Measuring modality sufficiency in automotive engine fault diagnosis: sensor ceilings and pre-trained encoder fusion

Abstract Sensor streams in automotive engine fault diagnosis carry intrinsic measurement limits that have not been systematically quantified. Using the EngineFaultDB benchmark (55 999 samples, 14 OBD variables, 4 fault classes) with controlled, author-constructed symptom descriptions as an idealized text-modality input, this paper reports three core measurements. (1) Unimodal empirical sensor ceiling and no detectable architecture contribution without informative text : five architecturally diverse classifiers converge to approximately 75% accuracy (10-seed GBDT: 0.7496 ± 0.0032, model spread 0.36 pp). The two most overlapping classes (lean mixture and ignition fault) exhibit a mean Jensen–Shannon (JS) divergence of only 0.005—near-total overlap in sensor space. A 5-seed placebo-text ablation yields 0.7495 ± 0.0009, essentially identical to the sensor-only baseline, confirming that the fusion architecture contributes no detectable gain absent informative text. (2) Pre-trained text encoders break the empirical sensor ceiling : a character-level Transformer trained from scratch reaches only 0.654, whereas frozen pre-trained encoders (MiniLM-L6: 0.880; MPNet-base: 0.918) break through the empirical sensor ceiling under the controlled symptom text; the text modality alone reaches 0.819—higher than the sensor-only baseline. The most safety-critical fault class (ignition fault) rises from per-class F 1 0.50 to 0.99, localizing the gain to pre-trained semantic knowledge rather than the fusion wiring. (3) Knowledge-graph gating and cross-modal knowledge transfer provide no additional benefit : a domain knowledge graph (35 entities, 44 relations) encoding symptom-to-sensor rules lowers aggregate accuracy relative to naïve fusion (0.871 vs 0.918), and three cross-modal knowledge-transfer methods did not transfer text information into a sensor-only model under the tested settings—indicating that the text modality must remain in the inference loop under these settings. These measurements establish a reusable JS-divergence-to-fusion measurement framework, presented as a proof of concept on controlled, noise-free, author-constructed text, for principled multimodal diagnostic system design.