Low-power laser back-scatter signals from nine industrial materials

Abstract

986 laser back-scatter scan records acquired from nine industrial materials with a sub-milliwatt Class II laser diode at 635 nm, together with a record-level manifest. The dataset accompanies the article Optical non-destructive identification of industrial materials from low-power laser back-scatter. Acquisition. A two-axis Cartesian stage positioned an M703A laser distance sensor over 40 × 40 × 2 mm specimens and scanned a 20 × 20 mm central area at normal incidence and constant stand-off. Each specimen was sampled at four static corner points and along the four edges and two diagonals joining them. Records were collected between 19 May 2025 and 6 September 2025. Fourteen records acquired with the laser inactive were removed before release, which is why the two PETG folders hold 93 files each. Contents. 986 comma-separated records, 3,943,999 samples in total, one file per scan, organised as data/ _ /. A manifest.csv gives one row per record (path, material, label, sample count, first timestamp, PETG colour variant). Materials and record counts. Glass 100 · PETG 186 (green and blue printed variants pooled into one class) · aluminium 100 · brass 100 · ABS 100 · plywood 100 · fabric-1 100 · fabric-2 100 · glazed ceramic tile 100. PETG was printed in two colours to test whether the measurement responds to the polymer matrix rather than to pigment. File format. UTF-8 CSV with a header row and one row per sample: - DIST — distance reported by the sensor, mm - AMP — back-scattered return amplitude, 12-bit ADC counts (0–4095); this is the measurement channel - TEMP — internal sensor temperature, 0.01 °C - VOLT — emitter drive level, arbitrary units - Timestamp — host-side acquisition time, ISO 8601 - Angle — probe orientation in degrees (90° = normal incidence) - ON — laser emitting - Movement — stage in motion during the sample The detector runs at an average physical rate of 56.7 Hz and the host writes samples irregularly, so timestamps are not evenly spaced; the accompanying analysis resamples each record onto a uniform 500 Hz grid. Reproducing the published analysis. Each record is divided into twelve non-overlapping epochs; 291 descriptors in eight physically interpretable families (radiometric, spectral, wavelet, cepstral, autoregressive, complexity, distribution and detector-state) are computed per epoch; a histogram-based gradient boosting ensemble classifies the epochs; and the record decision is the mean of the twelve epoch posteriors. Under five-fold cross-validation this gives 93.10 % accuracy and 93.42 % balanced accuracy over the nine classes. Funding. Scientific and Technological Research Council of Türkiye (TÜBİTAK), grant 1139B412405219. Licence. Creative Commons Attribution 4.0 International. Please cite both the dataset and the accompanying article.