Fault diagnosis in power management systems running multiple loads relies on experts manually inspecting current and voltage waveforms, which fails to scale across deployed units. We present TRACK, which pairs a time series encoder with a Small Language Model to produce structured, physics-grounded diagnostic reports, calling tools for exact, verifiable calculations. It operates on-device, keeping proprietary sensor data local and avoiding expensive API calls. We introduce a rubric-based framework that evaluates these reports by scoring structure, signal fidelity, physics reasoning, diagnosis, and actionability through per-criterion verdicts. On a 15-class multi-load simulated benchmark, TRACK achieves over 95% fault-category accuracy and generalizes to out-of-distribution faults, whereas the best encoder-probing baseline reaches only 58% accuracy.
Figures & tables
Figure 1: track architecture. The Transformer-CNN encoder processes 12 sensor channels per load; projected embeddings are interleaved with text tokens (including per-load baseline references) and fed to the frozen Qwen 2.5-3B with LoRA. The tool executor intercepts [CALL:] patterns and injects exact results via named variables. Shades of purple indicate curriculum stages, with darker shades denoting later stages approaching final SFT.
Figure 2: Reasoning trace comparison on a capacitive load. The direct generative (a) and equation tags (b) formats misdiagnose connector or contact degradation; only the tool-call format (c) computes the evidence with tools and correctly identifies capacitor ESR degradation.
Model
ID
OOD
Combined
Trans-CNN + MLP
45.53
26.10
40.16
Chronos + MLP
56.04
16.95
45.24
Trans-CNN + MLP with Baseline
56.58
28.95
48.95
Chronos + MLP with Baseline
65.71
19.29
52.88
Trans-CNN + MLP with Baseline & Load
63.58
29.33
54.12
Chronos + MLP with Baseline & Load
72.22
20.29
57.87
Table 1: MLP classifier accuracy (%) on encoder embeddings for multi-load faults. Combined is weighted by ID and OOD sample counts. Baseline appends 11 scalar references. Load adds one-hot load encoding.
#
Encoder
Calc. Style
St3 a
St4 b
Frozen c
Accuracy
Consistency
Composite
1
Trans-CNN BL
Direct generative
✗
✓
✗
92.70
95.52
77.98
2
Trans-CNN BL
Equation tags
✗
✓
✗
93.88
96.06
85.85
3
Trans-CNN BL
Tool call
✗
✓
✗
94.12
97.76
94.15
4
Chronos F
Tool call
✗
✓
✓
92.70
96.82
94.09
5
Chronos Tr.
Tool call
✗
✓
✗
94.64
95.48
94.95
6
Chronos F
Tool call
✓
✓
✓
95.42
97.97
95.01
Table 2: Training configurations and calculation style ablation (scores in %). Within each block, best scores are bold . a St3: fine-tuning on the CTS dataset with physics tool-call reasoning. b St4: fine-tuning for fault diagnosis on the Power Management Bench. c Frozen: encoder weights fixed during St4. Composite is the rubric-based score of Section 4 . Trans-CNN BL is randomly initialized, Trans-CNN-CTS is pretrained on CTS. F = frozen, Tr. = trainable at St4. ✓denotes the presence of a training setup and ✗ denotes the absence. All configurations share Stage 1 and 2.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Example from the Circuit Time Series dataset. (a) Transient waveforms from a buck converter circuit: output voltage Vout , inductor current IL1 , and load current Iload . (b) Structured reasoning traces showing physics tool calls (blue) with chained variable references (red), followed by a natural language interpretation. Tool calls invoke physics laws such as Ohm’s law, power calculation, and Kirchhoff’s current law to ground the reasoning in circuit physics.
Figure 4: Power management bench signals and extracted per-window features across six different conditions. (A) First 6 s of three synchronous channels ( VIN , IIN , Iresistance ) for five fault classes at 3200 Hz. (B) Mean ± 1 SD of three of extracted features over 116 sliding windows ( 50 samples per class); shaded bands show within-class variability.
Load Type
Condition
Train
Test
OOD
LED strip
Healthy
5,600
700
130
Connector degradation
800
100
95
Insulation damage
800
100
90
LED short circuit
800
100
185
Resistive
Baseline (reference)
2,800
350
–
Cap. ESR (minor)
406
51
50
Appendix
Table 3: Dataset split statistics. OOD samples are mapped to the closest ID label for evaluation.
Figure 5: Example from the Power Management Bench dataset illustrating a PWM frequency drift fault on a load. (a) Raw waveforms sampled at 3.2 kHz: supply voltage Vin , PWM channel current IJ4 , and total input current Itotal . (b) Extracted per-window features: voltage droop, rise time, and on-state current. (c) Structured chain-of-thought annotation showing physics tool calls (blue) with chained variable references, followed by a source diagnosis grounding the fault in circuit physics.
Figure 6: Comparison of reasoning traces formats for the same sample.
OOD Condition
Load
Wrong resistor (5 Ω )
Resistive load
Wrong resistor (20 Ω )
Resistive load
Heatsink missing
Thermal pad
Wrong pad type
Thermal pad
Thermal near-short
Thermal pad
Duty cycle stuck high
PWM load
Appendix
Table 4: OOD conditions and their affected loads.
Size
Acc.
Consistency
Composite
Latency (s)
0.5B
84.8
59.3
83.5
∼ 42 ∗
1.5B
93.3
75.8
88.8
∼ 53 ∗
3B
94.6
97.5
95.1
65
Appendix
Table 5: Effect of decoder model size on rubric evaluation (scores in %). All configurations use the same Trans-CNN-CTS encoder architecture and they are trained through St1-St4 with trainable encoder weights at St4. Latency is the inference time per sample. The 3B model is measured directly on Jetson Orin NX (16GB), and ∗ denotes expected run time for 0.5B and 1.5B scaled from their H100 batch latency ratios relative to 3B.
Fault diagnosis identifies, localizes, and explains failures in engineered and computational systems using sensor data, logs, models, and domain knowledge.
State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China · School of Information and Electrical Engineering, Zhejiang University City College, Hangzhou, China
Corporate Research Center, Midea Group, Foshan528311, People’s Republic of China · New Energy Business Division, Midea Group, Foshan528311, People’s Republic of China