TELEMETRY UC2 (Manufacturing) sample dataset for the Misuse Detection ML Toolkit: short 7-axis right-arm slices to illustrate model training

Overview This dataset contains one-second-resolution time-series telemetry recorded from the right arm of a 7-axis industrial robot during healthy manufacturing operation. It supports unsupervised detection of misuse or abnormal robot-arm behaviour (unexpected speed, force, or trajectory) on multivariate sensor time series, and is intended for use with the open-source Misuse Detection ML Toolkit (TELEMETRY Use Case 2 - Manufacturing).The three files are unlabeled normal operation. They are disjoint midday slices from the same programmed-speed recording (speed 0.15), two calendar days apart, so they can be used as train, validation, and test without leakage.This release is sufficient to smoke-test the toolkit pipeline (prepare → train/assess → infer). It is not a production training set. Sensor architecture Each sample is one timestamp plus 21 channels: speed, torque, and joint angle on axes 1-7 of the right arm. Values are as recorded by the robot controller. Group Columns Axes Speed RobotSpeed_axisN_right 1-7 Torque RobotTorq_axisN_right 1-7 Joint angle RobotJointAngle_axisN_right 1-7 The models learn to reconstruct these channels under normal motion. Misuse or abnormal behaviour is expressed as high reconstruction error on one or more channels. Files Each CSV has 12,000 rows (one row per second, ~3.3 hours) and 22 columns. Sampling is ~1 Hz. The files are unlabeled. Column Description timestamp Sample time (`YYYY-MM-DD HH:MM:SS.sss RobotSpeed_axisN_right Joint speed, axis N, right arm (N = 1…7) RobotTorq_axisN_right Joint torque, axis N, right arm (N = 1…7) RobotJointAngle_axisN_right Joint angle, axis N, right arm (N = 1…7) Splits robot001_normal_train.csv - Nominal operation (real data). Midday block on 26 Oct 2024 (12:00–15:20). Healthy motion at programmed speed 0.15. Training baseline for the 21 per-channel convolutional autoencoders. robot001_normal_validation.csv - Nominal operation (real data). Midday block on 28 Oct 2024 (12:00–15:20). Same operating condition as train; disjoint in time. Used for early stopping and reconstruction-error thresholds. robot001_normal_test.csv - Nominal operation (real data). Midday block on 30 Oct 2024 (12:00–15:20). Same operating condition; disjoint in time. Used for inference on held-out healthy motion. Supplementary figures Three figures are included (robot001_normal_train.png, robot001_normal_validation.png, robot001_normal_test.png). Each shows axis-1 speed, torque, and joint angle for the corresponding slice. The traces are periodic manufacturing cycles: repeating joint-angle paths, bursty speed, and torque with occasional sharp peaks that belong to normal operation.