Nasochec/ECRv1

Eddy Current Rail

Dataset description

This dataset contains multichannel discrete signals — eddy current defectograms — collected during non-destructive testing (NDT) of railway tracks. An eddy current flaw detector records the interaction of eddy currents with the conductive rail material at every millimeter of the track using an array of physical sensors, providing spatial structure to the measurements.

Automated defectogram analysis requires isolating useful signals (constructive track elements and rail defects) against a background of rail noise and hardware interference. Defect signals must be reliably distinguished from background noise, signals generated by constructive elements, and various forms of interference, such as low-frequency periodic or square wave noise. Such interference must be suppressed using appropriate methodological approaches.

The primary characteristic of a defect signal is typically a high amplitude value. Constructive elements, in addition to exhibiting high amplitudes, possess characteristic signal shapes. Consequently, accurately detecting constructive rail elements in multichannel signals is a critical prerequisite for automated flaw detection systems, allowing them to map track infrastructure and distinguish normal constructive elements from actual rail anomalies prior to defect identification.

This dataset can be utilized to train models for the task of object detection within one-dimensional, multichannel eddy current defectogram signals. Currently, ECRv1 exclusively contains annotations for constructive element signals: flash-butt welds, aluminothermic welds, bolted joints, axle counters, crossings, and manual arc welds. In the future, the dataset is planned to be supplemented with defect annotations. Beyond object detection training, the dataset may also be employed to investigate other fundamental properties of the signals.

• Curated by: Artemy N. Gladkov, Leonid Y. Bystrov, and Egor V. Kuzmin (Department of Theoretical Computer Sciences, P.G. Demidov Yaroslavl State University)
• Funded by: P.G. Demidov Yaroslavl State University
• License: CC-BY-4.0

Source data and acquisition equipment

The dataset does not assume a specific eddy-current flaw detector model; the hardware model is intentionally undisclosed to abstract the data from any specific equipment. It represents measurements from an abstract, generalized multichannel flaw detector. All recordings are uniformly 15-channel. Each channel is represented by an individual sensor, and all sensors are identical in their design. These sensors are positioned perpendicularly to the direction of travel along the rail head, which facilitates the extraction of information regarding the spatial distribution of defects on the rail. The scanning step is strictly fixed at 1 millimeter of the railway track. The raw signal values output by the equipment are encoded and stored as 16-bit signed integers (int16). The range of accepted values is bounded from -32,768 to 32,767.

The signal recordings were collected from various countries across different regions of the world at different times. However, the precise geographic coordinates of the track sections and their affiliations with specific infrastructure operators are not disclosed to maintain compliance with confidentiality agreements established with the providing railway companies. Recordings are split into anonymised passages (passage1 ... passage89): one passage constitutes a single testing series recorded by one device along a specific railway section at a given time. The dataset contains a total of 89 passages. Each passage contains signals of both rails (left, right), cut into numbered fragments (blocks).

The dataset annotation was conducted using a hybrid approach to ensure high quality and comprehensive coverage. The ground truth labels were obtained through a combination of three methods:

• Algorithmic Baseline: Initial labeling using existing analysis algorithms.

• Expert Domain Knowledge: Annotations and verified records provided directly by professional NDT (non-destructive testing) inspectors.

• Human-in-the-Loop Pseudo-Labeling: An initial neural network was trained on a small, manually annotated subset of the data. This model was then used to generate preliminary predictions (pseudo-labels) for the unannotated records. Finally, all model-generated predictions were thoroughly reviewed, verified, and manually corrected by human annotators.

Data format

Each signal fragment is a NumPy array stored as .npy:

property	value
dtype	int16
shape	(channels, length), channels = 15
sampling	1 sample = 1 mm along the track
length	50 000 samples (50 m) for most fragments, some are shorter (min 4 409)

Label format

For every data file there is a text file with the same name in labels/. One line per object:

type begin end

• type — class id (see the table below);
• begin, end — object boundaries in samples (= mm) from the start of the fragment, 0 ≤ begin ≤ end ≤ length. Objects that cross the fragment edge are clipped to it.

Lines are sorted by begin. All published fragments contain at least one object.

Classes

The annotations within the dataset represent 6 constructive elements classes.

• Flash-butt weld (1), Aluminothermic weld (2), and Bolted joint (3): These constitute short local objects with a minor spatial extent, extending up to 200 samples (20 cm). The signals of these objects are characterized by synchronous spikes across all defectogram channels and demonstrate a clear, distinct spatial center. Although the physical objects may slightly vary in actual length, their bounding box width is strictly unified to 200 mm (200 samples). For the detection task, they are strictly localized by their center point. Welds are highly prevalent in defectograms as they connect two rail fragments; they are typically situated at fixed intervals dictated by the specific type of rail in use. Bolted joints similarly connect rail sections but are encountered much less frequently than welded joints.
• Axle counter (4): An extended object (up to 3 meters) featuring a highly asymmetric signal. The signal is defined by a sharp, high-amplitude spike at the beginning of the object, which is then followed by a smooth exponential decay that blurs the right boundary.
• Crossing (5): An object of variable length recorded exclusively on one rail thread. The primary frog signal manifests as a low-frequency oscillation with a spatial shift between channels, and it is frequently flanked by spikes originating from adjacent joints.
• Manual arc (12): A local repair weld (overlay) utilized to eliminate surface defects on the rail head. These objects can vary in spatial extent from 50 to 655 samples. They exhibit sharp, high-amplitude spikes specifically at the boundaries of the object. Notably, these spikes appear only on several adjacent channels that physically coincide with the repair zone rather than all channels, as the repair overlay is applied exclusively to the damaged portion of the rail.

Class ids are kept as in the original markup, so they are not contiguous.

Directory structure

train/
  data/passage{K}_{side}_{NNNN}.npy
  labels/passage{K}_{side}_{NNNN}.txt
val/
  data/  labels/
test/
  data/  labels/
annotations/
  passage{K}.json
README.md

side is left or right, NNNN is the fragment number within the passage and side (4 digits, zero-padded). Numbers are not contiguous within a split: neighbouring fragments may belong to other splits or be unpublished.

Passage annotations

annotations/passage{K}.json describes all fragments of a passage, including those not published (split: null), so that gaps between published fragments can be estimated.

Passage-level fields:

field	description
format_version	annotation format version (currently 1)
passage	passage name
channels, dtype	array layout of the data files
sample_spacing_mm	distance between samples, mm (1)
offset_origin	origin of offset_mm: the earliest sample of both rails along the direction of travel
files	list of fragments (see below)

Fragment fields:

field	description
file	file name without extension (passage1_left_0002)
side, number	rail side and fragment number
split	train / val / test, or null if the fragment is not published
length	number of samples in the data file
block_length	fragment length reported by the acquisition system (may differ from length by a few samples)
offset_mm, end_offset_mm	position of the first and last sample along the direction of travel, mm from offset_origin
gap_to_prev_mm	distance from the previous fragment of the same side: 0 — contiguous, > 0 — gap, < 0 — overlap, null — first fragment
contiguous_with_prev	gap_to_prev_mm == 0

Left and right rails share the same origin, so their offsets are directly comparable. Absolute positions (odometer) are not published.

Splits and statistics

Fragments are assigned to splits by ranges of fragment numbers within each passage (with a few individual exceptions). The data is pre-partitioned into training, validation, and testing sets following a 71:19:10 ratio.

	train	val	test	total
passages	56	25	16	89
signals (files)	2099	560	294	2953
signal length, km	103.558	27.658	14.377	145.592
objects	8256	2183	1207	11646

id	class	train	val	test	total	share
1	Flash-butt weld	3401 (72%)	856 (18%)	481 (10%)	4738	40.7%
2	Aluminothermic weld	3538 (70%)	979 (19%)	543 (11%)	5060	43.4%
3	Bolted joint	253 (69%)	73 (20%)	43 (12%)	369	3.2%
4	Axle counter	475 (73%)	116 (18%)	58 (9%)	649	5.6%
5	Crossing	285 (71%)	77 (19%)	39 (10%)	401	3.4%
12	Manual arc	304 (71%)	82 (19%)	43 (10%)	429	3.7%

License

The dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Copyright (c) 2026 P.G. Demidov Yaroslavl State University. Developed by Gladkov A. N., Bystrov L. Y., Kuzmin E. V.

Citation

@article{Gladkov2027,
  title   = {Neural Network YOLO-1D for Useful Signals Detection on Eddy Current Rail Defectograms},
  author  = {Gladkov A.N. and Bystrov L.Y. and Kuzmin E.V.},
  note    = {Submitted to Trudy ISP RAN/Proc. ISP RAS}
}