We study industrial fault diagnosis as a classification task that assigns sensor time-series windows to predefined health states or fault types. Scarce labels and changing operating conditions make repeated training of specialized classifiers costly. Yet structures such as periodicity, trends, and transients can remain diagnostically informative and be examined through reusable analytical tools. LLM agents can select and interpret these tools, but fixed rules or training-derived guidance may not resolve each signal's ambiguity, while unconstrained interpretation may overstate weak or conflicting results. We introduce Agentic4PDM, a framework that separates adaptive investigation from control over evidence use. LLM agents compare each signal with labeled references from every candidate class, identify unresolved distinctions, and propose follow-up measurements. A deterministic software controller, called the host, enforces contract-checked gated adjudication (CCGA): it checks requests, requires commitment to a complete tool batch before results are observed, and assigns evidence roles based on tool checks against labeled references. These roles distinguish corroborative support from advisory information. The Judge Agent proposes a diagnosis, while the controller checks protocol compliance and records the investigation. Agentic4PDM supports training-free diagnosis using pretrained representations and labeled references without task-specific diagnostic models, and a standard setting that additionally investigates trained-model predictions. Experiments on selected bearing and motor diagnosis challenges and an unfiltered industrial test set supplied by our customer show gains in aggregate macro F1 over the evaluated diagnostic and agentic baselines. Ablations further support the value of the governed workflow.
Figures & tables
Figure 1: Agentic4PDM workflow. The Host constructs and freezes class-balanced comparative evidence before the Judge begins sample-specific investigation. The Judge proposes follow-up measurements or a diagnosis, while the Host governs which measurements may run and how returned evidence may influence the decision. The Host then publishes the reported class, marked as admitted or not admitted, together with the record of the checks behind it.
Figure 2: Performance comparison between Agentic4PDM and baselines from three model families agentic methods, time-series foundation models, and deep-learning models across three evaluation settings: (a) GRS, (b) BFH, and (c) the customer test set (Bel). M1.1R, M2S, C2, and C2S denote Moirai 1.1 R Large, Moirai 2 Small, Chronos 2, and Chronos 2 Small, respectively; XGB denotes XGBoost.
Figure 4: Observed plans across agentic methods on PU and MCC5. Bar segments show the frequency of each plan, the right number indicates the total number of distinct plans.
Enabled mechanisms
Performance (%)
Variant
CB retrieval
Analyst reports
Request gate
Packet & influence gate
Finali- zation
ACC ↑
F1-M ↑
A1 Full workflow
✓
✓
✓
✓
✓
47.24
38.54
A2 Global top- K
✗
✓
✓
✓
✓
37.20
29.89
A3 No Analyst reports
✓
✗
✓
✓
✓
46.35
37.72
A4 Ungoverned
✓
✓
✗
✗
✓
36.31
30.53
A5 No request gate
✓
✓
✗
✓
✓
39.96
33.35
Table 1: Ablation of the main components in Agentic4PDM on the PU data under GRS sampling.
Figure 5: Mean recorded time and total tokens per query for ZARA, FETA, and standard Agentic4PDM . Here, k denotes 1,000 tokens. Marker shapes identify methods; dashed lines guide the eye
Appendix figures & tables27 assets
Supplementary material from the paper’s appendix.
Appendix
Table 2: Performance comparison of Agentic4PDM against deep-learning, time-series foundation-model, and agentic baselines across three evaluation settings: BFH and GRS on the public benchmarks (UO, PU, and MCC5), and the full customer test set (Bel). ACC and F1-M denote accuracy and macro-F1, respectively; AVG is the unweighted mean across public datasets. Bold indicates the best results in each column.
Group
Method
Baseline-Failure Hard (BFH)
Geometry-Stratified Response-Stability (GRS)
AVG
UO
PU
MCC5
UO
PU
MCC5
BFH
GRS
ACC
F1-M
ACC
F1-M
ACC
F1-M
ACC
F1-M
ACC
F1-M
ACC
F1-M
ACC
F1-M
ACC
F1-M
ZARA [ 39 ]
45.00
39.30
39.00
40.31
35.00
24.90
63.43
65.26
41.94
43.23
47.66
44.22
39.67
34.84
51.01
50.90
Agent
FETA [ 55 ]
34.00
32.32
36.00
33.84
17.00
17.65
53.73
53.46
35.21
30.79
38.02
36.90
29.00
27.94
42.32
40.38
Ours
Agentic4PDM (TF)
63.00
62.12
44.00
42.30
42.00
40.73
81.52
81.40
47.68
40.78
50.69
50.14
49.67
48.38
59.96
57.44
Appendix
Table 3: Performance comparison of training-free (TF) Agentic4PDM against ZARA and FETA across two public benchmarks, BFH and GRS, on UO, PU, and MCC5. ACC and F1-M denote accuracy and macro-F1, respectively; AVG is the unweighted mean across datasets. TF uses no fitted-model predictions. Bold indicates the best results in each column, including ties.
Figure 6: Variation across three executions of standard Agentic4PDM , ZARA, and FETA on BFH and GRS. For each method, dataset, and metric, scores are centered on their mean across the three executions. The dots therefore show deviations from that mean in percentage points, with runs 1–3 ordered from left to right within each group. Shaded bands span ±1 sample standard deviation ( n=3 ); the corresponding standard deviation appears below each group. AVG denotes the unweighted mean of the three dataset scores in each execution. BFH and GRS use different vertical scales.
Method
MCC5
PU
UO
Flat source plurality
48.26
28.45
85.00
Agentic4PDM
54.09
37.90
87.38
Appendix
Table 4: Macro-F1 (F1-M, %) of Flat source plurality and Agentic4PDM on the same covered samples in each dataset. Bold marks the higher score.
Table 10
Tool (selected on) / method
ToolCard / scoring
MCC5
PU
UO
Mean
Windowed profile (MCC5)
quant_windowed_profile_windows_16_amplitude_l1
36.15
35.22
30.17
33.85
Crossing-run statistics (PU)
quant_crossing_run_statistics_q50_js
22.89
37.94
16.73
25.85
Distribution KS (UO)
quant_distribution_ks_raw
31.78
32.78
60.02
41.53
Agentic4PDM
Saved class
52.93
38.54
86.12
59.20
Appendix
Table 7: Macro-F1 (%) of the dataset-winning tools among the 25 selectable defaults and the full Agentic4PDM in the updated GRS audit. Agentic4PDM uses saved-class scoring. Bold marks the highest displayed score in each column.
Method
Time (s)
Input (k)
Output (k)
Total (k)
LLM calls
FETA
1,078.59
68.05
4.62
72.67
8.0
ZARA
98.20
5.11
2.72
7.83
4.0
Agentic4PDM (standard)
1,645.79
89.88
12.48
102.36
17.4
Appendix
Table 8: Mean recorded resource use per query sample. Token columns are in thousands (k); time is in seconds. Recorded timer scopes differ across methods, as described below.
Figure 10: Diagnostic macro-F1 versus estimated Qwen3.6-27B API cost per query at Alibaba Cloud’s Singapore/International list prices. Costs follow Eq. 6 , using separate input and output token rates. Each method’s mean cost estimate is reused across GRS, BFH, and Bel. These values estimate token-based API charges rather than observed API bills or total deployment costs.
Figure 11: Accuracy of FETA, ZARA, and Agentic4PDM across the displayed PU operating conditions and MCC5 motor speeds.
Dataset
Classes
Train
Fault generation / setting
Operating conditions
Rate (Hz)
PU [ 36 ]
3
150
Real bearing damage from accelerated lifetime tests
Condition 0: 900 rpm, 0.7 Nm, 1000 N. Conditions 1–3: 1500 rpm with torque/radial-load pairs 0.1 Nm/1000 N, 0.7 Nm/400 N, and 0.7 Nm/1000 N.
64,000
UO [ 25 ]
3
150
Inner-/outer-race defect specimens; fabrication method not specified in the source description
Time-varying speed: A, increasing; B, decreasing; C, increasing then decreasing; D, decreasing then increasing.
200,000
MCC5 [ 8 ]
4
200
Laboratory motor measurements with rotor, electrical, and bearing fault configurations
Frozen scored pool: 20 Nm at 1000, 2000, and 3000 rpm (Conditions 0–2).
12,800
Bel
2
20
Bearing vibration supplied by our customer; binary normal/abnormal labels, with damage type unspecified
-
–
Appendix
Table 9: Acquisition characteristics and retained-task statistics for PU, UO, MCC5, and Bel. Bel is evaluated separately from BFH and GRS; a dash indicates an unreported sampling rate.
(a) Bearing diagnosis
Dataset
Challenge
Total N
Healthy
Inner-race
Outer-race
PU
BFH
100
34
33
33
GRS
906
205
487
214
UO
BFH
100
34
33
33
GRS
763
197
221
345
Appendix
Table 10: Test-set sizes and per-class support for the frozen BFH/GRS manifests and the complete Bel test set. Each class count is a number of samples; class counts sum to the total N in each row.
Model / backbone
Prediction / reasoning approach
(a) Deep learning Evaluated baselines
CNN-TCC [ 18 ] QCNN [ 41 ] WDCNN [ 66 ]
Neural classifiers
(b) Foundation models Evaluated baselines
Moirai 1.1 R Large [ 59 ] Moirai 2 Small [ 42 ] Chronos-2 [ 3 ] Chronos-2 Small [ 3 ]
Fixed representations with KNN, SVM, or XGBoost (12 combinations), following MOMENT’s representation-plus-SVM evaluation [ 22 ] , StarEmbed’s fixed-embedding evaluation [ 37 ] , and the broader embedding comparison [ 27 ] .
(c) Agentic methods Evaluated baselines
FETA [ 55 ]
Training-set channel ranking, DTW exemplar retrieval, and channel-level LLM reasoning with confidence-based aggregation.
Appendix
Table 11: Baseline inventory grouped by model family. Each foundation-model backbone is paired with KNN, SVM, and XGBoost.
Protocol item
Setting
Labeled reference pool
50 samples per class: 150 each for PU and UO (3 classes), and 200 for MCC5 (4 classes)
Query sample
2048 observations, non-overlapping, using all available dataset channels
Appendix
Table 12: Shared labeled-reference and query-sample settings for the evaluated methods; BFH/GRS manifest sizes and per-class support are reported in Table 10 .
Component
Setting
Value
(a) Language model
Model and access
Qwen3.6-27B vision-language model, served locally through an OpenAI-compatible endpoint; the same model serves every agent role
Decoding
Temperature 0, top- p 1, top- k−1 , presence penalty 1.5, seed 0
Extended thinking
Enabled for the analysis and synthesis roles, 512-token budget
Output limit
38,048 tokens per request
Images
At most 10 per request, longest side 1400 px
Appendix
Table 13: Main Moirai 1.1 R Large Agentic4PDM configuration. One setting applies to standard and TF on both challenges and all three datasets; TF omits fitted-model outputs. Appendix E specifies the Chronos-2 Small substitution.
Optimizer
Schedule
Architecture
Algorithm
β1
β2
Learning rate
Epochs
Batch size
Bearing (3 channels)
Motor (8 channels)
Common
Adam
0.9
0.95
10−3
100
128
First convolution: kernel 16, stride 2
First convolution: kernel 8, stride 1
64 final channels; dropout 0.35
Appendix
Table 14: Training settings shared by the deep-learning baselines (CNN-TCC, QCNN, and WDCNN).
Group
Setting
Value
(a) Foundation-model representations
Backbones
Moirai 1.1 R Large, Moirai 2 Small, Chronos-2, Chronos-2 Small, all frozen
Evaluated on the common BFH/GRS task, split, manifests, reference data, and scoring procedure.
Appendix
Table 16: Implementation settings for FETA and ZARA in the BFH and GRS evaluations. The table summarizes their method-specific workflows under common evaluation conditions.
Figure 12: GRS construction: (1) allocate geometry clusters within class–condition strata; (2) group frozen-model responses into stability tiers; (3) retain one observed representative per nonempty tier. Bars show response variability Is ; counts show cluster allocations gs (illustrative total G=250 ). Stability is measured within each model across samples, not between models.
Symbol
Meaning
Inputs and reference retrieval
x⋆
Query multivariate time-series sample supplied for diagnosis.
xprev
Immediate predecessor of the query sample, available only through one-step temporal context.
T,C
Numbers of observations and sensor channels in one sample, respectively.
Y,Y,y
Diagnosis-label set, its number of classes, and a generic candidate class.
Dref,N
Labeled training set (the reference pool) and its number of samples.
Appendix
Table 17: Summary of main notation for the Agentic4PDM protocol and inference path.
#
Property
What it describes
1
amplitude_distribution_match
How the signal’s amplitude values are spread, i.e., the shape of their histogram.
2
severity_or_offset_distribution_shift
Whether the amplitude values are shifted, widened, or offset as a whole.
3
global_summary_match
Whole-sample summary statistics such as level, spread, and peakiness.
4
global_severity_or_impulsiveness
How strong and how spiky the signal is over the whole sample.
5
local_ordering_dynamics
The order in which neighbouring values rise and fall.
6
scale_invariant_local_shape
Local rise-and-fall patterns, regardless of amplitude and time scale.
Appendix
Table 18: Registered signal properties. Each property names an aspect of the signal that may remain unresolved; none refers to a class.
ToolCard
vibration_1 (high)
phase_current_1 (high)
phase_current_2 (medium)
Winning class
Step 1
Fractal scaling
0
0
2
Class 0
Usable
Local extrema
2
2
0
Class 2
Usable
Autocorrelation
0
2
1
None
Excluded
Appendix
Table 19: Step 1 on the round-1 packets of sample 107665: per-channel winners and the resulting winning class.
Sample
ToolCard
Class-0 control
Class-1 control
Class-2 control
gref(τ)
Role
106619
Autocorrelation
0
1
2
1
Corroborative
107665
Fractal scaling
0
none
none
0
Advisory
Appendix
Table 20: Reference preflight records: the class named by the ToolCard for each control, and the resulting role.
Action
Purpose and constrained request interface
compare_normal compare_abnormal
Re-check the normal-region overlay or the abnormal pairwise evidence. Request task_objective and judge_feedback ; the abnormal action may additionally bound target_classes .
view_image
Load one previously supplied figure for next-turn close inspection. Request image_path .
run_sandbox
Optional same-transform, full-sample comparative measurement. Request task_objective , sandbox_request , and metric_spec ; configurations that isolate the registered quantitative library disable this action.
select_quantitative_tools
Declare the unresolved properties Φt ; the Host returns the offer St from the 25 default ToolCards (Declare and Offer, Section 3.5.1 ).
run_quantitative_tools
Commit a batch Tt from the offer; the Host checks the request conditions and then runs it (Commit and Execute, Section 3.5.1 ).
finalize
Propose one terminal class decision. It must be standalone and include the Host-required fault-screen contract.
Appendix
Table 21: All seven general tool actions available to the Judge .
Family
Registered purpose
savgol
Light polynomial denoising that preserves local waveform shape.
median
Suppress isolated impulses without averaging their amplitude into neighbors.
ewm
Track a smooth causal-style local level with bounded lag.
median_mean_ewm
Extract a slow robust trend from mixed impulsive and broadband noise.
lowpass
Remove high-frequency content while avoiding phase shift.
The channel is a primary discriminator: it shows transient bursts or impulses, a clear envelope change, a region marked as deviant, or class-separating structure whose removal would change the class ranking.
medium
The channel is useful but secondary: it gives operating context, such as load or phase-current behavior, or a weak deviation that supports but cannot settle the class. This is the default tier for a kept channel.
low
The channel carries weak but real class information and serves only as tie-break support.
removed
The channel shows no visible class information.
Appendix
Table 23: Channel priority tiers assigned by the CSAgent .
ID
Selectable ToolCard
Fixed configuration
Measurement purpose
T01
quant_autocorrelation_lags_128_pearson
max_lag=128; similarity=pearson
Compare periodic recurrence across lags without requiring cycle-phase alignment.
T02
quant_autoregressive_coefficients_order_8_l1
order=8; similarity=l1
Compare short-memory linear dynamics using regularized autoregressive coefficients.
T03
quant_cepstral_shape_quefrency_128_pearson
max_quefrency=128; similarity=pearson
Compare repeated spectral spacing over normalized quefrency bins.
T04
quant_crossing_run_statistics_q50_js
similarity=js; threshold_quantile=0.5
Compare threshold crossings, run lengths, and state persistence.
Compare local residual shape after removing a declared slow trend.
T06
quant_distribution_ks_raw
kind=ks; standardize=false
Compare complete-sample amplitude distributions without requiring point alignment.
Appendix
Table 24: The 25 default quantitative ToolCards available to the agent. Each entry gives its exact identifier, fixed configuration, and measurement purpose. Eligibility remains subject to the input, requested properties, question scope, and budget.
Appendix
Protocol event shown
N
Agent selection of transform families and channel priorities, followed by Host parameter tuning, deterministic segmentation, and channel guards.
N
Frozen NAAgent and ABAgent comparisons that reject the healthy class but leave the two fault classes live.
In so-called "Industry 4.0," integrated sensors, analytics tools, and other digital technologies supplement industrial workflows, potentiating enhanced visibility and automation. Prognostics and health management (PHM) and natural language processing (NLP) practitioners alike have explored the application of large language models (LLMs) to industrial asset management, but many prior works' focus on exam-like benchmarks has limited their assessment of LLMs' ability to automate practical operations. Herein, we introduce CoffeeBench, a fault diagnostics task for LLM agents, using a simulated appliance. The LLMs we evaluate struggle to consistently identify faults from the sensor information alone, but the best-performing ones can iteratively refine the fault-option space to eventually obtain a successful diagnosis. Our subsequent analyses, therefore, focus on: differentiating models by diagnostic efficiency (RQ1); examining the lagging performance of the three worst models (RQ2); and identifying viable heuristics for selective prediction (RQ3), including that initial diagnostic uncertainty and LLM-judged soundness of reasoning are high-signal predictors of diagnostic success. In addition to our benchmark results (Section 3.1), we provide practical insights for improving the performance of the open-weight models on CoffeeBench, such as by explicitly tracking episode state on their behalf to reduce timeouts (Section 3.2) or applying a deliberation loop to refine diagnostic hypotheses (Section 3.3).
Diagnosing the root cause of anomalies is essential for safe industrial operation. Despite extensive sensor instrumentation, formulating hypotheses and gathering evidence remains a manual process, creating a major operational bottleneck. While existing data-driven approaches aim to automate this, two critical limitations restrict their deployment: their operate as black boxes unable to justify their diagnosis, and they require scarce labeled examples of faulty operation. To address this gap, we introduce AgentRCA, a zero-shot agentic framework for evidence-grounded root cause analysis. Rather than learning fault-specific mappings, AgentRCA performs inference-time reasoning by combining a data-driven digital twin (modeling normal system dynamics) with a tool-augmented large language model. The agent iteratively gathers statistical evidence, evaluates competing hypotheses, and identifies the physical fault that best explains the observed behavior. Evaluated on a real-world multiphase-flow facility and a large-scale chemical plant, AgentRCA achieves diagnostic performance competitive with fully supervised baselines without relying on fault-specific training. Crucially, it produces transparent reasoning traces that explicitly link observed symptoms to their underlying physical causes. These results establish autonomous hypothesis-driven reasoning as a practical foundation for scalable industrial root cause analysis.
Large language models are increasingly used as integration layers above specialized tools, but a stronger component does not necessarily produce a stronger combined system. Across five diagnostic datasets (bearing vibration, process monitoring, semiconductor equipment), we study whether an LLM can reliably use external diagnostic information; paired repeat calls separate advice effects from output instability. In all five, conflicting external information overturned initially correct LLM judgments. Among the four datasets with direct integration comparisons, none showed a consistent advantage for implicit LLM integration over the stronger standalone source. On a Tennessee Eastman confirmation set whose protocol was fixed before evaluation, unaided accuracy was 64.67%, implicit LLM-specialist integration 77.43%, and the specialist alone 83.33%. Specialist information improved the LLM by 12.8 points (95% interval 9.7 to 15.9), yet the integrated output stayed 5.9 points below the specialist (95% interval -12.0 to -0.7). A two-source selector oracle reached 92.76%, indicating complementarity that the integrated output did not fully realize. The integrated output missed 140 of 295 specialist corrections (47.5%) but lost 15 of 99 initially correct LLM judgments (15.2%). The deficit remained under prompt and specialist sensitivity analyses. Among CWRU cases solved under both evidence presentations, task-aligned physical evidence yielded lower estimates of susceptibility to incorrect advice in six of seven models (five intervals excluding zero); higher reasoning effort gave no reliable reduction in five models, and a separate four-model TEP analysis gave no clear evidence that it resolves the integration problem. Source quality and integration quality should be evaluated separately: an integration layer should be compared with its stronger standalone component, not only with the unaided LLM.