THULab/pudu_robot_operation_logs_bau_capstone_2026

PUDU Robot Operation Logs for Predictive Maintenance (TsFile)

Apache TsFile version of Lightcap/pudu-robot-operation-logs-bau-capstone-2026.

Overview

Raw robot-operation database snapshot used by the Bahçeşehir University Faculty of Engineering and Natural Sciences capstone project Fault Prediction and Diagnostics from Robot Operation Logs (2025-2026 Spring semester, Computer Engineering + Industrial Engineering). The dataset supports predictive maintenance, fault detection, diagnostics, robot fleet monitoring, KPI reporting, and reproducible evaluation of LSTM and Random Forest models trained on PUDU robot operational logs.

DOI: 10.57967/hf/8635

• Central error/event log: 147,488 raw records from 46 robots, covering 2025-05-03 to 2026-02-26.
• Robot info/status log: 3,893 records from 48 robots across the same deployment window.
• Access: the original dataset publishes the raw PostgreSQL tables as Parquet for dashboards and modeling without a live database connection.
• Model context: the registered LSTM model uses 10-timestep sequences with 8 engineered features (error_count, task_hour_num, day_of_month, day_of_week, robot_id_length, software_version_length, product_code_type, hourly_error_rate). Registered test-set performance: Accuracy 98.94%, Recall 95.76%, Precision 85.30%, F1 90.23%, AUC-ROC 99.33%. Training code: github.com/DrGb24/pudu_bot_model_training; dashboard: github.com/farukalpay/pudu-robot-operation-dashboard.
• Team: Faruk Alpay, Nazım Alp Batu Kardaş, Buğra Kılıçtaş (Computer Engineering); Bilge Ece Şentürk, Serra Uysal, Ayça Yeralp (Industrial Engineering). Advisors: Assoc. Prof. Cemal Okan Şakar and Assoc. Prof. Adnan Çorum.

Schema (TsFile structure)

robot_error.tsfile — hourly error/event log (147,488 rows, 46 devices)

• Time (INT64, milliseconds) — the source task_hour hourly bucket (UTC). Multiple records for the same robot-hour are separated by the minimum 1 ms offset so each device timeline is strictly increasing; no rows are dropped.
• robot_id (TAG) — robot identifier; query one robot with WHERE robot_id='8170G3C08050003'.
• ingest_id (FIELD, INT64), hourly_error_count (FIELD, INT64), pair_max_hourly_count (FIELD, INT64), noise_threshold (FIELD, INT64) — counts/thresholds.
• hourly_ratio (FIELD, DOUBLE) — error ratio for the hour.
• error_type (FIELD, STRING), error_level (FIELD, STRING), error_detail (FIELD, STRING) — fault category, severity and detail (NavigationSpeedLimit is the most frequent type).
• product_code (FIELD, STRING), sn (FIELD, STRING), mac (FIELD, STRING), soft_version (FIELD, STRING), hard_version (FIELD, STRING), os_version (FIELD, STRING) — robot identity and software/hardware versions.
• task_time, upload_time, inserted_at (FIELD, STRING) — exact task time, upload time and DB load time, preserved as ISO-8601 UTC strings. record_id, error_id, source_file, source_group (FIELD, STRING) — record provenance keys.

robot_info.tsfile — robot info/status log (3,893 rows, 48 devices)

• Time (INT64, milliseconds) — the source task_time (device check time, UTC).
• robot_id (TAG) — robot identifier; query one robot with WHERE robot_id='...'.
• is_success (FIELD, INT64) — 1/0 success flag of the self-check; check_result_count (FIELD, INT64) — number of check steps.
• check_result_json (FIELD, STRING) — JSON array of per-subsystem check results (kept as text; up to ~9 KB per row).
• sn, mac, product_code, ip, soft_version, hard_version, os_version (FIELD, STRING) — robot identity/network/version metadata.
• upload_time, inserted_at (FIELD, STRING) — upload and DB load timestamps as ISO-8601 UTC strings. ingest_id, record_id, source_file, source_group (FIELD, STRING) — provenance keys.

Type mapping: source integer columns stay INT64 and source double stays DOUBLE; text and all auxiliary timestamp columns become STRING. The two tables are independent files, each partitioned into one time series per robot_id.

Usage

Install the Apache TsFile Python SDK (pip install tsfile) and read a converted file:

from pathlib import Path
from tsfile import TsFileReader

path = Path("robot_error.tsfile")
with TsFileReader(str(path)) as reader:
    schemas = reader.get_all_table_schemas()
    print("tables:", list(schemas))
    table_name = next(iter(schemas))
    table = schemas[table_name]
    columns = [column.get_column_name() for column in table.get_columns()]
    print("columns:", columns)
    field_names = [
        column.get_column_name()
        for column in table.get_columns()
        if column.get_column_name() not in {"Time", "time"}
    ]
    if field_names:
        with reader.query_table(table_name, field_names[:3], batch_size=1024) as result:
            batch = result.read_arrow_batch()
            if batch is not None:
                print(batch.to_pandas().head())

Source & license

• Original dataset: https://huggingface.co/datasets/Lightcap/pudu-robot-operation-logs-bau-capstone-2026
• Author / publisher: Faruk Alpay, Nazım Alp Batu Kardaş, Buğra Kılıçtaş, Bilge Ece Şentürk, Serra Uysal, Ayça Yeralp (Bahçeşehir University)
• Paper: none (capstone project report)
• License: CC BY 4.0

Conversion notes

• Two time-series tables converted. public_robot_logs_error.parquet → robot_error.tsfile (147,488 rows; Time = task_hour, TAG = robot_id) and public_robot_logs_info.parquet → robot_info.tsfile (3,893 rows; Time = task_time, TAG = robot_id). No rows or columns were dropped from either table.
• The error table is the complete log, not one of the split files. The repository also ships public_robot_logs_error_training/validation/test.parquet, but they are an overlapping model-development partition: they share up to 15,573 rows with each other and together cover only 115,478 unique rows — i.e. they omit 32,010 rows that exist in the full 147,488-row log. Converting the split files separately would therefore be lossy, so the authoritative complete log is converted instead ("merged"). The split files are not published here as separate TsFiles.
• Model-training metadata tables are not converted. data/model_training_training_artifacts.parquet (3 rows), model_training_training_runs.parquet (1 row) and model_training_training_source_tables.parquet (5 rows) are model-training inventory/metadata with no time-series semantics; they are not included as TsFiles.
• Not included at all: model_artifacts/** (trained .h5 / .pkl / .npy / .json models, scalers and configs), assets/*.png|jpg (figures and field photos), reference/HATA_KODLARI_ROBOT.xlsx (error-code reference workbook) and SHA256SUMS.txt. These are not time series; the trained model artifacts and images remain available in the original dataset.
• Auxiliary timestamps (upload_time, inserted_at, and task_time when it is not the Time column) are stored as ISO-8601 UTC STRING fields rather than numeric timestamps; the Time column carries the primary event time.