acceptodds
Under review as a conference paper at ICLR 2027

KnoGML: Knowledge-Grounded Multimodal Learning for Fault Diagnosis

Abstract

Reliable fault diagnosis is essential to industrial safety and often relies on fragmented, variable-length time series. Many existing methods achieve diagnosis by learning a mapping from numerical features to fault labels, but face two key limitations: 1) they typically rely on lossy operations such as truncation or downsampling to accommodate fixed-length model inputs, limiting the use of complete recorded segments; 2) the domain-specific physical meaning of signals and their relationships is not explicitly utilized. To this end, we propose KnoGML, a knowledge-grounded multimodal learning framework that combines multivariate plot images with structured diagnostic priors for fault diagnosis. Specifically, domain-specific fault knowledge is organized into structured prompts describing variable meanings, cross-variable relationships, and fault associations within a three-stage diagnostic protocol comprising phase identification, symptom verification, and final decision. Meanwhile, multivariate time-series data are converted into fixed-resolution grid images, providing a unified visual representation for time-series segments of varying lengths. Building on this, a multimodal large language model jointly processes the structured prompts and visual representation, transforming fault diagnosis into a matching problem between fault descriptions and signal patterns. Finally, a classification head produces a deterministic fault label, providing a fixed output interface for industrial integration. In electric vehicle battery fault diagnosis, KnoGML improves Macro-F1 over the best traditional baseline by 7.44 and 1.97 percentage points across two battery chemistries, demonstrating the effectiveness of incorporating structured fault knowledge into multimodal learning. Code is available at https://anonymous.4open.science/r/KnoGML-77AB/.

Primary area:
learning on time series and dynamical systems
Keywords:
Fault Diagnosis, Knowledge-Grounded, Multimodal Learning, Time Series, Diagnostic Evidence

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rZ85OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KnoGML for battery fault diagnosis from fragmented, variable-length time series. It combines synchronized multivariate plots with phase-dependent diagnostic priors and adapts a pretrained multimodal model using LoRA and a classification head to provide fixed-label outputs without free-form generation. On the semi-synthetic CH-BatteryGen benchmark, it achieves Macro-F1 scores of 90.24 and 87.61 percent on LFP and NCM, respectively. Prompt and component ablations assess the complete design’s contributions.

Critical strengths
  1. KnoGML renders variable-length multivariate segments as synchronized plots, providing a common input format that covers the recorded interval and supports temporal comparisons across battery signals.

  2. Phase-dependent diagnostic priors connect cross-variable symptoms with fault descriptions, enabling a pretrained multimodal model to incorporate battery knowledge into classification through LoRA adaptation.

Critical weaknesses
  1. The source of the image-free performance is unclear. Table 4 reports Macro-F1 scores of 81.61%/79.84% on LFP/NCM without images, while Section 3.2 includes vehicle IDs and operating status in the prompt. Please clarify which inputs remain and report results without vehicle IDs to explain what drives these scores.

  2. Please report the classification results of the original Qwen2.5-VL-7B-Instruct model without fine-tuning, using the same images, contextual information, and complete prompt in Table 7.

3.The paper has clear practical value, but its methodological novelty mainly lies in combining existing components. Time-series visualization, diagnostic prompts, LoRA, and classification heads all build on established techniques, and the paper does not clearly explain what new methodological contribution their combination provides.

Questions for authors
  1. Which inputs remain in the image-free variant in Table 4, how is the operating-status description generated, and how does performance change when vehicle IDs are removed?

  2. Could you report the classification results of the original Qwen2.5-VL-7B-Instruct without fine-tuning, using the same images, contextual information, and complete prompt in Table 7?

  3. Finally, what is the main methodological contribution beyond combining time-series visualization, diagnostic prompts, LoRA, and a classification head, and which experiments demonstrate its benefit?

Reviewer_W9mROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KnoGML, a knowledge-grounded multimodal framework for industrial fault diagnosis. The method converts variable-length multivariate time series into plot images and combines them with structured diagnostic knowledge through a multimodal LLM. A classification head is used to produce deterministic fault predictions. Experiments on two battery chemistries demonstrate improvements over numerical and multimodal baselines.

Critical strengths

Motivation: I appreciate the motivation to incorporate physical knowledge into fault diagnosis rather than relying entirely on data-driven feature learning. This is particularly relevant when fault labels are scarce.

Evaluation: The improvements on both battery chemistries are promising. The prompt ablations also provide useful evidence that the quality of diagnostic knowledge matters.

Critical weaknesses

Limited methodological novelty and unclear knowledge-guided reasoning. My main concern is that KnoGML primarily combines several existing techniques, including time-series visualization, structured prompting, LoRA, and a classification head. While this combination may be practically useful, I am not fully convinced that it constitutes a new knowledge-guided learning mechanism. In particular, the proposed phase identification, symptom verification, and final decision stages are not independently supervised or verified. It remains unclear whether the model actually follows this diagnostic procedure or simply learns correlations between fault descriptions and labels. I would find the contribution more convincing if the authors could isolate the effect of the three-stage organization from simply providing the same diagnostic knowledge in an unstructured prompt.

The variable-length motivation is not sufficiently supported. The paper argues that numerical models require truncation or downsampling, while converting sequences into images preserves the recorded observations. However, I find this comparison somewhat misleading because modern numerical time-series models can also process variable-length inputs through masking, pooling, or temporal patching. Moreover, mapping sequences containing tens of thousands of observations into a fixed-resolution image inevitably compresses temporal information. The resolution sensitivity experiment does not establish whether short-duration fault patterns remain distinguishable. A comparison with numerical encoders that have access to the complete recorded intervals would be important for validating this motivation.

The contribution of visual observations versus contextual knowledge is unclear. Table 4 shows that removing visual input still yields 81.61% and 79.84% Macro-F1 on LFP and NCM, respectively. This is surprisingly strong for a fault-diagnosis task that supposedly depends on recognizing signal patterns. Since the prompt also includes battery type, vehicle identifier, and operating status, I wonder how much predictive information is already available from these contextual fields. Although the authors use vehicle-disjoint splits, this does not clarify which contextual variables drive the predictions. More detailed ablations would help establish whether the model is learning fault-specific visual evidence rather than relying heavily on contextual information.

Limited generalization evidence and baseline competitiveness. Both evaluated battery chemistries come from the same semi-synthetic benchmark with controlled fault injection. The diagnostic rules may therefore align closely with the fault-generation mechanisms, making it difficult to assess whether the method generalizes to independent real-world failure patterns. In addition, the improvement over KNN on NCM is relatively modest (87.61% versus 85.64% Macro-F1), despite the substantially greater model complexity. I would like to see stronger evidence that the improvements are attributable to knowledge-grounded learning rather than model capacity, task-specific adaptation, or characteristics of the synthetic faults.

Questions for authors

see weakness

Reviewer_kVp3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KnoGML, a knowledge-grounded multimodal learning framework for fault diagnosis from fragmented, variable-length time series. It converts multivariate signals into fixed-resolution grid images and organizes diagnostic knowledge into structured prompts covering phase identification, symptom verification, and final decision. A pretrained multimodal model processes the images and prompts, with LoRA adaptation and a classification head producing fault labels. Experiments on the semi-synthetic CH-BatteryGen benchmark show Macro-F1 scores of 90.24% for LFP and 87.61% for NCM batteries. Ablations show the value of explicit fault associations, visual input, and adaptation. Fixed image resolution and validation on field data remain limitations for deployment.

Critical strengths
  1. The structured prompts provide clear guidance on variable meanings, cross-variable relationships, and fault associations. The ablations show that fault associations help more than descriptions of signal changes alone.
  2. KnoGML performs well on both battery chemistries, achieving the highest Macro-F1 and Weighted-F1 among the evaluated baselines. Keeping all segments from the same vehicle in one subset also helps reduce data leakage.
Critical weaknesses
  1. The evaluation uses only the semi-synthetic CH-BatteryGen benchmark. Testing on two battery chemistries is useful, but it does not show whether KnoGML can handle real faults and changing operating conditions in field data.
  2. The no-image variant still achieves Macro-F1 scores of 81.61% and 79.84%, although the fault descriptions are shared across samples. This raises questions about how much information comes from contextual fields, such as vehicle identifier and operating status. Removing these fields would help clarify what the model learns from signal patterns.
Questions for authors
  1. Have you tested KnoGML on real field data or under changing operating conditions? How well do the diagnostic priors match real faults beyond the controlled fault injection in CH-BatteryGen?
  2. Why does the no-image variant still achieve relatively high Macro-F1 when fault descriptions are shared across samples? Could you report results without vehicle identifiers and operating status to clarify how much these contextual fields contribute?
  3. Given the complexity of the framework, have you considered its real-time inference efficiency and practical challenges in industrial deployment?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.