Semantic Multi-Modal Framework for Early Stage Anomaly Detection in Lithium-Ion Batteries.

This paper presents the Battery Anomaly Ontology (BAO), a framework that combines static knowledge and stream reasoning capabilities to enable interpretable anomaly detection in lithium-ion batteries. The paper presents both an extended version of a previously presented ontology [1] and the framework for processing the data, which was presented in more detail before by the same authors[2].
The authors show that utilizing a multimodal approach, where the outputs of time series analysis and thermal imaging solutions are mapped to RDF Streams, which are then processed by the C-SPARQL RDF Stream Processing engine (in combination with the BAO ontology), results in more accurate anomaly detection.

The work targets a timely topic, i.e., monitoring lithium-ion batteries using a multimodal approach.

However, in its current form, the paper has some limitations:

1) The exact contribution of the paper is unclear. Is it the ontology itself, the reasoning framework or the framework for validation questions?

It seems that the ontology was (mostly) published before by the same authors ([1]). There seems to be a small addition, though, but far from sufficient to make the ontology the contribution of the paper.

The reasoning framework was published before in [2], and without proper position of the authors, it is unclear to me how this work goes beyond their previous work.

I would suggest the authors to better highlight the novelty of their contribution and differentiate it from their own previous work. In its current form, it is hard to assess the novelty of the proposed approach.

2) the limited theoretical foundations and lack of examples throughout the paper make it hard to understand.
- The definition for an 'ontology' seems rather unusual. 1) Subsumption is typically not defined as a function. 2) there are no individuals (or ABox), 3) this definition does not allow to define axioms beyond subsumption. I would suggest the authors to use either Description Logic notation or build upon the OWL2 specification.
- The definition for stream reasoning is very minimal. Also a definition for stream reasoning is not needed here. It is enough to explain what stream reasoning is.
- A definition for a stream, RDF stream, Window operators could be introduced, instead of the def. for stream reasoning.
- For the definition of a C-SPARQL query, I suggest the authors to look at the RSP-QL paper [3]
- the authors claim "C-SPARQL provides the theoretical foundation for combining static background knowledge (ontology) with dynamic observations (sensor streams)", it would be good to define what is meant with "static background knowledge" is this the abox or tbox of the ontology or both?
- Abbreviations are often introduced incorrectly. For example "State of Health SOH", should be "State of Health (SOH)"

- the lack of examples throughout section 4 make it hard to understand. I would suggest to add some examples, e.g. how does the data look like, what is a RDF stream in this case, show some examples of the used axioms in the ontology, how is the spatial reasoning done,
etc.

References for each definition is missing.

3) There seem to be a lot of vague statements in the paper that should be further clarified. I have listed some below:
- the authors seem to position their work in accordance with the core vision of Industry 5.0 (which AFAIK is shifting the focus towards human-centric, sustainable, and resilient approaches), and claim that their solution enables collaboration between humans and machines. These bold claims are presented in the introduction, but never clarified or any evidence is presented throughout the paper to support these claims. I would suggest the authors to better explain what the relation with Industry 5.0 is and if it is substantial, validate their approach according to these claims.
- the authors claim that the time series sensor readings and thermal image segments are merged into a single stream. It is unclear why this is necessary. I would imagine each has different frequencies and different analytics might apply to them.
- the authors claim to support spatial reasoning, but from the text it is unclear how this is achieved. Is this done by a separate reasoner, inside C-SPARQL? (The presented C-SPARQL queries do not seem to include any geosparql statements)
- In 4.4 the authors claim that their approach enables both temporal and spatial reasoning within a unified semantic context. It is unclear how the semantics of both temporal and spatial reasoning are unified. I would suggest the authors to either elaborate or rephrase.
- the authors claim to support causal relationships. It would be good if the authors would provide some more insights on how this is achieved. From the ontology presented in figure 3, there does not seem to be a relation between the anomalies and other parts of the ontology (except for the time module).
- In 4.4 the authors claim that this layer performs multimodal reasoning. I believe this claim to be too strong and might set wrong expectations. The C-SPARQL engine performs logical reasoning upon RDF annotated data resulting from multimodal sources.
- page 15, the authors show in listing 2 that the results of the c-sparql query results in an overheat event, however, it is unclear how that event is created, given that the query is a select query and not a construct query.

- Figure 2 seems to be missing.

4) It would have been interesting to see how the C-SPARQL engine copes with these datastreams and what is the overhead compared to other approaches. While RDF gives you interoperability, it comes with a cost.

5) It is unclear to me where the Validations Questions come from. I understand it allows to validate the ontology, but without sufficient support, it seems a bit arbitrary. For example, based on this, the authors claim that it is possible to distinguish transient from persistent anomalies. It would be good to have more explanation of where these questions came from.

6) In terms of code availability, the authors claim that the stream reasoning queries, ontology, and datasets utilized for anomaly detection, along with the experimental results, are available on their github page (https://github.com/Marwa-Zitouni/SAFE-LIB), however, I was unable to find the used queries, the ontology or the framework. There are some notebooks, the c-sparql starters pack and some ontologies, however 1) no documentation in the notebooks, 2) there is no ontology that represents the ontology in the paper (a small subset yes, but the new concepts of the image module are not there), 3) no queries are provided to run c-sparql with. In its current sthate, this is not an artifact that enables reproducability or reusability.

Minor remarks:
- many abbreviations are not or wrongly introduced
- BAO is introduced any various places, once the first time it is presented is enough.
- Figure 2 is missing.
- page 8 line 23 claims that figure 3 describes the ontology engineering approach, but it does not.
- page 10 line 1 has a formatting issue.

[1]Zitouni, M., Giustozzi, F., Samet, A. and Mesbahi, T., 2024. Toward Anomaly Representation in Lithium-Ion Batteries: An Ontology-Based Approach. Procedia Computer Science, 246, pp.1319-1328.
[2]Zitouni, M., Giustozzi, F., Samet, A. and Mesbahi, T., 2025. Anomaly Detection in Lithium-Ion Batteries via Stream Reasoning on Structured Knowledge and Time-Series Data. Procedia Computer Science, 270, pp.3688-3697.
[3] Dell'Aglio, D., Della Valle, E., Calbimonte, J.P. and Corcho, O., 2014. RSP-QL semantics: A unifying query model to explain heterogeneity of RDF stream processing systems. International Journal on Semantic Web and Information Systems (IJSWIS), 10(4), pp.17-44.

First of all, the anomaly forecasting and detection of Lithium-Ion batteries are crucial, as the consequences of failure could be devastating. Beyond just performance, addressing the opacity of the monitoring system is also a critical issue. Therefore, from my point of view, white-box analysis becomes essential, and I want to express my gratitude to the authors for tackling this topic using a semantic framework.

However, I believe that the paper must definitely distinguish between explainability and interpretability in this regard as they have been employed interchangeably. Moreover, if we are strictly talking about the opaque nature of these systems, the framework is using Neural Network (NN) models during the temperature-based feature extraction. I have strong suspicions about their opacity too, regardless of the reported MAE score of the model. Because of this black-box step at the beginning of the pipeline, I am not sure we can still call the total system fully explainable.

Furthermore, I believe that different SOTA techniques could be observed and analyzed for the image-based feature extraction, such as using Mixture Models and Multiple CRFs. Finally, I strongly believe that we need further empirical details regarding the explainability of the model results such as faithfullness, rather than just qualitative examples.

As an example, I suggest to look at the following paper for the image-based feature extraction: 10.1109/TIP.2016.2552401

This paper presents an ontology-based framework — the Battery Anomaly Ontology (BAO) — for multimodal anomaly detection in lithium-ion batteries, combining time-series sensor streams and thermal imaging within a modular semantic architecture. Stream reasoning is performed via C-SPARQL queries, enabling real-time detection of safety-critical events with interpretable, rule-traceable explanations. Ontology quality is assessed using the TOMM framework, and performance is benchmarked against data-driven baselines from Anomalib and Merlion.
The paper addresses a genuine and important problem in battery management systems. The combination of symbolic knowledge representation and real-time multimodal analytics is a valid and underexplored research direction. However, several methodological gaps must be addressed before the manuscript meets journal publication standards.

Strengths
Relevance of the problem statement. The identified gap — the inability of conventional monitoring approaches to integrate multimodal data and contextual knowledge in real time — is well-motivated and clearly articulated.
Architectural coherence. Combining heterogeneous data sources (time-series and thermal imaging) within a modular ontology constitutes a consistent and domain-relevant architectural contribution.
Rigorous ontology evaluation. The use of the TOMM framework to assess structural ontology quality reflects commendable methodological rigor. The reported results — cohesion of 0.86, high independence, and appropriate size (158 entities) — are consistent and correctly interpreted.
Transparent evaluation protocol. Algorithm 1 explicitly details the anomaly score computation, severity weights, and thresholding procedure. This level of methodological transparency is uncommon and praiseworthy.
Honest acknowledgment of limitations. Section 6.2.4 explicitly recognizes that the framework can only detect anomalies that have been anticipated and formally encoded. Integrating this self-critique into the main discussion, rather than relegating it to a closing limitations section, strengthens the paper's credibility.
Competitive performance. The multimodal BAO framework achieves an AUROC of 0.935, outperforming most tested baselines, including PatchCore (0.890) and the LSTM Detector (0.931).
Value of interpretability. The focus on interpretable explanations represents a genuine added value in safety-critical applications where black-box approaches are operationally unacceptable.

Major Weaknesses
1. Unjustified severity weights.
The weights assigned to constraint violations (0.4 for temperature, 0.3 for current, 0.3 for thermal gradient) are presented without justification — neither empirical, nor expert-elicited, nor derived from any optimization procedure. These hyperparameters directly determine the anomaly score and, consequently, all reported metrics. A sensitivity analysis or, at minimum, a domain-expert rationale is required.
2. Unmotivated decision threshold.
The binarization threshold is fixed at τ = 0.5 without discussion. Given that the anomaly score is a weighted sum with a theoretical maximum of 1.0, this choice is not neutral and directly shapes the precision-recall trade-off. A Precision-Recall curve with an analysis of the optimal threshold selection should be provided.
3. No ablation study.
The framework is evaluated solely in its full multimodal configuration (time-series + thermal imaging), while the baselines are either exclusively temporal (Merlion) or exclusively thermal (Anomalib). It is therefore impossible to determine whether the observed performance gain stems from multimodal fusion or from the symbolic reasoning approach itself. An evaluation of isolated components is necessary.
4. Dataset not quantitatively described.
The paper provides no statistics on the experimental dataset: total number of sliding windows, class distribution, proportion of anomalies per category, or sequence duration. Without this information, the reported metrics cannot be reliably interpreted, as significant class imbalance could artificially inflate precision or recall.
5. Absence of comparative justification for C-SPARQL.
C-SPARQL is adopted as the stream reasoning engine without comparative positioning against existing alternatives. This omission weakens the motivation for the architectural choice.

The work rests on a solid motivation and a well-conceived architecture. The honest treatment of limitations and the rigor of the ontology evaluation are genuine strengths. However, the identified gaps, unjustified weights, unmotivated threshold, absence of ablation study, and lack of dataset statistics, undermine the robustness of the conclusions and preclude acceptance in the current form.