With the advent of Industry 4.0 and the ability to monitor production lines in real time, new possibilities in terms of product quality management have emerged. One of them is Quality 4.0 [1], an extension of Industry 4.0 to quality assurance which allows to combine quality data with data from other sources (machine sensors, manufacturing, etc). As industries need to reduce risks and costs and ensure the quality of products, predictive models can use data collected by machine sensors to anticipate breakdowns or manufacturing errors. However, these models are not all inherently explainable and can entail huge difficulties in tracking root causes of anomalies [2].
This work is part of the XQuality 1 project whose main goal is to implement intelligent and automated quality assurance to assist operators in manufacturing companies. We propose to use a hybrid approach to detect and explain quality loss by reasoning over an ontology that integrates all the available knowledge such as results of predictive models, technical documentation and expert knowledge.
Quality assurance in manufacturing companies is an essential process for ensuring that products meet standards. It contributes to customer satisfaction and reduces the costs of defects. To tackle quality issues, we propose to adapt the approach presented in [3]. The idea is to detect situations that may lead to quality losses by observing products and tracking abnormal sensor values. Expert knowledge will allow the identification and assessment of the root cause that led to a detected abnormal situation.
The framework proposed in Figure 1 is based on [3]. Three modules are used to detect and explain quality issue situations reasoning over an ontology: Translation, Temporal Relations and Cause Determination. This reasoning is performed in real time (stream reasoning) or offline (classical reasoning over an ontology). To detect quality issues, an ontology based on the Context Ontology described in [4] is used. This ontology is composed of three core ontologies (Sensor Ontology, Time Ontology, Location Ontology and Situation Ontology) and three domain ontologies (Resource Ontology and Process Ontology). Therefore, we propose to extend the Situation Ontology with a Quality Assurance Ontology. As the Situation Ontology concerns only situations on machines, we want to extend it with quality issues situation detection on products. The Quality Assurance Ontology is based on TOVE Traceability Ontology [5] which provides representation to identify and trace a quality problem. As TOVE is a core ontology, we mainly use the idea of traceability between products and activities. This allows to link products to the machine that produced or modified them when and where the default occurred.
The Translation module is responsible for collecting sensor data and converting it to RDF streams thanks to Stream Generators. This component performs a semantic enrichment of raw data using the concepts and relations defined in the ontology. One Stream Generator is used per sensor.
The Temporal Relations module is used to explore the data. RDF streams are continuously queried into the Stream Reasoner using the ontology to contextualize the streams. The queries represent different quality issue situations to detect. For this component, we used RSP4J [6] to query the streaming data. The data from the detected situations is then formatted and put into the ontology.
The Cause Determination module identifies the root cause and explains the problem. The Reasoner is used over the ontology to check the consistency of the ontology and then, infer new knowledge. The cause of the quality problem can then be determined using different classifiers as proposed in [7].
In this section, we present an illustrative case study created to test our framework. We consider a manufacturing production line named PL1 composed of two machines M 1 and M 2 which produce products named P 1 , P 2 and P 3 . The production line is equipped with sensors that observe machine and product properties. The sensors collect data on properties in Table 1. Constraints ranging from π 1 to π 8 relate to machines and the other ones relate to products. Abnormal situations that could lead to machine failures are defined from expert knowledge and expressed as a set of constraints in Table 2. Quality issues situations are also defined from expert knowledge and expressed as a set of constraints in Table 3. They describe quality issues detected on products and the associated machine and product constraints.
Data streams are then created by the Stream Generator and continuous queries are performed on them. To do this, the Stream Reasoner is used with an ontology containing information on the production line. Once a situation is detected, is it added to the ontology. An off-line reasoning is done to check the consistency of the ontology. Another Reasoner is then used to provide an explanation thanks to the information contained in the ontology which allows to link defects found with expert information and technical documentation.
A semantic approach to quality loss explanation in the manufacturing industry is presented. Data streams are processed with stream reasoning allowing real-time situation detection. A context ontology is used to help detect quality issues by enriching the information contained in the streams. In future work, explanations of abnormal situations and their causes will be adapted according to the end user.
This work was supported by the French National Research Agency [grant number ANR-22-CE92-0007].