News
On June 2, 2026, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim met online to present and discuss the current progress of their work within the X-Quality project.
INSA Strasbourg
Slimane presented recent progress on anomaly detection and counterfactual explanations for manufacturing quality enhancement using the CETIM dataset. A unified multi-branch framework for real-time multi-rate CNC anomaly detection was introduced, combining classical machine learning and deep learning approaches for the analysis of multi-sensor manufacturing data, including vibration, power consumption, and RMS vibration signals. The proposed anomaly detection approach has been accepted for publication at the DEXA Conference 2026.
Furthermore, a counterfactual explanation framework based on an LSTM-VAE model was presented for intelligent product quality enhancement in CNC machining. The method generates realistic counterfactual sensor profiles that transform anomalous parts into normal-quality parts by reducing reconstruction error below an anomaly threshold. The approach enables the identification of actionable root causes influencing product quality, such as excessive tool overhang and feedrate changes caused by operator overrides.
The framework was evaluated on part 62, operation 160, using the GENO-TOPSIS optimization approach and compared with state-of-the-art counterfactual explanation methods. Results demonstrated that the generated counterfactuals provide interpretable process adjustments that reveal how manufacturing parameters need to change to convert anomalous parts into normal-quality products. Additionally, the approach links machine learning explanations with manufacturing knowledge graphs and ontologies to enable semantic interpretation of anomalies and root cause analysis.
Furtwangen University
Rudolf presented the completed optimization of Dynamic LIFT, introducing the new method called Interval-based Dynamic LIFT (IDLIFT). The method replaces Temporal Truth Tables (TTTs) with the Interval-based Temporal Event Table (ITET), enabling a more expressive representation of manufacturing events by preserving duration, supporting multiple event occurrences, and explicitly modeling event intersections.
To validate the approach, a systematic evaluation was performed using 18 abstract Dynamic Fault Tree (DFT) structures. Synthetic datasets were generated under varying temporal event probabilities and confidence levels, and reconstructed DFTs were evaluated by comparing triples consisting of nodes, gates, and child relations. The evaluation prioritized recall, as missing true failure mechanisms was considered more critical than generating superfluous relations.
Results showed that IDLIFT consistently outperformed the original Dynamic LIFT approach, achieving substantially higher reconstruction recall across all experimental settings despite increased structural complexity and variation. Furthermore, comparisons on event logs demonstrated that Dynamic LIFT tends to learn deeper and partially incorrect structures due to TTT limitations, whereas IDLIFT produces more temporally consistent and meaningful fault structures.
Finally, Rudolf presented an orchestration concept for combining multiple X-Quality methods. In the proposed workflow, an LSTM-AE first detects anomalous parts, counterfactual explanations identify influential signals and required parameter changes, Dynamic Fault Trees explain causal propagation and root causes, and RAG-based document validation supports the verification of identified relationships using technical documentation.
On May 5, 2026, the consortium partners from Rouen and Furtwangen met online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen
Gabriel presented progress on the RAGTree and NeoOLAF frameworks. The benchmarking of RAGTree was completed across multiple gold-standard datasets, and the associated KBS survey paper was finalized. The evaluation was extended with analyses of computational cost, execution time, and estimated carbon footprint. Furthermore, the literature review was broadened to include Semantic RAG, Ontology-Guided RAG, Knowledge Graph RAG, and Agentic RAG approaches.
For NeoOLAF, Gabriel reported that the framework reached a stable implementation stage. Ablation study infrastructure was completed to evaluate the contribution of individual modules and processing layers. NeoOLAF was applied to the XQuality Machine32 dataset, generating ontology and knowledge graph artifacts from industrial technical documentation. Additionally, fault description data were processed to investigate the alignment between automatically extracted semantic structures and Rudolf’s fault tree representations. Future work includes evaluating NeoOLAF on five datasets from different domains, integrating non-gold-truth evaluation metrics, and adding multilingual relation extraction capabilities.
Léa presented an ontology-based approach for identifying and explaining quality issues using the XQuality dataset. The work focused on Rudolf’s annotated events from operation OP120. A quality assurance ontology was developed by extending an existing Industry 4.0 context ontology with quality control outcomes and product traceability features. The ontology was then populated with manufacturing data by defining thresholds to detect abnormal machine situations, identifying threshold violations and missing events, and linking machine situations to manufactured products and quality outcomes.
To determine potential root causes, concept induction techniques were applied to ontology individuals. The results identified several highly relevant concepts associated with feed rate reduction, including middle AI1 behavior, middle FZ situations, and spike-down FY patterns. These findings demonstrate the potential of ontology-driven concept induction for explainable quality diagnosis and root cause identification in manufacturing.
Furtwangen University
Rudolf presented improvements to Dynamic LIFT for learning Dynamic Fault Trees (DFTs) from manufacturing event data. A key limitation of Temporal Truth Tables (TTTs) was identified: they cannot adequately represent temporal behavior due to missing duration information, lack of support for multiple event occurrences, and persistence assumptions. To address this, an Interval-based Temporal Event Table (ITET) was introduced, preserving event duration, explicitly modeling overlaps, and supporting repeated occurrences.
Furthermore, Dynamic LIFT was optimized by replacing TTTs with ITETs, removing the greedy search strategy in favor of a continuous full search for the best event subset and gate combination, and introducing candidate filtering to remove irrelevant features. Structural comparisons showed that DFTs learned with ITET produced more meaningful and temporally consistent explanations than TTT-based approaches. Future work includes preparing datasets for publication, orchestrating DFT explanations with additional explainability methods, and finalizing a publication on the improved IDLIFT approach.
On April 7, 2026, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim met online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen
Gabriel presented the benchmarking of the RAGTree framework, evaluating 12 RAG-based approaches for document-level relation extraction across multiple datasets. Results show that knowledge graph-based and agentic methods perform best in structured scenarios, while simpler LLM-based approaches are less effective. He also reported on the application of the OLAF ontology learning framework to industrial X-Quality documents, which generated a large ontology but revealed issues such as noisy extractions and limited structural consistency.
INSA Strasbourg
Slimane presented work on counterfactual explanations applied to the CETIM use case. Two main categories were explored: genetic-based approaches (e.g., NSGA-II with TOPSIS) and model-driven techniques using gradients and controlled perturbations.
In the CETIM scenario, a relationship between feed rate and vibration behavior was identified. Specifically, reducing the feed rate increases RMS vibration levels. This supports counterfactual explanations showing how adjusting vibration-related parameters can shift production from abnormal to normal states, highlighting feed rate as a key controllable factor.
Furtwangen University
Rudolf presented a pipeline for lightweight event abstraction from time series data. Continuous signals (power, force, vibration) are discretized into interpretable states using statistical thresholds derived from reference parts. Extracted events (with timestamps) are stored and transformed into Temporal Truth Tables (TTTs), which are then used to learn Dynamic Fault Trees (DFTs) via Dynamic LIFT. The approach was applied to 16 operations and 3 issue types, resulting in 48 DFTs that explain failure patterns in the event space.
On March 3, 2026, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim met online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen
Léa presented a pipeline to detect abnormal situations using stream reasoning over RDF data streams and SPARQL queries. Detected situations are stored in a triple store and analyzed with concept induction to identify potential root causes. Experiments on a smart farming dataset showed that disease risks can be linked to environmental conditions such as high humidity, low temperature, and reduced sunlight exposure.
Gabriel presented work on ontology learning and causal relation extraction. Progress was reported on benchmarking the RAGTree framework for extracting causal relations from manufacturing documents. In addition, the OLAF ontology learning framework was evaluated on technical documentation, automatically generating domain-specific concepts and relations. Based on these results, a new modular framework called NeoOLAF is being developed.
INSA Strasbourg
Slimane presented experiments on a CNC screw manufacturing dataset with multi-sensor time-series data. A preprocessing pipeline was developed to handle heterogeneous sensor sampling rates, and several anomaly detection models were evaluated, including classical machine learning and deep learning approaches.
Furtwangen University
Rudolf presented the current status of a joint paper developed by all project partners. The paper proposes an orchestration framework that combines multiple explanation methods for quality diagnosis, integrating sensor data, expert-annotated event logs, and technical documents to generate structured quality assurance reports.
Cetim
Cetim presented the current machining experiments and demonstrated the acquired signals, including cutting forces, spindle power, vibration, and acoustic data. The partners discussed which experiments should be used as reference data for further analysis, and a metrology report will be provided soon.
On February 3, 2026, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim met online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen presented progress on knowledge management, stream reasoning, and causal knowledge extraction.
Léa reported on the development of the global architecture and the application of stream reasoning to detect abnormal situations using sensor data from a smart farming dataset. Different types of phenomena were defined and monitored, including water stress, heat stress, vegetation stress, and disease risk, based on environmental variables such as soil moisture, temperature, humidity, NDVI, and sunlight exposure. Thresholds were determined using statistical indicators such as means, quantiles, and standard deviations. SPARQL queries were used to identify relevant conditions, and the results were integrated into a triple store. In addition, concept induction techniques were applied to determine potential root causes from detected situations. This process included feature selection, system configuration, root cause identification using RCIQ 4.0, concept induction via ECII, metric extraction, weighted-sum scoring, and ranking of candidate concepts.
Gabriel presented work on LLM- and RAG-based causal relation extraction. The objective is to establish a unified survey and benchmarking framework for semantic-guided and explainable causal extraction from manufacturing documents. The technical pipeline is fully implemented and includes dataset preparation, ontology integration, preprocessing, retrieval modules, LLM-based relation extraction, and evaluation against ground truth annotations. Initial benchmarking results indicate that LLM-only baselines perform well on constrained datasets but struggle with recall and stability on larger, noisier corpora. Semantic-guided approaches improve precision and consistency, while agentic methods are currently being evaluated. Ongoing work focuses on consolidating benchmarking results, finalizing a survey article, and integrating ontology learning workflows with the RAG-based pipeline.
INSA Strasbourg presented an overview of the CNC machining dataset provided by CETIM. The dataset describes the production of screws through 16 sequential operations, including bar feeding, turning, grooving, threading, drilling, boring, milling, engraving, and cutting-off. Multi-sensor time-series data are recorded for each operation, including cutting forces, spindle power, vibration signals, and contextual machine parameters. Sampling rates vary significantly, ranging from 20 Hz for RMS vibration values up to 50 kHz for raw vibration signals. The dataset structure includes per-part folders, per-operation recordings, and multiple acquisitions per operation. The presentation highlighted the diagnostic relevance of different sensor modalities, the multi-rate sampling architecture, and the distinction between normal production data and controlled abnormal production scenarios with intentionally varied parameters. These datasets are intended to support predictive maintenance, anomaly detection, and process analysis tasks.
Furtwangen Universityreported progress on conceptual work for a joint scientific publication involving all project partners. The proposed contribution focuses on structuring and combining complementary explanation approaches for quality diagnosis in manufacturing. The central idea is that quality assurance requires answers to different diagnostic questions, which cannot be addressed by a single explanation method. Instead, the project aims to orchestrate multiple explanation perspectives into a structured diagnostic report. These include preventive insights, pattern-based insights, causal explanations, and knowledge-based justification. The working title under discussion is “Orchestrating Complementary Explanation Approaches for Quality Diagnosis in Manufacturing.”
On January 13, 2026, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim met online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen reported progress on knowledge management and explainable diagnosis.
Léa presented her ongoing work on instance deletion in ontologies, motivated by the need to avoid overpopulation of knowledge bases in continuous and real-time quality monitoring scenarios. The work focuses on defining principled deletion strategies to remove obsolete or low-value instances while preserving knowledge base consistency. Several approaches from the literature were reviewed, including syntactic deletion, semantic erasure, metaproperty-guided deletion, trigger-based approaches, and data-quality-driven strategies. The analysis highlighted trade-offs between simplicity, computational cost, and semantic correctness. For the X-Quality use case, metaproperty-guided deletion was identified as particularly relevant, as it allows domain-specific deletion policies aligned with engineering practices. Future work will focus on the design of a unified framework for instance deletion and its evaluation on representative datasets.
INSA Strasbourg presented progress on data analysis, preparation, and anomaly detection for machining process monitoring. An initial exploratory analysis of the CETIM dataset revealed limitations related to incomplete parts, sparse output labels assigned only at the end of the process, and highly heterogeneous sensor sampling frequencies. To address these challenges, a data preparation pipeline was developed, including signal resampling to a common reference frequency, multivariate time-series construction, and sliding-window segmentation. Two complementary anomaly detection strategies were investigated: an autoencoder-based approach operating directly on multivariate time-series windows, and a feature-based approach using statistical descriptors combined with an Isolation Forest. Both methods were trained on data from Part 5, assumed to represent normal operating conditions, and evaluated on data from Part 7. Experimental results showed consistent detection of anomalous behavior toward the end of the machining operation across both approaches, suggesting a genuine deviation in process behavior. While the precise cause of the anomalies remains unknown due to missing ground truth and post-process quality measurements, the consistency between methods significantly narrows the investigation scope. Further validation with CETIM and additional labeled data is planned.
Furtwangen University reported progress on Dynamic Fault Tree (DFT) generation using the Dynamic LIFT algorithm. The focus was on analyzing the computational complexity of DFT generation and its impact on practical applicability. The team discussed the use of reduced-depth DFTs, in particular depth-1 DFTs, as an efficient alternative to full DFTs. While full DFTs provide deeper explanatory structures, depth-1 DFTs are theoretically logically equivalent in terms of predictive capability, as deeper levels primarily serve explanatory purposes. Preliminary results indicate that depth-1 DFTs can be generated significantly faster while achieving comparable predictive performance. Ongoing work focuses on systematically comparing predictive performance across different DFT depth levels and assessing robustness under noisy, statistically derived causal relations.
Cetim presented the current status of part production and data acquisition. Data acquisition is ongoing, with the first manufactured parts identified as non-conforming. Production has since transitioned to good parts, and future production will intentionally include parts with quality issues to support downstream analysis and validation. These datasets will serve as the basis for further modeling, anomaly detection, and explainable diagnosis within the project.
On December 2, 2025, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim met online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen reported significant progress in two main research directions.
Léa presented the experimental evaluation of the normalized (structural) depth score. The evaluation compared three configurations: the original ECII algorithm, ECII combined with the structural depth score, and ECII combined with an ontological depth score. Six benchmark datasets from the literature (Forte, Yingyang, Poker, Moral, Trains, and ADE20k), ranging from 150 to nearly 11,000 axioms, were used. The experiments, conducted with the Pellet reasoner on a high-performance machine, showed similar runtime between ECII and ECII with structural depth, comparable accuracy across all approaches, and a similar number of solutions between the structural and ontological variants. The structural depth score was validated through strong alignment with ontological hierarchies on Poker and Yingyang, and moderate correlations on Forte and Trains due to semantic modeling choices such as disjointness and mutual exclusivity. Results on larger datasets are ongoing.
Gabriel reported progress on consolidating the state of the art for LLM- and RAG-based causal relation extraction in manufacturing. Five datasets were fully retrieved and normalized into a unified JSONL schema, including text processing and entity extraction, while CausalBank is nearing completion due to its scale. Ontologies were converted into Turtle format to ensure interoperability. A structured pipeline was defined, covering dataset normalization, ontology and knowledge graph preparation, entity recognition and linking, retrieval modules for multiple RAG variants, LLM-based relation extraction, and alignment and faithfulness validation. Steps covering setup, normalization, ontology preparation, entity linking, and relation extraction were completed, with twelve RAG variants selected and the first three implemented. The project now enters the evaluation phase, with upcoming work focusing on large-scale normalization, comprehensive benchmarking, survey submission, OLAF to NeoOLAF integration, multilingual extraction, and preparation for processing original X-Quality documents.
Furtwangen University reported progress on event-based preprocessing and Dynamic Fault Tree generation. The team evaluated extracted events using the SKAB time-series dataset, including accelerometer, current, pressure, temperature, thermocouple, voltage, and volume flow rate signals. Semantic events were extracted from signals using the RBAθ method and transformed into Temporal Truth Tables (TTTs). The resulting representations were applied to an XGBoost classifier and to Dynamic LIFT for DFT generation. Preprocessing focused on anomaly-centered episodes, comprising 19 pre-anomaly windows and one anomaly-onset window per anomaly period, resulting in 320 episodes with a highly imbalanced class distribution (304 normal, 16 anomalous). Initial model evaluations were performed without hyperparameter tuning. A first Dynamic Fault Tree was generated. Next steps include further optimization of semantic event extraction, including windowing and threshold parameters, and the introduction of confidence levels for DFT generation.
Cetim presented the current status of part production and data acquisition. A first dataset has been uploaded to the cloud. Part production is ongoing and will continue throughout December, with data updates planned every week. Metrological setup of the coordinate measuring machine is scheduled for the following week, with dimensional measurements planned for December. A first version of the technical report is being finalized and will be shared with project partners.
On November 4, 2025, the consortium partners from Rouen, Strasbourg, Furtwangen, and Cetim convened online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen reported two major research directions.
Léa Charbonnier presented the formal definition and properties of the newly developed normalized depth score for ranking candidate class expressions based solely on their structural specificity. The score is defined over ALC expressions, is deterministic, invariant to ACI flattening, monotonic, and normalized between 0 and 1. It replaces reasoning-based depth evaluation with a scalable syntactic method while preserving meaningful relative rankings. An illustrative example demonstrated alignment between the score and the ontological hierarchy.
Gabriel Henrique Alencar Medeiros presented ongoing work toward a semantic-guided framework for explainable causal relation extraction in manufacturing. The team is establishing a benchmark of datasets and ontological resources and designing workflows for entity recognition, ontology linking, and embedding generation. A comparative study between RAG, GraphRAG, and OntoRAG showed that OntoRAG improves explainability through conceptual reasoning. Upcoming work includes ontology integration, retrieval configuration, and evaluation using metrics such as Entity F1, Relation F1, alignment quality, and faithfulness.- Furtwangen University presented progress on evaluating preprocessing strategies for event extraction. Two models were developed to compare datasets, revealing that many do not yield reliable predictive results. The SKAB dataset was identified as a suitable candidate for further experiments. The team selected the Ramp Behaviour Analysis (RBAθ) method to extract events from time-series signals, converting sensor data into compact event representations used for temporal Boolean transformation. RL was applied to optimize threshold-values, significantly improving model performance. Current challenges preventing Dynamic LIFT from generating DFTs include anomaly persistence and active gaps in event sequences. Addressing these issues defines the next steps.
Cetim initiated its data acquisition phase, demonstrating the first successful collection. Full acquisition is planned to conclude in early December, after which new data will be made available weekly. Technical documentation for the machinery has been provided.
On September 30, 2025, the consortium partners from Rouen, Strasbourg, and Furtwangen convened online to present and discuss the current progress of their work within the X-Quality project.
INSA Rouen presented two ongoing research directions. Léa Charbonnier reported on the concept induction work, recently published at RuleML + RR 2025, introducing a novel scoring approach that balances coverage, precision, and ontology hierarchy depth. The method identifies logical root causes for product defects using the AI4I 2020 predictive maintenance dataset. A new syntactic depth calculation has been developed to replace reasoning-based evaluation, improving scalability while preserving relative accuracy. Future work will focus on supporting subclassed situations and expanding applicability to any concept.
Rouen also advanced stream reasoning, adding timestamped observations, deletion mechanisms for outdated instances, and the use of a triple store (GraphDB) for synchronized ontology and data processing.
Gabriel Henrique Alencar Medeiros introduced his post-doctoral research on explainable root-cause detection from technical documents, which extracts ontology-guided causal relations from maintenance logs and reports. The approach integrates state-of-the-art methods in retrieval-augmented generation (RAG), ontology-based reasoning, and language models. Upcoming work includes comparative experiments between LLM + RAG methods and ontology-driven baselines.- INSA Strasbourg will continue its contributions in analog and multimodal data analysis for manufacturing quality, aligned with previous work on sensor and image-based defect characterization.
Furtwangen University reported progress on learning Dynamic Fault Trees (DFTs) from data using the Dynamic LIFT algorithm applied to Temporal Truth Tables (TTTs). The team emphasized the importance of preprocessing and discretizing time-series data into symbolic event sequences to enable DFT learning. Selection criteria for the discretization method include unsupervised operation, temporal order preservation, simultaneity, non-parametric formulation, and scalability. The next steps involve selecting an appropriate change-point detection method, preprocessing real datasets, and applying Dynamic LIFT to evaluate the learned DFT structures.
On June 20, 2025, the consortium partners from Rouen, Furtwangen, Strasbourg, and Cetim convened in Rouen to present and discuss the current status of their work in the X-Quality project.
INSA Rouen updated on their two main contributions: quality issue detection via stream reasoning and quality cause determination using concept induction. The concept induction method employs logic-based learners (ECII), operating over ontologies with a reasoner. They developed a new ranking score for concept evaluation that balances coverage, precision, and ontology hierarchy depth. Experiments using the AI4I 2020 predictive maintenance dataset demonstrated improved performance in identifying logical root causes for defective products. Future work will include contextual concept enrichment.
- NSA Strasbourg presented their analog and multi-modal manufacturing datasets collected at the Future Factories Lab. These datasets contain time series and image data from an industrial assembly line, with defects introduced manually. Their models generate counterfactual explanations based on time series data, supported by domain ontologies. They use both gradient-based and genetic approaches for generating counterfactuals. Ongoing work investigates using LLMs to support interpretation and reduce ontology dependency.
Furtwangen University demonstrated progress in extracting Temporal Truth Tables (TTTs) for Dynamic Fault Tree learning. They use change detection (e.g. increase, decrease) on sensor features from the MS Azure Predictive Maintenance dataset. Reinforcement Learning (Q-learning) is applied to optimize the thresholds for Boolean transformation. Using Boolean features, they train LSTM classifiers and use the F1-score as reward signal. Suboptimal thresholds yield inconsistent DFTs. Next steps involve defining state- and pattern-based events, including expert-defined features, and extending the TTT logic to tolerate gaps between event occurrences.
- Cetim shared updates on their machining data collection. Using a stainless-steel workpiece machined via turning and milling, they acquired data with and without quality defects across 100 parts. Various sensors (spindle power, cutting forces, vibration, audio) captured signals at 40–55 kHz. The dataset includes multistep operations with different tools. The data may be made publicly available later.
This date, the Consortium partner from Rouen, Furtwangen and Cetim came together to discuss their progress in the X-Quality project.
INSA Rouen presented advancements in their Concept Induction approach. They developed a new scoring method for rule evaluation, which combines coverage, rule precision, and hierarchy, offering a more balanced assessment of rules. The new method has shown improved results on a predictive maintenance dataset, and an article for ISWC is in preparation.
Furtwangen University shared their latest developments in applying Temporal Truth Tables (TTTs) for predicting failures in a black-box model. They utilized a thresholding approach based on C4.5 for TTT creation and are currently testing their method with the Microsoft Azure Predictive Maintenance dataset.
The next consortium meeting is scheduled for June 20, 2025, in Rouen.
This date, all Consortium partner came together to discuss their progress in the X-Quality project.
INSA Strasbourg applied their counterfactual explanation methods to a real-world manufacturing dataset. Using a domain-specific ontology, they identified critical sensor signals from robotic arms responsible for introducing defects. Their results offer intuitive insights into how small changes can shift a product from normal to defective classification.
INSA Rouen presented new results from their concept induction approach. They improved how rules are created and evaluated using clear scores like coverage and F1, and added a way to measure how detailed each rule is. The updated method creates simpler and more reliable rules within their ontology. Next, they plan to test it on more realistic machine situations and production setups.
Furtwangen University reported on their latest publication:
“Learning Dynamic Fault Trees from Data for Interpretable Failure Prediction” (submitted to the journal Engineering, special issue on Intelligent Manufacturing). They extended the LIFT algorithm to learn Dynamic Fault Trees (DFTs) using Temporal Truth Tables (TTTs) and statistical tests for inferring PAND and SEQ gates. Their findings highlight the trade-offs between dataset noise and the strictness of statistical confidence thresholds.Cetim is currently preparing machines for upcoming experiments. Once setup is finalized, they will present their data acquisition system before beginning data collection across multiple processes (e.g., turning, milling).
Key topics and updates:
INSA Strasbourg shared progress using two new industry-grade datasets collected at the Future Factories Lab, University of South Carolina. The datasets simulate defects introduced during an assembly process. A DL model trained on the sensor data achieved 91% accuracy in predicting product quality.
INSA Rouen introduced an approach for concept induction based on description logic. The goal is to automatically identify classes of faulty situations by analyzing previously detected anomalies in their ontology framework.
Furtwangen University tested their LIFT-based algorithm using abstract FTs. The evaluation highlighted the importance of confidence thresholds and balanced datasets for generating interpretable and precise FTs.
On January 21st, 2025, the X-Quality consortium held its first meeting of the year in an online format. The partners from INSA Rouen Normandie, INSA Strasbourg, Furtwangen University, and Cetim came together to share their current progress and align on the next project steps.
INSA Strasbourg presented two new methods for generating counterfactual explanations for multivariate time series and introduced the CEVO ontology to ensure consistency of these explanations.
INSA Rouen shared progress in recovering data from an educational production line at Schneider Electric, including a rich technical documentation package and fault simulation capabilities.
Furtwangen University showcased their Fault Tree Surrogate Model for explaining ML predictions. This approach supports model adaptation over time by combining real and simulated data, aiming to evolve from static to dynamic FTs.
The meeting concluded with technical discussions around data interpretation, defect grouping, and demonstrator requirements. The next consortium meeting is scheduled for March 4th, 2025.
In the recent XQuality project meeting, we presented updates on our work and discussed progress across different research areas.
INSA Strasbourg:
Slimane is focusing on the roughing dataset derived from the DQ-Meister Project as a use case. Amel, on the other hand, is developing a Symbolic AI-based approach for Explainable AI (XAI). The approach involves encoding machine learning models using propositional logic to generate contrastive explanations. These explanations aim to highlight alternative sets of facts that would change a decision outcome. Additionally, Amel's work focuses on reducing the size of explanations, improving scalability, and addressing classification problems, with plans to extend the approach to regression problems. The underlying data for this task involves time series datasets.
INSA Rouen:
Léa compared two ontologies: MALFO and FOLIO. MALFO provides an ontological understanding of malfunctions by creating a formal taxonomy that ensures semantic interoperability of engineering knowledge. It builds upon a pre-existing causation ontology. In contrast, FOLIO focuses on describing anomalies and their root causes, particularly for Failure Mode and Effect Analysis (FMEA).
Furtwangen University:
Rudolf worked on combining continuous-valued datasets with categorical data for fault tree generation. By applying the C4.5 principle, they extracted thresholds from continuous data and converted it into a Boolean dataset. Using this data, they generated a Fault Tree as a surrogate model using the LIFT algorithm to explain complex machine learning models globally. For interpretability, they emphasized the need to simplify the fault tree and visualize the most relevant branches for predictions locally. Additionally, they discussed updating the fault tree dynamically with each input to adapt over time.
Cetim:
Roger Busi visited the meeting, because he is responsible for the data acquisition at Cetim. Roger presented their data acquisition system and provided insights into the types of data available for machining operations.
This day, some of the Consortium partners attended the URAI conference in Offenburg to present their published paper with the title: "Artificial Intelligence for Quality Assurance and Troubleshooting in Industry".
In the recent XQuality project meeting, we presented updates on our work and discussed progress across different research areas.
INSA Strasbourg: The team is developing a Counterfactual Explanation Library for multivariate time series and normal datasets using methods for generating, mutating, evaluation and visualization. Initial tests were conducted on the Adult Income Dataset to predict income categories and are being expanded to electrical power consumption datasets.
INSA Rouen: The MALFunction Ontology (MALFO), a formal taxonomy for engineering malfunctions, was highlighted. This ontology incorporates definitions for terms like failure, fault, and root cause and reuses an existing ontology of causation to semantically link engineering knowledge. MALFO is being applied to create a quality assurance ontology, providing semantic interoperability for analyzing malfunctions.
Furtwangen University: The team presented their work on Learning Fault Trees (LIFT) from observational data. The LIFT algorithm constructs fault trees iteratively, starting from a Top Event (TE), by statistically evaluating relationships among events. Preliminary results show LIFT effectively reconstructs fault trees, with plans to test on real-world datasets to provide interpretable explanations for system predictions.
Cetim: A meeting is scheduled for November 20th to focus on data collection, aiming to align datasets with consortium needs for predictive modeling and quality assurance applications.
In the X-Quality project meeting, we shared new advancements in our respective areas of research.
INSA Strasbourg introduced a new drilling dataset for explainable AI (XAI) applications, enabling multi-class classification for failure detection. They addressed class imbalance by adjusting the loss function to penalize misclassifications of rare failure events more heavily, improving the model's sensitivity to failures.
INSA Rouen is applying Description Logics (DL) to enhance the quality assurance ontology, aiming for formal and expressive descriptions of quality issues. DLs provide robust semantics, reasoning capabilities, and interoperability with OWL standards, allowing for reusability and modularity in ontology development. This approach is intended to improve linking quality issues to abnormal situations in a structured way.
Furtwangen University presented an approach to using Fault Trees (FT) as a surrogate model for XAI, transforming complex model explanations into understandable fault trees. By using fault trees as recommendation systems, HFU aims to provide actionable insights into parameter adjustments that could prevent failures. Future steps include selecting XAI methods that meet FT requirements, such as rule-based and model-agnostic approaches, and developing metrics to evaluate fault trees.
Further Discussion: The team discussed potential synergies, such as simulating quality issues at another campus to generate new datasets and exploring collaborations between Amel’s rule-based explanations and Rudolf’s FT transformation approach.
In the X-Quality project meeting, we discussed progress across various focus areas and outlined plans for upcoming work.
INSA Strasbourg presented an approach for mining chronicles to identify abnormal drops in product quality, initially tested on lithium-ion batteries and then applied to a roasting machine dataset. They preprocess data to form an input matrix of 17 features and 60 time steps, employing the PrefixSpan algorithm to identify patterns. However, the process is resource-intensive, and they are exploring optimization methods.
INSA Rouen continued developing a stream reasoning approach with RDF Streams and ontological modeling. They introduced an ontology based on F. Giustozzi’s Context ontology, incorporating quality assurance concepts and spatiotemporal characteristics to support quality issue tracing. Upcoming work includes enriching the ontology with quality assurance data and integrating explanations from other partners.
Furtwangen University has been organizing manufacturing failure data into structured fault trees within Excel, creating a systematic knowledge base of failure causes and solutions. They also shared a grinding dataset from the DQ-Meister Project, featuring 36 recordings of grinding experiments with 67 categories, to enhance machine learning for fault detection.
Cetim introduced a milling dataset with features such as force, power, and tool wear, covering 800 parts. They discussed generating additional data through test campaigns for non-conforming parts to support predictive models.
Further Discussion: The consortium emphasized working on a unified use case by the end of 2024 and publishing a project dataset as an additional outcome.
During the X-Quality workshops in Strasbourg, the consortium divided into groups to focus on distinct use cases, data requirements, and research questions.
Group Léa explored integrating machine and product data for quality monitoring. Their focus included observing machine and product interactions and ensuring temporal alignment. Key data needs include machine types, operational parameters (like speed and vibration), extra sensor data, and detailed failure descriptions. They aim to understand how to merge product and machine data to describe quality issues and link them to causes.
Group Amel & Slimane concentrated on anomaly detection for quality degradation, such as battery State of Charge and roasting machine quality. They are working with time series data, both proprietary and public, to generalize anomaly detection and make approaches scalable for large datasets. Their research questions involve optimizing explanations for detected anomalies.
Group Rudolf focused on using Fault Tree Analysis (FTA) and AI to prevent chatter in milling processes. They aim to identify and address Basic Events (BEs) in real-time, using machine parameters like spindle speed and sensor data for predictive modeling. Their goal is to assess how combining FTA and AI can prevent chatter in machining.
Group Christoph & Cecilia are working on a collaborative use case with Cetim, integrating expert knowledge into an ontology alongside AI/XAI models. The goal is to enable predictions and explanations that help operators detect and understand quality issues, combining data-driven insights with expert knowledge.
Workshop 2 Outcomes: The consortium identified benefits of their approaches for Cetim, such as process optimization through XAI, preventive actions with machine feedback, and enhanced expert support. They identified a need for failure descriptions and data associated with these failures to enhance predictive capabilities and operator guidance.
In the X-Quality project meeting, Cetim Grand Est presented new opportunities in defect detection and quality control. They are exploring vision-based defect detection with MobileNet V2 and Yolo v5/8 for transfer learning. Additionally, they aim to enhance quality control using magnetic fields by revealing scratches with a magnetic spray and assessing via a phone app. Data needs were discussed: Prof. Reich expressed interest in multimodal data, combining images with sensor data, as well as time series data. Suggestions included troubleshooting documents and fault tree analysis to assist with root cause identification.
Use Cases: The consortium discussed the challenges in defining general use cases, highlighting a demonstration using a printer and combining AI with Fault Trees for transparent fault prediction.
Dissemination: The team plans to showcase their work at the URAI 2024 Conference, with Ahmed presenting a poster and creating a new project rollup. The deadline for this is July 15, 2024.
HFU’s Current Approach: HFU continues to develop automatic Fault Tree (FT) generation using computer vision for feature extraction, machine learning classifiers for fault type identification, and predicting Basic Events (BEs) for Top Event (TE) determination. Current work includes using the Steel Plates Faults Detection dataset, which features data on fault types and descriptors, to refine FT learning and enhance fault detection methods.
In the X-Quality project meeting, we shared advancements and current challenges.
INSA Strasbourg developed a Dense-layer model for product quality prediction using the roasting machine dataset, achieving a Mean Absolute Error (MAE) of 16.05, which is considered acceptable within a quality range of 200–500.
INSA Rouen continued work on stream reasoning with RSP4J, enabling simultaneous querying of data streams and linked ontologies. They created RDF collections from query results within time windows, preparing to integrate new detected cases into the ontology. Bibliographic research for an article on quality detection in Industry 4.0 highlighted comparisons of RSP engines and preventive maintenance approaches. An internship using the OLAF platform is planned to explore automated root cause extraction from technical documentation.
Furtwangen University focused on translating Machine Learning models (Decision Trees and Bayesian Networks) into Fault Trees (FT) for data-driven fault analysis. Early experiments using a generated dataset demonstrated that translating from a Decision Tree simplifies the FT structure. Future steps involve translating models into open-PSA format and applying a realistic public dataset for further validation.
In the X-Quality project meeting, we presented updates on various technical approaches and challenges.
INSA Strasbourg worked on analyzing a roasting machine dataset, featuring 15 temperature sensors, humidity, and height data. They resampled data to hourly aggregates and developed regression models (LSTM, CNN) to predict quality outcomes. By comparing time-series data to high-quality reference points, they categorized deviations and mined chronicle patterns to track quality degradation events.
INSA Rouen advanced stream reasoning with RSP4J, linking ontologies to SPARQL queries for enhanced data selection. Although initial results are promising, integrating data in RDF tables remains a challenge. They are also exploring ontology integration techniques, such as alignment and versioning, to synchronize various ontologies, including MASON.
Furtwangen University submitted a paper to the International Conference on Cloud Computing, focusing on combining Machine Learning (ML) with Fault Tree Analysis (FT) for explainable failure detection in cloud systems. Experiments used two datasets (SOFI and SMART) to predict Basic Events (BEs) and subsequently determine Top Event (TE) probabilities, improving interpretability and root cause analysis in fault diagnosis.
In the X-Quality project meeting, the consortium reviewed the latest progress across partners.
Consortium Agreement: All partners have signed the agreement. Vincent Arnoux will distribute a scanned copy and return the originals.
INSA Strasbourg focused on quality prediction models, particularly for lithium-ion batteries, exploring the MIT dataset for SOH and SOC estimation. Challenges include sourcing suitable time series data, essential for understanding temporal quality patterns. They are implementing E-LSTM and CNN-LSTM models and using SHAP and Matrix Profile for explainable data mining.
INSA Rouen is advancing ontology merging using PrOM and switching from OntopStream to RSP4J for reliable RDF stream processing. They are refining the integration of ontologies and anomaly detection on sensor data streams, creating link ontologies with improved class hierarchy and consistency.
Furtwangen University is exploring Machine Learning (ML) and Fault Tree (FT) combinations for defect detection in manufacturing. Five potential methods combine ML predictions with fault trees, from predicting specific events to generating simplified fault trees from observational data. These approaches aim to enhance defect detection and root cause analysis while reducing false positives and negatives.
In the X-Quality project update, we shared progress on various use cases and methodologies.
INSA Strasbourg has worked with a dataset regarding lithium-ion batteries, focusing on estimating the State of Health (SOH) of cells to assess quality. They tested E-LSTM and CNN-LSTM models using the MIT battery dataset, targeting SOH estimation based on current, voltage, and temperature data. They defined abnormal SOH decreases and established levels of degradation for monitoring.
INSA Rouen is working on defect detection and explanation across machine and production line levels, combining data-driven and knowledge-based approaches for adaptive anomaly detection. Their work includes using FLAGS for root cause analysis on sensor data and testing frameworks like C-SPARQL for querying streaming data, as well as OntopStream for heterogeneous data sources. Plans include hiring an intern to explore automated root cause extraction from technical documentation using OLAF.
Furtwangen University focused on a Fault Tree Analysis (FTA) use case, utilizing the GC10-DET dataset for steel defect detection. They created a fault tree representing root causes, with ML model confidence values as probabilities to calculate defect likelihood. This FTA serves as a surrogate model, providing a conceptual proof for further applications in manufacturing.
In the X-Quality project meeting, we presented our progress. INSA Strasbourg has hired two researchers: Amel Hidouri, working on explainable data mining for failure analysis, and Slimane Arbaoui, focusing on explainable deep learning models for degradation prediction using pattern mining. INSA Rouen hired PhD student Léa Charbonnier, contributing to hybrid AI models, and is planning an intern recruitment to explore interpretable models.
Rudolf Hoffmann (Furtwangen University) shared the publication of a survey paper in MDPI's Electronics journal, detailing a systematic review of AI and XAI for Visual Quality Assurance (VQA) in manufacturing. His review identified trends in VQA practices, methods, and applications, highlighting gaps in XAI adoption and emphasizing the need for transparent AI methods in manufacturing quality assurance.
Moreover Rudolf provided 8 possible use cases for VQA in manufacturing that he found in the literature.
In the X-Quality project online meeting, we discussed the Data Management Plan (DMP), prepared by Cecilia Zanni-Merk on the Opidor platform, which allows collaborative editing. Key considerations included data privacy, defining storage timeframes, and ensuring data security through encryption and storage on HFU servers.
Rudolf Hoffmann shared updates on his Systematic Literature Review (SLR) examining AI in quality management for manufacturing. The review identified 59 relevant studies, with a focus on visual quality control in metals, electronics, and additive manufacturing. Few studies utilized Explainable AI (XAI), highlighting a gap this project aims to address.
Additionally, Rudolf conducted experiments using GradCAM with a VGG16 model on a steel defect dataset. GradCAM’s visualizations helped identify defect regions, demonstrating practical XAI applications in defect detection.
The X-Quality project started with a kickoff meeting in Strasbourg, bringing together all consortium partners. The project aims to develop innovative semantic XAI methods to explain quality issues in manufacturing processes. The core objectives include integrating time-series and text data mining, creating a hybrid reliability model, and applying these approaches to a real industrial use case for explainable diagnostics. Consortium members, including Furtwangen University, INSA Strasbourg, INSA Rouen, and Cetim Grand Est, discussed use cases, administrative tasks, and hiring plans. The meeting set clear project timelines, roles, and responsibilities.