--- license: mit tags: - telemetry - time-series - sls - 3d-printing - additive-manufacturing - inova-mk1 configs: - config_name: ticks data_files: - split: train path: data/ticks/*.parquet default: true - config_name: peregrine data_files: - split: train path: data/peregrine/*.parquet dataset_info: features: - name: frame_chamber dtype: image - name: frame_galvo dtype: image - name: frame_thermal dtype: image --- # Inova-Mk1-Telemetry Time-aligned printer-state recordings from Inova Mk1 SLS 3D print runs. One row per 10 Hz **tick** — the recorder's `/state/snapshot` poll — with the full sensor state snapshot (~64 columns: temperatures, position, power, lights) on every row, the nearest camera frame embedded inline when one fell within the prior 100 ms window, and any 1 kHz position-stream samples from that window collected as a nested list. 25 parquet files across builds spanning 2026-05 through 2026-07. Build metadata (job name, profile, start/end time) is denormalized into every row, so each file is self-sufficient for ML — no joins needed. ```python from datasets import load_dataset ds = load_dataset("ppak10/Inova-Mk1-Telemetry", split="train") row = ds[0] # row["frame_chamber"] → PIL.Image.Image (or None) # row["bedmatrix"] → {width, height, values (768 °C floats), path} (or None) # row["powderBed.temp.current"] → float (°C) # row["positions.position.z2"] → float (µm) # row["position_hf_burst"] → list of {ts_offset_ms, x, y, z1, z2, r, has_homed} ``` Full image bytes are embedded in each row — no separate frame download. Filter to rows with a particular frame kind: ```python with_chamber = ds.filter(lambda r: r["frame_chamber"] is not None) ``` ## `peregrine` config — defect-labeled layers A second, sparse config shaped like the [ORNL Peregrine dataset](https://huggingface.co/datasets/ppak10/Peregrine-Dataset-v2023-11): **one row per layer that carries a defect label**, for Peregrine-style transfer learning. Only labeled layers are recorded, so it's small and defect-focused (vs. the dense per-tick `ticks` config). ```python peregrine = load_dataset("ppak10/Agentic-SLS-Telemetry", "peregrine", split="train") row = peregrine[0] # row["image_after_powder"] → JPEG bytes, post-recoat chamber still (Peregrine after_powder) # row["image_after_melt"] → JPEG bytes, post-scan chamber still (Peregrine after_melt) # row["part_mask"] → PNG bytes, galvo scan mask (Peregrine part_ids) # row["labels"] → list of {class, bbox [x0,y0,x1,y1] normalized, polarity} # row["has_debris"], ... → bool per class (PEREGRINE_ALL_CLASSES) # row["printBed_temp"], row["laser_power_w"], ... → process scalars at the layer ``` Images are raw encoded bytes (`binary`, like Peregrine's `image_after_*`) — decode with `PIL.Image.open(io.BytesIO(...))`. Labels are bbox-level (from the live defect detector / operator verdicts), not per-pixel masks; `polarity` is `positive` (defect present) or `negative` (operator marked the alert a false positive). One parquet per build (`data/peregrine/{build:03d}.parquet`), so a new build never rewrites existing files. --- ## Timelapse previews Per-build layer-by-layer timelapse GIFs (one frame per detected print layer, 25 fps, ≤ 12 s). Below is the composite (chamber | thermal | galvo) for build 012, the first full Layers-phase run: ![Build 012 galvo timelapse — laser scan trace per layer](previews/012/timelapse_galvo.gif) Galvo (scan-mirror) trace for build 012 — Hex Coasters, 221 layers. Each frame is the last captured view of one print layer just before the next powder spread. Full composite (chamber | thermal | galvo) at `previews/012/timelapse_composite.gif`. Individual-kind GIFs (`timelapse_chamber.gif`, `timelapse_thermal.gif`) and full-speed MP4 previews are also available under `previews/{build_id:03d}/`. See the **Previews** section below. --- ## Row shape Each row is one moment in time (a single 10 Hz tick). ~80 columns: | Group | Columns | Notes | |---|---|---| | **Build context** | `build_id`, `job_name`, `started_at`, `ended_at`, `phase`, `print_profile_name`, `inova_session_id` | Denormalized — same value on every row in a given file. `inova_session_id` is null until the upstream sidecar CSV is curated by hand. | | **Timestamp** | `ts` (`timestamp[us, UTC]`) | `respondedAt` from `/state/snapshot`. | | **Position (10 Hz)** | `positions.position.{x,y,z1,z2,r}` | Stage position in **microns** (firmware native; divide by 1000 for mm). Always present. `z2` advances with each deposited layer and is used for layer detection in the timelapse scripts. | | **Lights** | `lights.lights.{enabled,count}` | | | **Power (W)** | `{laser,fanGalvo,buzzer,laserSafety,io-en,wd-en,wd-in}.power`, `powerman.power.{current,required,max}` | Per-component draw and manager state. | | **Temperature (°C)** | `{powderBed,printBed}.temp.{current,average,target}`, `{powderChamber1..4,printChamber1..4}.temp.{current,average,target}`, `{quadrant1..4,surface,surfaceAvg,surfaceMin,surfaceMax,testTemp1}.temp.{current,average}` | ~51 columns. | | **Frame images** | `frame_chamber`, `frame_galvo`, `frame_thermal` | HF `Image` feature (struct of `{bytes, path}`). Null when no frame of that kind was captured in the 100 ms window. Loads as `PIL.Image` via `datasets`. `path` preserves the original filename for traceability. **`frame_thermal` is a legacy pre-rendered IR heatmap GIF and is null for builds recorded after 2026-07-12 — use `bedmatrix` for thermal from there on** (see below). | | **Bed temperature matrix** | `bedmatrix` | Struct `{width: int32, height: int32, values: list, path: string}`. The raw IR bed-surface temperature grid (32×24 = 768 cells, row-major, °C) attached nearest-before-tick in the same 100 ms window; null when none landed. Reshape `values` to `height × width` and colormap to render a heatmap. This supersedes `frame_thermal` and is available on all builds (the stream has run since build 013). | | **Position burst** | `position_hf_burst` | `list>`. 1 kHz position events that fell in `(tick_ts − 100 ms, tick_ts]`. Empty list during heating/idle. | ### Null frames `frame_*` columns are null when no frame of that kind was captured within the 100 ms tick window — this preserves the "is this image fresh?" signal. Forward-fill for display or training: ```python import polars as pl df = pl.read_parquet("data/ticks/026.parquet") df = df.with_columns( pl.col("frame_chamber").forward_fill(), pl.col("frame_thermal").forward_fill(), pl.col("frame_galvo").forward_fill(), ) ``` --- ## Build inventory 25 builds total. Very short failed-heating runs (2–24 ticks) produce a parquet but represent only seconds of recording. The **timelapse** column shows the layer-by-layer composite GIF (chamber | thermal | galvo) for builds where the printer reached the sintering phase (z2 > 0). | build_id | timelapse | job_name | ticks | date | |---:|:---:|---|---:|---| | 001 | | Hex Coasters 2026_05_31 | 102,805 | 2026-05-31 | | 002 | | Hex Coasters 2026_05_31 | 68 | 2026-05-31 | | 012 | | Hex Coasters 2026_05_31 | 201,660 | 2026-05-31 | | 013 | | D790 and D638 and Benchy 2026_06_02 | 13,811 | 2026-06-02 | | 014 | | D790 and D638 and Benchy 2026_06_02 | 2 | 2026-06-02 | | 016 | | D790 and D638 and Benchy 2026_06_02 | 14 | 2026-06-02 | | 017 | | D790 and D638 and Benchy 2026_06_02 | 2 | 2026-06-02 | | 025 | | D790 and D638 and Benchy 2026_06_02 | 24 | 2026-06-02 | | 026 | | D790 and D638 and Benchy 2026_06_02 | 355,580 | 2026-06-02 | | 028 | | D790 and D638 2026_06_07 | 340,162 | 2026-06-07 | | 029 | | D790 and D638 and Cards 2026_06_09 | 285,092 | 2026-06-09 | | 030 | | D790 and D638 Recycled Powder 2026_06_24 | 228,723 | 2026-06-24 | | 031 | | D790 and D638 Recycled Powder 2026_06_25 | 161,379 | 2026-06-25 | | 032 | | Hex Coasters and Nameplates and Benchies | 19,954 | 2026-06-27 | | 033 | | Hex Coasters and Nameplates and Benchies | 1,559 | 2026-06-27 | | 034 | | Hex Coasters and Nameplates and Benchies | 18,251 | 2026-06-27 | | 035 | | D790 and D638 and other objects 2026_06_27 | 430,997 | 2026-06-27 | | 036 | | Hex Coasters and Nameplates and Benchies | 734,064 | 2026-06-28 | | 037 | | D790 and D638 Debug Run 2026_06_29 | 776 | 2026-06-29 | | 038 | | D790 and D638 Debug Run 2026_06_29 | 8,294 | 2026-06-29 | | 039 | | D790 and D638 and Nameplate 2026_06_29 | 334,454 | 2026-06-29 | | 040 | | D790 and D638 and D256 and Benchy 2026_07_01 | 446,764 | 2026-07-01 | | 041 | | Unknown 2026_07_09 ¹ | 134,790 | 2026-07-09 | | 042 | | Nameplates 2026_07_09 | 203,425 | 2026-07-10 | | 043 | | D790 and D638 and Nameplate 2026_07_10 | 373,863 | 2026-07-10 | | **total** | | | **4,900,289** | | ¹ Build 041 was recovered from the NVMe spool after the recorder's Postgres database was reset; `job_name` is a placeholder. --- ## Files ``` data/ticks/{build_id:03d}.parquet # one file per build, zero-padded (001–043) previews/{build_id:03d}/ chamber.mp4 # optical — real-time 10 fps, forward-filled thermal.mp4 # IR bed heatmap (from bedmatrix; inferno, fixed 20–200 °C) galvo.mp4 # scan-mirror trace composite.mp4 # 1×3 panel: chamber | thermal | galvo timelapse_chamber.gif # layer-by-layer, 25 fps, ≤ 12 s timelapse_thermal.gif timelapse_galvo.gif timelapse_composite.gif # 1×3 panel GIF ``` The HF glob `data/ticks/*.parquet` loads the full dataset across all builds. Builds that never started printing (z2 = 0 throughout) produce a timelapse GIF with zero or one frame. --- ## Previews ### Real-time MP4s Four MP4s per build under `previews/{build_id:03d}/`: `chamber.mp4`, `thermal.mp4`, `galvo.mp4`, and `composite.mp4`. Playback at 10 fps (the tick rate), so video duration equals build wall-clock time — build 026 → ~9h54m of video. Null frames are forward-filled for viewing continuity. ### Timelapse GIFs Four animated GIFs per build sampled one frame per detected print layer: - **Layer detection**: `positions.position.z2` quantized in 100 µm buckets. Only z2 > 0 rows are used, so pre-print heating is automatically excluded. - **Representative frame**: the *last* non-null frame within each z2 bucket — the most-recent view of the layer just before recoating begins. - **Playback**: 25 fps, capped at 300 frames (12 s max). Null-frame levels are forward-filled within the GIF. - **Canvas**: 240 px height (half the MP4 canvas) for web-friendly file sizes. Both previews are sourced from this dataset's own parquet (not the recorder's raw frame files), so they see only the per-tick-attached frame subset (~36% chamber, ~37% galvo attachment rate). The **thermal** panel is rendered from the numeric `bedmatrix` IR grid — inferno colormap over a fixed 20–200 °C range, so the same color means the same temperature across every build — rather than the legacy `frame_thermal` GIF. Builds 001/002/012 predate the bedmatrix stream and fall back to the old GIF. --- ## What's lost vs the raw exporter format This dataset is **tick-anchored**, not a strict outer join. Three things are reduced compared to the upstream flat exports: - **Between-tick position resolution**: the 1 kHz `position_hf` stream is preserved *during* motion via `position_hf_burst`, but not as standalone rows between ticks. For the raw 1 kHz timeline fall back to the upstream `position_hf/{build_id}.parquet`. - **Sparse events without telemetry**: no row exists for a camera frame or `events` record that doesn't land near a 10 Hz tick. This dataset represents "state at each tick", not "every recorded event". - **`plotter_commands`**: not surfaced — historical builds predate this recorder feature and the table is empty. --- ## Joining back to Inova-Mk1-Database Each row carries `print_profile_name` — the string the printer was running when the build started, taken from the `build_start` event payload. To recover the matching `PrintProfile.Id` UUID, join against [`ppak10/Inova-Mk1-Database`](https://huggingface.co/datasets/ppak10/Inova-Mk1-Database): ```python import json from pathlib import Path profiles_dir = Path("Inova-Mk1-Database/source/PrintProfiles") name_to_id = { json.load(p.open())["Name"]: json.load(p.open())["Id"] for p in profiles_dir.glob("*.json") } ds = ds.map(lambda r: {**r, "print_profile_id": name_to_id.get(r["print_profile_name"])}) ``` --- ## Upstream Raw data lives in [`Agentic-Additive-Manufacturing-Process-Optimization`](https://github.com/ppak10/Agentic-Additive-Manufacturing-Process-Optimization). Its `scripts/export.py` produces flat JSONL/parquet files under `data/exports/`. This dataset reads from those directly — no live database connection required. --- ## Regenerate ```sh # Full ETL (all builds) uv run scripts/ticks/01_extract.py # Specific builds only uv run scripts/ticks/01_extract.py 26 28 # MP4 previews uv run scripts/previews/01_render.py # all builds uv run scripts/previews/01_render.py 26 28 # specific builds # Timelapse GIFs uv run scripts/previews/02_timelapse.py # all builds uv run scripts/previews/02_timelapse.py 26 28 # specific builds ```