Ishikawa diagrams, also known as fishbone diagrams, are an established tool in the manufacturing domain for conducting root cause analysis. The Ishikawa diagram ontology was developed to explicitly model Ishikawa diagrams as visual artifacts, their encoded knowledge and the process of their creation. While formalization is an important step for making encoded knowledge accessible, another challenge is to establish reasoning mechanisms to provide decision-support to users. In this paper, we present a reasoning pipeline which resulted from a case study concerning fabric fault detection. The reasoning pipeline is intended to be used in conjunction with the Ishikawa diagram ontology.
Ishikawa diagrams, also known as fishbone diagrams, are established tools in the manufacturing domain
for conducting a root cause analysis (RCA) with little to no training [
In the domain of industrial knitting (a subdomain of the textile industry), the vast amount of diferent yarn qualities, designs, and knitting machine characteristics paired with small lot sizes leads to the problem that training documents are incomplete and knitters (who operate circular knitting machines) must develop heuristics to quickly identify and fix fabric faults. In addition, knitters are exposed to diferent problems due to assigned production orders. As a consequence, knitters are not equally trained on the same fabric fault and novice knitters struggle to acquire knowledge missing in the training material.
Broken inlay yarn
Consequently, there is a need to automatically support knitters in the context of fabric fault detection. Researchers have proposed various applications of artificial intelligence to address specific problems, e.g., the automatic visual detection of faults [7, p. 172]. However, the efectiveness of the automated visual inspection is challenged by the consequences of mass customization and novel design choices. For instance, a design with intentionally irregular 3D patterns might visually resemble a fabric fault of a flat-knitted design.
Due to the unreliability of computer vision applications, our collaboration partner relies on regular quality workshops in which the 5-Why method is applied to stimulate knowledge exchange among knitters. However, a number of issues were observed that limit the efectiveness of the workshops. First, knitters tend to list only direct causes instead of longer cause-efect chains. Second, the knowledge exchange is limited to knitters that actually attend the same workshop, which excludes the knowledge of knitters from other shifts.
In agreement with the textile manufacturer, we developed a reasoning pipeline to provide automatic decision support for five fabric faults in the scope of an exploratory case study. The case study consisted of four steps: (1) Acquisition of the ground truths, (2) building a knowledge graph, (3) Implementation of the reasoning pipeline, and (4) Evaluation of the implementation. The fabric faults under investigation can be easily distinguished by an experienced knitter during visual inspection. Samples of the fabric faults are shown in Table 1.
We have interviewed two knitting experts to elicit relevant knowledge on how to identify, analyze, and resolve the occurrence of the five named faults while operating on the knitting machine. As the mode of communication (which includes the expression of technical terminology) was in Turkish for both interviewees, a translator mediated the interviews by translating between English and Turkish. Each interview lasted an hour, which led to a total English audio track of 50 minutes and a Turkish audio track of 70 minutes. In each interview, the experts had to identify and visually describe the fabric fault, describe its root causes, and what actions must be taken to resolve the fault. For the visual inspection, we provided a fabric sample for each fault. Based on the interviews, we have formalized a set of 31 relationships which we consider the ground truths and from which we can directly generate diferent Ishikawa diagrams.
As the Ishikawa diagram is “a guide to concrete action”[3, p. 29], we must add weights to the causeefect relationships to prioritize causes to define the order of necessary actions. The prioritization process follows a behavioral aggregation approach where the priority order might be influenced by the participating group of knitters. As knitters must complete their daily orders, it cannot be ensured that all knitters have the capacity to attend the focused discussion. Also, the elicitation and prioritization process might become less efective if too many knitters at once participate which increases the efort to coordinate the session.
In order to collect individual weights of the cause-efect relationships, we have recruited eight
experienced knitters who provided their perceived frequency of occurrence for each relationship as a
numerical value and as a verbal probability expression (VPE) [
As a result, we have captured eight individual knowledge bases which address subsets of the known knowledge space (here: 31 relationships and their entities) and which form a semi-connected knowledge graph.
Ishikawa diagrams can be formalized and generated from any individual knowledge base or a set of
knowledge bases using the Ishikawa diagram ontology as basis. As the next step, we aimed for an
interactive and complete overview for knitters to explore known cause-efect relationships as the desired
decision support. To infer the relevant causes for a fabric fault, we applied the IDP-Z3 system [
Two limitations of the described implementation must be pointed out. First, the IDP-Z3 system
cannot directly reason on RDF data, but requires the data in FO(· ) syntax which prevents the described
implementation to be used on a larger scale. Nonetheless, there already exists on-going work to make
RDF data interoperable with FO(· ) [
Due to the case study’s small scope, the captured knowledge space was insuficient to meaningfully support knitters in production. Instead, we conducted a SWOT analysis with two knitting experts for the evaluation. One identified strength was that Ishikawa diagrams can be understood by people with diferent roles and technical backgrounds, which helps the communication between these diferent parties. Additionally, the approach can be used to detect knowledge gaps of knitters, for which appropriate training material could then be provided. The provision of an expert knowledge base as decision support might also shorten the training time of novice knitters by manifesting the expert’s mental model. Weaknesses were that knitters require initial training and the help of a moderator before they can correctly apply the new approach. Also, it may not always be possible to arrive at detailed, formalized knowledge for each use case. The main threat for the approach is that new fabric designs and qualities might lead to irregular behavior contradicting the existing knowledge base. In summary, the experts see the greatest benefit in the approach to shorten the training time of knitters by visualizing the mental model of experts for novice knitters to adopt.
We described an exploratory case study with an implemented reasoning pipeline that is intended to be used in conjunction with the Ishikawa diagram ontology to illustrate how Ishikawa diagrams can be made actionable to provide decision support to knitters (or any other users). Additionally, we highlighted current limitations and challenges of such a reasoning pipeline that must be addressed by new tools to harness the full potential of the Ishikawa diagram ontology.
The author(s) have not employed any Generative AI tools.