cs.LGSep 30, 2026
SaveSCORE-LM: State-Space Radar Representations with Language Models for Fault Diagnosis
Organizations: Georgia Institute of Technology, Atlanta, GA, USA
Abstract
Radar hardware faults threaten automated perception, motivating accurate, compact diagnosis and understandable maintenance guidance. We introduce SCORE-LM, which couples a small scatterer-conditioned operator-response encoder (SCORE) to an adapted local language model. SCORE combines self-referenced complex trajectories, physical descriptors, and a selective state-space branch, with source-only self-supervision and directional fault inference. On eight capture-excluded Rad-R fault recordings, it achieves state-of-the-art performance within the evaluated nine-model comparison: 88.39% mean capture recall and 88.20% four-fault macro-F1 at ten frames. Its 39,520 radar inference coefficients are 119.7 times fewer than RadrNet-DS-CI's, while recall is 15.56 percentage points higher than this strongest competitor. In a separate low-label protocol, SCORE reaches 71.58% recall with one labeled source window per class. A nonlinear projector converts four frozen fault similarities into five soft tokens, linking compact diagnosis to class-conditioned maintenance guidance. On 75 development questions covering 24 radar windows, language adaptation raises correct-fault answers from 45 to 62 (60.0% to 82.7%) relative to removing the co-trained adapters, while retaining the same projector. SCORE-LM thus combines a compact radar specialist with a language interface for communicating fault-specific inspection guidance.
Figures & tables
Fig. 1: Rad-R capture setup, arranged from the authors’ dataset assets [ 1 ] . From left: front and side vehicle views, followed by the two sensor-plate views. The hardware includes the cascade radar, camera, inertial and environmental sensors, positioning, storage, and controlled-vibration hardware.
Fig. 2: Rad-R acquisition dashboard: yaw2, radar frame 2122, snapshot at 217 s [ 1 ] . Radar views, camera frames, and auxiliary-sensor traces illustrate the recorded context.
Fig. 3: Frozen SCORE signal pipeline, with the selective-state-space block and source self-supervision. Each peak contributes 512 temporal and 54 descriptor coordinates. Temporal and descriptor branches are fused before mean pooling over peaks. Mean-plus-last pooling over radar loops is distinct from frame or peak averaging. The three legacy descriptors are uncalibrated heuristics. The diagram is schematic; reconstruction is training-only. The language connection in Fig. 4 is added after the frozen diagnostic readout.
Fig. 4: Visual overview of score-to-language alignment. (a) Frozen SCORE and its directional readout produce four raw similarities; projected tokens and a question condition generation. (b) Three-layer nonlinear projector, with widths shown for Qwen. (c) Low-rank updates inside frozen decoder linears. (d) Source-only answer-token training updates the projector and LoRA, not SCORE or base-language weights. Tensor shapes and traces are schematic; no textual fault label is supplied as input.
Fig. 5: Ten-frame confusion matrices using a common row-percentage scale and five seeds. Each seed contributes 40 vibration, 40 misalignment, 38 blockage, and 40 degradation decisions. There is no true Healthy row; SCORE’s empty Healthy prediction column follows its four-fault directional readout.
Fig. 6: Nine-method low-label transfer with common source-window selections. Error bars are population SD across five resample means. SCORE additionally uses unlabeled source pretraining. One label is one ten-frame window. The 20-window budget uses seven folds because one fold has only 18 source blockage windows; other budgets use eight.
| Method | 1 frame | 10 frames | 50 frames | 100 frames | 10-frame BA | 10-frame F1 |
|---|---|---|---|---|---|---|
| RD-CNN ERM | 69.68 | 71.94 | 75.62 | 80.00 | 72.12 | 71.36 |
| CORAL (adapted) | 68.94 | 70.18 | 72.50 | 73.75 | 70.29 | 68.85 |
| MixStyle | 61.99 | 63.29 | 65.00 | 73.75 | 63.34 | 63.45 |
| DARM (adapted) | 61.31 | 64.62 | 68.75 | 73.75 | 64.68 | 63.22 |
| DDDG (adapted) | 67.51 | 69.40 | 70.83 | 71.25 | 69.39 | 69.27 |
| BDC (adapted) | 59.00 | 63.69 | 67.71 | 67.50 | 63.72 | 64.48 |
TABLE I: Primary capture-mean recall by frame budget and pooled ten-frame balanced accuracy (BA) and macro-F1, in percent. All five seeds retained. Adapted methods use radar-specific implementations.
Fig. 7: Before and after SCORE encoding for 178 matched ten-frame windows: 160 source windows and 18 held-out Blockage1 windows. Healthy samples are source-only. The inputs have 566 coordinates and the embeddings 128. Both t-SNE fits use cosine distance, perplexity 30, 1,000 iterations, and identical random initialization; labels only set plot colors. Each panel has its own two-dimensional coordinates.
| Input | 1-frame recall | 10-frame recall | 10-frame F1 |
|---|---|---|---|
| Full | 84.56 (1.09) | 88.39 (0.70) | 88.20 (0.67) |
| Physical only | 84.62 (1.60) | 87.47 (1.47) | 87.29 (1.37) |
| Temporal only | 27.45 (1.17) | 44.58 (4.85) | 39.61 (5.32) |
| No legacy geometry | 74.34 (3.24) | 86.11 (2.25) | 85.86 (2.26) |
TABLE II: Component ablations: mean recall and macro-F1 (%). Full and Physical only use the five-seed directional protocol (population SD). The last two rows retain the earlier three-seed input-ablation protocol (sample SD).
| Runtime | Graph | GPU median / p99 (ms) | DSP median (ms) | DSP + GPU median / p99 (ms) |
|---|---|---|---|---|
| Windows | DDDG | 1.04 / 1.93 | 132.83 | 138.13 / 156.87 |
| WSL | RadrNet-DS-CI | 8.51 / 14.68 | 89.18 | 147.82 / 197.58 |
| WSL | SCORE directional graph | 9.97 / 15.03 | 79.69 | 92.86 / 122.19 |
TABLE III: Processing measurements from separate runs: batch one, one CPU thread, 20 warm-ups and 200 synchronized repeats. Composed timing is measured directly.
Fig. 8: Radar model size versus ten-frame fault accuracy, measured as mean capture recall over eight held-out recordings. Error bars: population SD across five seeds. Size counts active scalar inference weights and fitted readout coefficients; scaler statistics, running buffers, training-only heads, and all language modules are excluded. SCORE includes its 640 center/direction values; RadrNet excludes its unused severity head. Adapted methods and readout/supervision differences follow Sec. VI. The highlighted gain is relative to RadrNet-DS-CI, the highest-recall competitor.
| Score transform | Projector | True fault | Decoder agreement |
|---|---|---|---|
| Softmax | Linear | 20/25 | 22/25 |
| Sigmoid | Linear | 20/25 | 22/25 |
| Raw | Linear | 21/25 | 20/25 |
| Sigmoid | Full MLP | 6/25 | 5/25 |
| Raw | Full MLP | 23/25 | 24/25 |
TABLE IV: Historical projector development study, fixed 25 short diagnosis questions. The full MLP and linear projectors differ in capacity and initialization.
| Measure | With rank-8 LoRA | LoRA removed |
|---|---|---|
| Correct fault, all questions | 62/75 (82.7%) | 45/75 (60.0%) |
| Decoder agreement, all | 67/75 (89.3%) | 47/75 (62.7%) |
| Correct fault, canonical windows | 20/24 | 17/24 |
| Decoder agreement, canonical | 22/24 | 18/24 |
| Output cap reached | 0/75 | 20/75 |
| Unsupported observations † | 0/75 |
TABLE V: Current maintenance evaluation. Same scores, questions, jointly trained projector and decoding; only language LoRA is removed in the control. Counts cover 75 questions on 24 radar windows.
Fig. 9: Matched maintenance dialogue, Q005 (Blockage1, W11). Both models receive the same radar scores, question and jointly trained projector; the control removes language LoRA. The adapted answer is complete; the control is an exact excerpt. The signal icon is schematic. Highlighted boxes are post-hoc review annotations, not model outputs. Both fault labels are correct in this example, but the control asserts echo absence that the four-score input does not establish. Aggregate unsupported-observation counts exclude fault-label errors: the adapted model still makes 13/75 incorrect fault statements.
| Backbone | Correct fault / 75 | Agreement / 75 | Unsupported / 75 | Capped / 75 | Time / answer |
| Qwen3.5-4B | 62 | 67 | 0 † | 0 | 4.75 |
| Phi-4-mini-instruct | 59 | 66 | 0 † | 0 | 2.98 |
| Ministral-3-3B-Instruct-2512 | 54 | 57 | 0 † | 0 | 2.67 |
| Granite-4.2-3B | 65 | 70 | 0 † | 0 | 3.88 |
| † Answer-level assistant review of unsupported observations beyond the fault label. | |||||
TABLE VI: Matched maintenance QA with rank-8 language adapters and trained projectors. Greedy decoding, 384-token cap; 75 questions on 24 windows. Time is mean generation latency in seconds per answer on an RTX A4000, excluding model loading.