RelaySec-FEL

Relay Secondary-Circuit Fault Element Localization Dataset

Production field measurements · 20,000 diagnostic cases · 8 substations · 2021–2025

Versioned archive · All-version DOI · Browse files

RelaySec-FEL contains measurements and operational records collected from relay-protection secondary circuits during production operation. It supports element-level fault localization: ranking the devices, ports, physical links, and virtual circuits most likely to explain the observations in a diagnostic window.

The release brings together circuit structure, communication and protection events, inspection records, and fault labels. It is intended for research on graph-based diagnosis, multi-source evidence fusion, candidate ranking, and evidence-chain explanation under incomplete or conflicting observations.

Dataset overview

Property	Release v1.0.1
Diagnostic cases	20,000
Record period	January 2021–December 2025
Substations / bays	8 / 128
Unified event observations	156,523
Static graph nodes	3,944
SCD structural relations	7,608
Stored graph edges, including reverse relations	15,216
Fault-element labels	20,000
Inspection records	19,798
Defect and maintenance work orders	8,917
Data format	UTF-8 CSV; graph triples in TSV
License	CC BY 4.0

The static graph comprises 528 devices, 1,536 ports, 768 physical links, 600 virtual circuits, and 512 logical functions. Logical functions provide structural context; the localization targets are devices, ports, physical links, and virtual circuits.

Exact table counts and structural validation statistics are provided in metadata.json.

Data source and organization

The underlying data are real measurements and records from production operation. The release organizes these records into relational tables and diagnostic cases, together with a static graph derived from the circuit-structure tables. Measurement records, derived representations, and supervision tables have distinct roles in the diagnosis workflow.

The records cover protection events, sequence-of-events (SOE) logs, communication status, GOOSE and sampled-value records, waveform features, device self-checks, configuration checks, normal or recovered states, and inspections. Observations retain source, time, target, quality, and evidence-reliability fields where provided.

Controlled perturbations introduced for robustness experiments should be reported separately from the original production records.

Task definition

For each case in fault_cases.csv, combine the relevant circuit structure and available observations to return a ranked list of candidate fault elements:

Circuit structure + case observations + pre-fault inspection context
                              ↓
             Ranked candidate fault elements
                              ↓
          Supporting and opposing evidence paths

The candidate types are Device, Port, PhysicalLink, and VirtualCircuit. An observation target identifies where a symptom was recorded; the reference fault element identifies the failed component. These may differ because secondary circuits couple physical connections with logical communication mappings.

fault_labels.csv supplies the reference element for supervision and evaluation. Labels and maintenance outcomes are kept separate from inference inputs.

Files and relationships

Data layer	Files	Role
Circuit structure	stations.csv, bays.csv, asset_devices.csv, asset_ports.csv, physical_links.csv, logical_functions.csv, virtual_circuits.csv, scd_relations.csv	Describe assets, physical connections, logical functions, and communication mappings.
Entity alignment	entity_alignment.csv	Map business-table identifiers to graph nodes.
Diagnostic cases	fault_cases.csv	Define case identifiers, time windows, station/bay context, and evaluation splits.
Unified evidence	event_observations.csv	Provide the common event representation and observation targets.
Source-specific records	protection_events.csv, soe_events.csv, communication_records.csv, goose_records.csv, sampled_value_records.csv, waveform_features.csv, self_check_records.csv, configuration_records.csv, normal_status_records.csv	Preserve event details and measurements from individual record types.
Inspection context	inspection_records.csv	Supply health observations available before a case begins.
Supervision and verification	fault_labels.csv, defect_workorders.csv	Supply fault-element labels and maintenance outcomes for training, evaluation, and post-hoc verification.
Derived static graph	static_graph/base_nodes.csv, static_graph/base_edges.csv, static_graph/base_triples.tsv	Provide node, edge, and triple representations of the circuit structure.

Core fields

Table	Key fields	Interpretation
fault_cases.csv	sample_id, station_id, bay_id, case_start_time, case_end_time, window_start_time, window_end_time, split	Case identity, operating context, diagnosis window, and prescribed partition.
event_observations.csv	event_id, sample_id, event_time, ingest_time, target_node_id, target_node_type, evidence_source, event_type, value_state, quality_flag	Event identity, case membership, timing, target, source, state, and data quality.
event_observations.csv	source_reliability, time_consistency, conflict_degree, credibility	Evidence attributes available for reliability-aware modeling; these are not fault probabilities.
entity_alignment.csv	node_id, node_type, field_table, primary_key	Correspondence between graph nodes and business records.
fault_labels.csv	sample_id, fault_node_id, fault_node_type, fault_mode, label_source, verification_method, label_confidence	Reference fault element and label documentation.
inspection_records.csv	node_id, inspection_time, health_score, status_code, defect_level	Time-stamped inspection context associated with a graph node.

Join cases and unified observations through sample_id. Match target_node_id, fault_node_id, and inspection node_id to graph node_id; use entity_alignment.csv for business identifiers. The unified evidence and source-specific tables provide related views of the records and should not be treated as independent evidence without checking their identifiers.

Complete table inventory and row counts

File	Rows
fault_cases.csv	20,000
asset_devices.csv	528
asset_ports.csv	1,536
bays.csv	128
communication_records.csv	113,441
configuration_records.csv	10,469
defect_workorders.csv	8,917
entity_alignment.csv	3,944
event_observations.csv	156,523
fault_labels.csv	20,000
goose_records.csv	17,051
inspection_records.csv	19,798
logical_functions.csv	512
normal_status_records.csv	4,249
physical_links.csv	768
protection_events.csv	6,793
sampled_value_records.csv	7,133
scd_relations.csv	7,608
self_check_records.csv	4,958
soe_events.csv	3,438
static_graph/base_edges.csv	15,216
static_graph/base_nodes.csv	3,944
static_graph/base_triples.tsv	19,160
stations.csv	8
virtual_circuits.csv	600
waveform_features.csv	7,133

Load the data

Local archive

Extract RelaySec-FEL-v1.0.1.zip and read the tables from the relaysec-fel/ directory:

from pathlib import Path
import pandas as pd

root = Path("relaysec-fel")
cases = pd.read_csv(root / "fault_cases.csv")
events = pd.read_csv(root / "event_observations.csv")
nodes = pd.read_csv(root / "static_graph" / "base_nodes.csv")
labels = pd.read_csv(root / "fault_labels.csv")  # Supervision/evaluation only

Hugging Face Datasets

Each table has its own configuration because the tables have different schemas. The default configuration is fault_cases.

from datasets import load_dataset

repo_id = "james12221/RelaySec-FEL"
cases = load_dataset(repo_id, "fault_cases", split="data")
events = load_dataset(repo_id, "event_observations", split="data")

train_cases = cases.filter(lambda row: row["split"] == "train")
val_cases = cases.filter(lambda row: row["split"] == "val")
test_cases = cases.filter(lambda row: row["split"] == "test")

Here, data is the storage split for a complete table. The benchmark partitions are the values in the split column of fault_cases.csv, not the storage split name. Other configurations use the file stem, such as asset_devices, fault_labels, base_nodes, and base_edges.

pandas

For direct access to individual CSV tables:

import pandas as pd

base = "https://huggingface.co/datasets/james12221/RelaySec-FEL/resolve/main"
cases = pd.read_csv(f"{base}/fault_cases.csv")
events = pd.read_csv(f"{base}/event_observations.csv")
nodes = pd.read_csv(f"{base}/static_graph/base_nodes.csv")

case_id = cases.loc[cases["split"].eq("test"), "sample_id"].iloc[0]
case_events = events.loc[events["sample_id"].eq(case_id)]

The release contains data and graph representations. Model implementations are supplied separately. For reproducible experiments, report the dataset version and cite the versioned archive below.

Evaluation protocol

Partition	Diagnostic cases	Share
Training (train)	14,027	70.135%
Validation (val)	2,948	14.740%
Test (test)	3,025	15.125%

Use the supplied case partitions for comparable results. Report Top-1 accuracy, Hit@k, and mean reciprocal rank (MRR) over the candidate ranking, together with the candidate-scope rule and evaluation sample count. If MRR is truncated, state the cutoff explicitly. For different partitions, document how they were formed; the supplied case split should not be assumed to establish chronological or unseen-station generalization.

To prevent information leakage:

1. Use fault_labels.csv for supervision and evaluation, and defect_workorders.csv for post-hoc verification. Exclude both from inference features, evidence aggregation, and candidate selection.
2. Restrict observations to the information available at the diagnosis cutoff. Account for event time and ingestion time when defining online evaluation.
3. Use an inspection record only when inspection_time < case_start_time.
4. Fit learned preprocessing and model parameters on training data. Do not derive input features from validation/test fault labels or later maintenance outcomes.

Data quality and scope

Production records contain missing observations, suspect quality flags, delayed timestamps, repeated events, normal or recovered states, and cross-source conflicts. These attributes support evaluation of diagnosis methods under imperfect evidence and should be handled explicitly rather than silently discarded.

The release metadata reports zero event targets missing from the graph and zero labels without a corresponding diagnostic case. These are structural consistency checks; they do not imply that every record is complete or every label is certain. Source, quality, verification, and confidence fields should be considered when interpreting individual cases.

Results on the released cases describe performance under the reported evaluation protocol. Generalization to other substations, configuration changes, and operating conditions requires separate evaluation. Candidate rankings and evidence paths support inspection prioritization; they do not, by themselves, establish a physical fault cause.

Citation

Please cite the versioned dataset archive:

Zeus. (2026). RelaySec-FEL: Relay Secondary-Circuit Fault Element Localization Dataset (Version 1.0.1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23492959

@dataset{relaysec_fel_2026,
  author    = {Zeus},
  title     = {{RelaySec-FEL}: Relay Secondary-Circuit Fault Element Localization Dataset},
  year      = {2026},
  version   = {1.0.1},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23492959},
  url       = {https://doi.org/10.5281/zenodo.23492959}
}

The all-version DOI resolves to the latest archived release.

License and contact

The dataset is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0), as registered in the Zenodo archive. Attribute the dataset creator and cite the dataset when using or redistributing it.

For questions about the tables, evaluation protocol, or data quality, open a thread in the dataset discussion area.