--- license: eupl-1.2 pretty_name: "LeakDB: Leakage Diagnosis Benchmark for water distribution networks" language: - en task_categories: - tabular-classification - time-series-forecasting tags: - water-distribution-networks - leak-detection - leakage-diagnosis - anomaly-detection - fault-diagnosis - benchmark - time-series - EPANET - hydraulic-simulation - critical-infrastructure - synthetic size_categories: - 10M) for the Hugging Face Hub. Values are copied unchanged; the conversion only restructures files, adds types and timestamps, and adds quality-control flags. ## Quick start ```python from datasets import load_dataset import pandas as pd train = load_dataset("KIOSCoE/LeakDB", "net1_timeseries", split="train") # also "validation", "test"; or "hanoi_timeseries" scen = load_dataset("KIOSCoE/LeakDB", "scenarios", split="train").to_pandas() leaks = load_dataset("KIOSCoE/LeakDB", "leaks", split="train").to_pandas() leaks = leaks[~leaks.stale_file] # leaks that belong to each simulation # one scenario as a (time x sensors) frame, reading only that scenario's row group df = pd.read_parquet("hf://datasets/KIOSCoE/LeakDB/data/net1_timeseries", filters=[("scenario", "==", 42)]) ``` Each Parquet row group holds exactly one scenario, so filtering by `scenario` reads only that scenario. `splits.csv` lists the split of every scenario. ## Networks and simulation | | Net1_CMH | Hanoi_CMH | |---|---|---| | Scenarios | 1000 | 1000 | | Nodes with pressure/demand series | 11 | 32 | | Links with flow series | 13 | 34 | | Sources | reservoir 9, tank 2 | reservoir 1 | - Duration: one year per scenario (17,520 steps of 30 min; time stamps use the synthetic year 2017). - Simulator: WNTR, pressure-driven demand; EPANET units CMH (pressures in m, demands and flows in m³/h). - Per scenario: pipe length, diameter and roughness perturbed within 5–20 %; synthetic demands; 0–2 leaks (abrupt or incipient) at random nodes, start and end times. - Base network models are in `networks/`. Each scenario’s perturbed EPANET model is in `models//` (gzipped .inp). ## Data structure Every config has the splits `train`, `validation` and `test`, assigned at **scenario** level: all rows of a scenario (time steps, leaks, leak demands) are in the same split, so no information leaks between splits. Assignment is 70/15/15 per network, stratified by the scenario's valid leaks (none / abrupt / incipient / two leaks), seed 42; the full list is in `splits.csv` and in the `assigned_split` column of `scenarios`. | network | train | validation | test | |---|---|---|---| | Net1_CMH | 700 (514 with leaks) | 151 (111 with leaks) | 149 (109 with leaks) | | Hanoi_CMH | 700 (534 with leaks) | 150 (115 with leaks) | 150 (114 with leaks) | Use `train` to fit models, `validation` to tune thresholds and hyper-parameters, and report results on `test`. For comparability with the original 2018 benchmark (which scored detectors on all scenarios), also report scores on the concatenation of all three splits. ### `net1_timeseries` Net1_CMH: one row per scenario and 30-min time step with the leak label and all node pressures, node demands and link flows. 17,520,000 rows, 21 file(s), 0.99 GB. | column | type | description | |---|---|---| | `scenario` | int16 | Scenario number (1-1000) within the network. | | `time_index` | int32 | 1-based time-step index as in the original CSV files (1-17520). | | `timestamp` | timestamp[ms] | Simulation time stamp (synthetic year 2017, 30-min steps, no time zone). | | `label` | int8 | 1 if at least one leak is active at this time step, else 0 (original Labels.csv). | | `pressure_node_` (11 columns) | double | Pressure head at each node, m. IDs: 2, 9, 10, 11, 12, 13, 21, 22, 23, 31, 32. | | `demand_node_` (11 columns) | double | Actual (pressure-driven) demand at each node, m^3/h; negative at reservoirs/tanks supplying the network. IDs: 2, 9, 10, 11, 12, 13, 21, 22, 23, 31, 32. | | `flow_link_` (13 columns) | double | Flow at each link, m^3/h; sign gives direction relative to the link's start node. IDs: 9, 10, 11, 12, 21, 22, 31, 110, 111, 112, 113, 121, 122. | ### `hanoi_timeseries` Hanoi_CMH: one row per scenario and 30-min time step with the leak label and all node pressures, node demands and link flows. 17,520,000 rows, 20 file(s), 2.96 GB. | column | type | description | |---|---|---| | `scenario` | int16 | Scenario number (1-1000) within the network. | | `time_index` | int32 | 1-based time-step index as in the original CSV files (1-17520). | | `timestamp` | timestamp[ms] | Simulation time stamp (synthetic year 2017, 30-min steps, no time zone). | | `label` | int8 | 1 if at least one leak is active at this time step, else 0 (original Labels.csv). | | `pressure_node_` (32 columns) | double | Pressure head at each node, m. IDs: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32. | | `demand_node_` (32 columns) | double | Actual (pressure-driven) demand at each node, m^3/h; negative at reservoirs/tanks supplying the network. IDs: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32. | | `flow_link_` (34 columns) | double | Flow at each link, m^3/h; sign gives direction relative to the link's start node. IDs: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34. | ### `scenarios` One row per scenario: parameter-uncertainty settings, leak counts and quality-control flags. 2,000 rows, 3 file(s), 0.00 GB. | column | type | description | |---|---|---| | `network` | string | Network name: Net1_CMH or Hanoi_CMH. | | `scenario` | int16 | Scenario number (1-1000) within the network. | | `assigned_split` | string | Split this scenario belongs to (train, validation or test); same as splits.csv. | | `duration_hours` | int32 | Simulated duration in hours. | | `time_step_min` | int32 | Hydraulic and reporting time step in minutes. | | `uncertainty_topology` | string | Topology uncertainty setting (NO for all scenarios). | | `uncertainty_length_pct` | double | Maximum relative perturbation of pipe lengths, percent. | | `uncertainty_diameter_pct` | double | Maximum relative perturbation of pipe diameters, percent. | | `uncertainty_roughness_pct` | double | Maximum relative perturbation of pipe roughness coefficients, percent. | | `scenario_label` | int8 | 1 if any time step of the scenario is labelled as leaking. | | `n_labelled_steps` | int32 | Number of time steps with label 1. | | `n_label_events` | int32 | Number of contiguous runs of label 1. | | `n_leak_files` | int32 | Number of leak info files in the original Leaks folder. | | `n_valid_leaks` | int32 | Leak info files that belong to this scenario's simulation (n_leak_files minus stale files). | | `n_stale_leak_files` | int32 | Leak info files left over from an earlier simulation run (see stale_file). | | `labels_explained_by_leaks` | bool | True if the label series equals the union of the valid leak windows. | | `qc_notes` | string | Human-readable quality-control notes added during conversion. | ### `leaks` One row per leak info file: location, type, size, start/peak/end and quality-control flags. 1,993 rows, 3 file(s), 0.00 GB. | column | type | description | |---|---|---| | `network` | string | Network name: Net1_CMH or Hanoi_CMH. | | `scenario` | int16 | Scenario number (1-1000) within the network. | | `leak_node` | string | Node ID where the leak is located. | | `leak_type` | string | abrupt (constant size from onset) or incipient (growing until peak time). | | `leak_area_m2` | double | Leak (orifice) area at its maximum size, m^2. | | `leak_diameter_m` | double | Leak (orifice) equivalent diameter at its maximum size, m. | | `start_index` | int32 | Original 'Leak Start' time-step index. Abrupt leaks start one step later (start_index + 1). | | `end_index` | int32 | Original 'Leak End' time-step index (last labelled step). | | `peak_index` | int32 | Original 'Peak Time' index: step at which an incipient leak reaches its maximum size. | | `start_time` | timestamp[ms] | Timestamp of start_index. | | `end_time` | timestamp[ms] | Timestamp of end_index. | | `peak_time` | timestamp[ms] | Timestamp of peak_index. | | `first_outflow_index` | int32 | First time step with non-zero leak outflow in the leak demand series (derived). | | `last_outflow_index` | int32 | Last time step with non-zero leak outflow in the leak demand series (derived). | | `max_outflow_m3h` | double | Maximum leak outflow, m^3/h (derived). | | `label_coverage` | double | Fraction of the leak window that is labelled 1 in Labels.csv (derived). | | `stale_file` | bool | True if this leak file is older than the scenario's simulation output, i.e. left over from an earlier run and not part of the simulated scenario. | ### `leak_demands` Leak outflow time series, one row per leak and time step. 34,917,360 rows, 41 file(s), 0.32 GB. | column | type | description | |---|---|---| | `network` | string | Network name: Net1_CMH or Hanoi_CMH. | | `scenario` | int16 | Scenario number (1-1000) within the network. | | `leak_node` | string | Node ID where the leak is located. | | `time_index` | int32 | 1-based time-step index as in the original CSV files (1-17520). | | `timestamp` | timestamp[ms] | Simulation time stamp (synthetic year 2017, 30-min steps, no time zone). | | `leak_demand_m3h` | double | Leak outflow at the leak node, m^3/h. | ## Evaluation and baselines `label` is the per-time-step ground truth (1 while any leak is active). Detectors are scored with the LeakDB metrics, computed over all time steps of all scenarios of a split: true-positive rate (RTP), true-negative rate (RTN), F1 and the early-detection score (SED), which rewards detecting each leak event soon after it starts (within the leak plus 5 h) and only counts detections that persist. `scripts/leakdb_score.py` is a Python port of the original MATLAB scoring function (identical results with `--matlab-compat`, checked in GNU Octave), and `scripts/leakdb_baselines.py` contains the paper's reference detectors. ```bash pip install pandas pyarrow huggingface_hub # predictions: Parquet/CSV with columns scenario, time_index, prediction (0/1), for every scenario of the split python scripts/leakdb_score.py --predictions my_preds.parquet --network hanoi --split test --data hf://datasets/KIOSCoE/LeakDB # reference baselines (MNF tuned on validation) for both networks python scripts/leakdb_baselines.py benchmark --data hf://datasets/KIOSCoE/LeakDB ``` Baseline results (`baselines/baseline_results.json`; MNF λ and w tuned for F1 on `validation`): **Net1_CMH** — test split (149 scenarios) and all 1000 scenarios; scores in %. | detector | RTP | RTN | F1 | SED | RTP (all) | RTN (all) | F1 (all) | SED (all) | |---|---|---|---|---|---|---|---|---| | Ideal | 100.00 | 100.00 | 100.00 | 97.73 | 100.00 | 100.00 | 100.00 | 98.15 | | Null | 0.00 | 100.00 | 0.00 | 0.00 | 0.00 | 100.00 | 0.00 | 0.00 | | Random | 50.07 | 49.99 | 30.79 | 0.00 | 50.02 | 50.00 | 30.31 | 0.10 | | MNF λ=0.10, w=10 (paper) | 84.67 | 40.90 | 43.23 | 69.58 | 86.50 | 42.61 | 44.01 | 69.55 | | MNF λ=0.20, w=10 (paper) | 53.95 | 98.67 | 68.03 | 44.86 | 58.76 | 99.01 | 72.40 | 46.61 | | MNF λ=0.30, w=10 (paper) | 51.83 | 99.87 | 68.07 | 44.84 | 56.81 | 99.90 | 72.28 | 46.10 | | MNF λ=0.25, w=7 (tuned on validation) | 52.49 | 99.68 | 68.33 | 45.02 | 57.33 | 99.82 | 72.57 | 46.57 | **Hanoi_CMH** — test split (150 scenarios) and all 1000 scenarios; scores in %. | detector | RTP | RTN | F1 | SED | RTP (all) | RTN (all) | F1 (all) | SED (all) | |---|---|---|---|---|---|---|---|---| | Ideal | 100.00 | 100.00 | 100.00 | 97.62 | 100.00 | 100.00 | 100.00 | 98.77 | | Null | 0.00 | 100.00 | 0.00 | 0.00 | 0.00 | 100.00 | 0.00 | 0.00 | | Random | 50.06 | 49.99 | 32.45 | 0.00 | 49.98 | 49.99 | 31.37 | 0.07 | | MNF λ=0.10, w=10 (paper) | 76.36 | 68.26 | 55.14 | 52.09 | 69.99 | 67.59 | 50.12 | 43.60 | | MNF λ=0.20, w=10 (paper) | 49.12 | 99.77 | 65.56 | 33.96 | 40.92 | 99.78 | 57.78 | 29.73 | | MNF λ=0.30, w=10 (paper) | 41.88 | 99.89 | 58.90 | 29.99 | 33.40 | 99.94 | 50.00 | 25.04 | | MNF λ=0.20, w=21 (tuned on validation) | 53.26 | 97.67 | 66.32 | 36.59 | 45.78 | 97.36 | 59.19 | 31.81 | The Minimum Night Flow (MNF) detector flags a day when its minimum inflow exceeds the lowest minimum of the previous `w` unflagged days by more than λ; inflow is `flow_link_1` for Hanoi and `-(demand_node_9 + demand_node_2)` for Net1. Like the reference code, it labels the whole day from 00:00, i.e. it looks ahead by up to one day. SED is normalised by the number of leak events, as in the MATLAB code; the ideal detector scores below 100 because it is silent during the 5-h extension window. `leakdb_score.py` also reports `SED_ideal_normalised` (paper eq. 6). ## Known issues in the source data (flagged, not altered) | network | scenarios | with leak label | with stale leak files | labels not explained by leak info | |---|---|---|---|---| | Net1_CMH | 1000 | 734 | 24 | 26 | | Hanoi_CMH | 1000 | 763 | 0 | 9 | - Synthetic data from two small benchmark networks; real sensor noise, missing data and sensor faults are not modelled. - Net1_CMH: 31 leak files in 24 scenarios are left over from an earlier generator run (flagged stale_file); the time-step labels do not include them. - Some scenarios had two leaks at the same node; only the later leak's info file survived, so labels contain events with no leak info row (labels_explained_by_leaks = false). - For abrupt leaks, outflow and labels start one step after start_index. - For incipient leaks the leak outflow series is intermittent and reaches zero at or before peak_index, while labels remain 1 until end_index. - Time stamps are a synthetic calendar year (2017) at 30-minute resolution. Use `stale_file == False` in `leaks` (and `leak_demands` joined on network/scenario/leak_node) to get the leaks that belong to each simulation; `qc_notes` in `scenarios` explains every flagged scenario. The time-step `label` is consistent with the valid leaks in all other cases. ## Provenance and processing Fully synthetic. Each scenario was generated by the LeakDB toolkit (Python, WNTR WNTRSimulator, pressure-driven demand mode) on the Net1 and Hanoi EPANET benchmark networks: pipe length, diameter and roughness were perturbed uniformly within a scenario-specific percentage, consumer demands were synthesised from weekly and yearly patterns with noise, and leaks were inserted as orifice emitters at randomly chosen nodes and times. - Original per-series CSV files (Index,Value) were merged into one wide table per network, one row per scenario and time step; values are copied unchanged as float64. - Labels were cast from 0.0/1.0 to int8; time indices were joined to Timestamps.csv. - Leak and scenario info files (key/value CSV) were parsed into typed tables. - Quality-control flags were derived (stale leak files, labels not explained by leak info files); no records were removed. ## Intended uses and limitations Training and benchmarking leak detection, localisation and anomaly detection algorithms on water distribution networks; scoring with the LeakDB metrics (F1, true-positive/negative rate, early detection score). Leak sizes, locations and timings are sampled from uniform distributions and do not reflect real failure statistics. ## Personal and sensitive information None. The data are simulated and contain no personal information. ## Authors and contact - Stelios G. Vrachimis ([ORCID 0000-0001-8862-5205](https://orcid.org/0000-0001-8862-5205)) - Marios S. Kyriakou ([ORCID 0000-0002-2324-8661](https://orcid.org/0000-0002-2324-8661)) - Demetrios G. Eliades ([ORCID 0000-0001-6184-6366](https://orcid.org/0000-0001-6184-6366)) - Marios M. Polycarpou ([ORCID 0000-0001-6495-9171](https://orcid.org/0000-0001-6495-9171)) KIOS Research and Innovation Center of Excellence, University of Cyprus, Nicosia, Cyprus. Questions and issues: Stelios G. Vrachimis (corresponding author), ; KIOS Research and Innovation Center of Excellence, . You can also open a discussion on this dataset's Community tab. ## Funding - European Union Horizon 2020 programme, grant agreement no. 739551 (KIOS CoE). - Interreg V-A Greece-Cyprus 2014-2020 programme, project SmartWater2020, co-financed by the European Regional Development Fund (ERDF) and national funds of Greece and Cyprus. ## Related resources - [DOI of this Hugging Face release](https://doi.org/10.57967/hf/11171) - [Zenodo archive of the original release (CC BY 4.0)](https://doi.org/10.5281/zenodo.13985057) - [Paper (WDSA/CCWI 2018)](https://doi.org/10.5281/zenodo.1313116) - [Generator, MATLAB scoring and detectors](https://github.com/KIOS-Research/LeakDB) ## License European Union Public Licence v1.2 (EUPL-1.2), as in the original release. ## Citation If you use LeakDB, please cite both this dataset release (DOI [10.57967/hf/11171](https://doi.org/10.57967/hf/11171)) and the paper that introduced the benchmark: ```bibtex @misc{kioscoe2026leakdb, author = {{KIOS Research and Innovation Center of Excellence, University of Cyprus} and Eliades, Demetrios and Vrachimis, Stelios and Kyriakou, Marios and Polycarpou, Marios}, title = {{LeakDB} (Revision 059b043)}, year = {2026}, url = {https://huggingface.co/datasets/KIOSCoE/LeakDB}, doi = {10.57967/hf/11171}, publisher = {Hugging Face} } @inproceedings{vrachimis2018leakdb, title = {{LeakDB}: a benchmark dataset for leakage diagnosis in water distribution networks}, author = {Vrachimis, Stelios G. and Kyriakou, Marios S. and Eliades, Demetrios G. and Polycarpou, Marios M.}, booktitle = {Proceedings of the 1st International WDSA/CCWI 2018 Joint Conference}, address = {Kingston, Ontario, Canada}, year = {2018}, doi = {10.5281/zenodo.1313116} } ```