.306 respectively(the calculation formula can be seen in Appendix), and preventive measures can be taken in advance.

Construction and application examples of multi-layer causal knowledge network

According to the correlation-subordination relationship between nodes at different levels, connecting the constructed concrete causal knowledge network at the component level, the concrete causal knowledge network at the vehicle model level, and the constructed abstract causal knowledge network can construct MCKN. For the convenience of differentiation, this paper has reordered and recolored the nodes in the component-level concrete causal knowledge network, the vehicle model-level concrete causal knowledge network, and the abstract causal knowledge network, as shown in Fig. 12.

Fig. 12.

MCKN. The figure was generated using Gephi 0.10.1(URL link: https://gephi.org/).

In Fig. 12, the component-layer concrete causal knowledge network (with red nodes, located at the bottom layer) and the vehicle model-layer concrete causal knowledge network (with blue nodes, located at the middle layer) belong to CCKN. The abstract causal knowledge network with only one layer constructed (with green nodes, located at the top layer) belongs to ACKN. MCKN combines CCKN and ACKN in a coordinated and unified manner. When encountering quality problems, specific analysis of specific problems can be carried out through CCKN to quickly find the specific causes of the problem so as to formulate temporary containment measures. According to the affiliation relationship among nodes, it can be found out which category the problem belongs to. With the help of level transitions, the causal relationships behind this kind of problems can be analyzed in depth in ACKN, enabling the problem to be solved effectively in the long term.

Also taking the necking problem as an example, through Fig. 9, the specific causes of the necking problem when producing different components can be analyzed. Through Fig. 10, the specific causes of the necking problem when producing different vehicle models can be analyzed. According to the correlation-subordination relationship between necking problems in different situations in CCKN, the two can be combined to help enterprise employees analyze the specific causes of the necking problem in different situations more meticulously and specifically. Through the correlation-subordination relationship between nodes in CCKN and ACKN, the necking problem in specific situations can be transferred to the abstract causal knowledge network, and the deep-seated causes of the necking problem can be deeply analyzed.

Thus, it can be seen that CCKN and ACKN in MCKN are independent but closely related, which can help enterprise employees analyze quality problems comprehensively and deeply. Meanwhile, as production proceeds, the “Third Data” will gradually increase, and MCKN will also be further improved to better help enterprise employees analyze the causal relationship of quality problems.

Conclusion

Given the complexity and practical demands of causal analysis for quality problems, this paper makes full use of the causal knowledge and problem contexts embedded in “Third Data” to systematically consider multiple factors that influence causal analysis, including the duality of cause analysis, the contextual dependence of quality problems, the diversification of contexts, the similarity among actual quality problems, and the differences between containment measures and long-term solutions. Based on this, MCKN is proposed, and its corresponding construction process and steps are provided. The main conclusions are as follows:

1. “Third Data” is a knowledge-based data category formed by enterprise employees recording quality problem-solving processes. It possesses both objectivity and subjectivity, and exhibits both structured and unstructured characteristics. This distinguishes it from the objective, structured “First Data” automatically generated by machinery and equipment, as well as the subjective, unstructured “Second Data” generated by consumers exchanging product usage experiences via the Internet. The rich knowledge embedded in “Third Data” can provide valuable support for causal analysis and problem-solving in quality management.

2. The proposed MCKN is a composite knowledge network consisting of two parts and multiple levels. It can effectively organize and represent the causal knowledge mined from “Third Data,” assisting enterprise employees in conducting multi-perspective, multi-level, and multi-dimensional causal analysis of quality problems, thereby improving both the efficiency and effectiveness of causal analysis.

3. Taking quality problems encountered in the stamping production process of the body-in-white at an automobile manufacturing enterprise as an example, this paper demonstrates that MCKN can be successfully constructed from actual “Third Data.” The constructed model intuitively presents causal pathways among quality problems and supports multi-perspective, multi-level and multi-dimensional causal analysis, thereby verifying the scientificity, rationality, and effectiveness of MCKN.

The MCKN proposed and constructed in this paper is a composite knowledge network comprising two parts and multiple levels, with each level interconnected through correlation-subordination relationships among nodes. When quality problems need to be analyzed from different perspectives, these connections enable navigation to the most appropriate level for causal analysis, assisting enterprise employees in conducting multi-perspective, multi-level, and multi-dimensional causal analysis. This effectively addresses the challenge of comprehensively and deeply analyzing causal relationships in quality problems solely through personal experience, overcomes the limitations of experience-dependent causal analysis, and mitigates the impact of organizational knowledge loss caused by employee turnover. By integrating knowledge management with causal analysis, this model provides new research perspectives and methodological tools for the field of quality management, enriches the theoretical system of Quality 4.0, and contributes to advancing the knowledge-based and intelligent transformation of quality management.

Nevertheless, this study has certain limitations. For instance, the case study serves primarily as a proof-of-concept, and the generalizability of its findings requires further validation across more enterprises and industries. Additionally, the causal knowledge mined from “Third Data” is inherently experience-based and inductive, and there is currently a lack of mature validation mechanisms for systematically assessing the credibility of such causal knowledge and handling conflicting causal records. Therefore, future research will focus on the following directions: First, developing causal reasoning algorithms based on MCKN parameters to enable automated inference from specific problems to surface causes and root causes, and validating these algorithms in real-world scenarios. Second, establishing credibility assessment mechanisms for causal knowledge, exploring methods such as expert review or statistical causal inference to enhance the reliability of causal relationships within the model. Third, investigating deployment mechanisms for MCKN in actual enterprise environments, including data standardization, knowledge validation protocols, and integration with existing quality management systems, while further deepening collaboration with the case enterprise and other partner organizations to promote the practical application and dissemination of MCKN.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, Y.W. and F.L.; methodology, Y.W. and X.Y.; software, R.F.; validation, F.L. and R.F.; formal analysis, X.Y.; investigation, Y.W.; writing—original draft preparation, Y.W. and R.F.; writing—review and editing, F.L. and X.Y.; funding acquisition, Y.W. All authors have read and agreed to the published version of the manuscript.

Data availability

Data cannot be shared openly but are available on request from F.L. (fangzhouli@lpubatangas.edu.ph).

Competing interests

The authors declare no competing interests.

Ethical approval

This article does not contain any studies with human participants and/or animals performed by any of the authors.