SE Inspect 1.0.0-alpha.4 — tells foreign parts apart by appearance

2026-11-01 · fourth public alpha · licence PolyForm Noncommercial 1.0.0 (LICENSE, commercial use: COMMERCIAL.md)

This alpha is about telling parts apart honestly. A cell that only sees outlines will call anything SD-card-shaped an SD card; alpha.4 adds what it takes to say "unknown" instead. It also contains a fix for a regression we introduced and caught on our own bench — described below, because that is how this project works.

Measured on the bench

Pi 5, four AKD1500 M.2 modules, Camera Module 3, white paper, 48-scene test set (15 × SD card, 15 × Allen key, 8 × both, 5 × empty, 5 × dark plastic foreign parts shaped like an SD card). Confidence threshold 60 %.

	before (build 1.9.75, 4 Oct)	alpha.4 (build 1.9.81, 1 Nov)
Single parts named correctly	30 / 30	29 / 30
Empty table, nothing reported	5 / 5	5 / 5
Foreign parts reported as unknown	0 / 5	4 / 5
Two parts, both named, nothing extra	2 / 8	2 / 8
False-confident names (all scenes)	5 / 61	1 / 59

One test set, one bench, two object types — a measurement, not a benchmark. Builds 1.9.82–1.9.83 (also in this release) add diagnostics and a switch that is off by default; they do not change these numbers.

New: foreign parts with the same outline

• Appearance check: from the training photos the cell learns what each object looks like — brightness, texture, bright and dark areas, colour, edge density inside the part — and compares a named part with its nearest training photo. The limit comes from the training photos themselves (how far each photo is from its nearest neighbour). A part with the right outline but foreign appearance becomes unknown, with a readable reason ("appearance does not match colour, bright areas, texture"). On the bench: real cards 0.39–1.01, plastic chips 3.0–3.4, limit 2.66.
• Stray parts can win: parts taught as Background now win when they match clearly better (≥ 10 points) and are reported as stray part (verdict reject).

Fixed: a regression we caught ourselves

Builds 1.9.76–1.9.79 measured the appearance of training photos and of live parts with different masks. Real SD cards were rejected as "foreign", and the M3 detector was then allowed to rename the rejected part — as "Allen key". Single parts dropped from 30/30 to 17/30 on the bench. Fixed in 1.9.80: same mask on both sides, and a part rejected by appearance can no longer be renamed by the detector. A bench with real images from the Pi (tools/test_aussehen_echt.py) now guards this.

New: one part stays one part (partly)

• Confidently named parts are not split by the M3 detector any more unless the pieces are clearly stronger and say something else; detector boxes that cover a part already found are dropped as duplicates; small gaps from glare are bridged.
• Not solved: a chrome Allen key on white paper mirrors the paper and is partly invisible against the empty image (gap of about a quarter of the image width, see /api/diag_maske). No software bridging closes that. Use a mid-grey or matt black base for shiny metal parts. An edge-based switch (kanten) is included but off until measured.

New: see which scene fails, and why

• Individual scenes under the threshold curve: thumbnail, expected, found, status; filter "errors only"; view at the current threshold or the suggestion.
• Diagnostics: switches for appearance check, gap bridging and edge silhouette (/api/erkennung_schalter), test runs on a selection of scenes ({aktion: start, szenen: [ids]}, does not overwrite calibration), the decision trail per scene, and the last foreground mask as an image (/api/diag_maske).

Known limitations

• Shiny metal on white: see above. Parts that differ neither in outline nor in appearance cannot be told apart — SE Inspect does not claim otherwise.
• Two parts touching or close together: 2 of 8 scenes fully correct on the bench.
• Digital I/O, the Modbus I/O-module driver and the PLC interface remain untested on real I/O hardware. Modbus has no authentication: isolated machine networks only.
• Alpha: interfaces, file layouts and the vorsa_daten/ format may change between alphas without migration. Nothing in this software is a safety function (SAFETY.md).

Upgrading from alpha.3

Pull, then sudo systemctl restart vorsa. Appearance profiles are rebuilt automatically within about a minute (aussehen.json; the earlier format is discarded). Run the test set once afterwards.

Changes

CHANGELOG.md lists builds 1.9.75 → 1.0.0-alpha.4; each has a reproducible patch script in tools/. New benches: test_fusion, test_glanzluecke, test_aussehen, test_aussehen_echt (real images from the bench in beispielbilder/). Specification: SE_INSPECT_1_1_SPEC.md.

SE Inspect 1.0.0-alpha.3 — learns from operation, measured threshold, trend

2026-10-04 · third public alpha · licence PolyForm Noncommercial 1.0.0 (LICENSE, commercial use: COMMERCIAL.md)

This alpha is about the cell learning from its own operation and about measuring instead of guessing. Everything below was exercised on a Pi 5 with four AKD1500 M.2 modules against a real 48-scene test set (SD card, Allen key, foreign parts, empty table) — which is also where most of the fixes in this release come from.

New: the cell learns from operation (1.1-A)

• Parts that were unknown during inspection land in the inspection archive. Teach › Suggestions from operation groups them by shape and colour and shows each group with thumbnails: As new object, To object …, Background or Discard — then Train …. A slider sets the grouping finer or coarser.
• Also reachable from the "Re-teach recommended" card in Inspect.

New: threshold from the test set (1.1-B)

• Settings › Recognition › Threshold from the test set runs every test-set scene through recognition and shows, for thresholds 30–95 %, hits, foreign parts detected, known parts reported as unknown and the false-confident rate. The suggestion is the threshold with the most hits at ≤ 1 % false-confident; if no threshold achieves that, it is marked as a compromise with a note on what actually helps.
• Calibration per object: each object reaches a different confidence when recognised correctly (on the bench: SD card ~0.6, Allen key ~0.9). The test run measures this and rescales so that a typical hit sits at 90 % for every object — one threshold then fits all. Shown in the settings, discardable, marked as outdated after a new training.
• The test set now has a Foreign part chip (expected: unknown) and is visible in both profiles under Inspect › Details; "Open test set" jumps there.

New: statistics and trend (1.1-C)

• A persistent inspection journal (pruefjournal.jsonl) survives restarts. Inspect › Trend: today's numbers, stacked bars for the last 24 hours, a 14-day table, and a drift warning ("confidence dropping") when the mean confidence of the last 50 named parts falls ≥ 8 points below the first 50 after training.

Fixed

• M3 training looked stuck at "Isolating photos 0 %": it was isolating every photo in all 32 augmentation variants (3 872 passes on 687-px images, no progress). Rotation variants were useless for M3 (scene synthesis rotates anyway). Now original + mirror only, images capped at 384 px, progress per photo, and an overall progress bar across all phases with percent and colour (blue running, green done, red failed).
• Test run measured a different chain than live: test scenes were 687-px crops, so the empty image was rejected by a shape check and empty scenes produced phantom parts. The empty image is now cropped to the capture window and scaled to the scene. The test run also measures strictly raw (calibration locked) and refuses mode switches while it runs.
• Empty image vs. rotation: the empty image remembers the rotation it was taken at; if the image has been rotated since, it is not used and Setup/Inspect say so.
• Image rotation survives a restart (drehung.json); rotate button in the zoom bar (locked during Inspect).
• Two UI elements that jumped on every state refresh (Teach-zones button, First steps) are quiet now.

Known limitations

• Alpha: interfaces, file layouts and the vorsa_daten/ format may change between alphas without migration.
• Silhouette matching cannot separate foreign parts with the same outline as a taught object (bench: a dark plastic chip vs. an SD card). Teaching such parts as Background is the only remedy in this release.
• Digital I/O, the Modbus I/O-module driver and the PLC interface remain untested on real I/O hardware. Modbus has no authentication: isolated machine networks only.
• The M3 training chain (TensorFlow → cnn2snn → AKD1500) has not yet been run end-to-end on the bench since the speed-up; report what you see.
• Nothing in this software is a safety function (SAFETY.md).

Upgrading from alpha.2

Pull, then sudo systemctl restart vorsa. Learned objects and settings in vorsa_daten/ stay. New files: pruefjournal.jsonl, kalibrierung.json, drehung.json, vorschlaege_erledigt.json — all optional. install.sh --config "<link from spikingedge.com/configure>" writes the configuration files for a fresh cell.

Changes

CHANGELOG.md lists every build 1.9.63 → 1.0.0-alpha.3; each has a reproducible patch script in tools/patch_19xx_*.py. New benches: test_selbstlernen, test_trend, test_kalibrierung, test_m3_fortschritt, test_1_1_server; new UI smokes: smoke_elf, smoke_drehen, smoke_ruhig, smoke_laufbalken.

SE Inspect 1.0.0-alpha.2 — guided workflow and machine integration

2026-10-03 · second public alpha · licence PolyForm Noncommercial 1.0.0 (LICENSE, commercial use: COMMERCIAL.md)

Two things in this alpha: the operator interface was reorganised around the way people actually work, and SE Inspect can now be wired into a machine. Everything below runs in simulation without hardware and is benched in CI; the hardware paths for digital I/O and Modbus are untested on real I/O — reports welcome.

New: guided workflow

• The sidebar is the workflow: ① Setup → ② Teach → ③ Inspect, each step ticks itself off. A "next step" card leads on; First steps is a six-point assistant that opens once on a fresh cell.
• One primary button per page (INSPECT / Take photo / Record empty image), everything else in the "⋯" menu. Line and arm buttons only appear in line mode; a running line stays stoppable everywhere.
• One settings window (gear): General · Camera · Recognition · Jobs · I/O · PLC/Modbus · Users | Hardware · Diagnostics · Maintenance.
• Jobs (formerly "recipes") bundle the setup per product; the old "save setting" is part of it.
• Teach: objects as cards with photo progress and status light, "Background" instead of "nothing / foreign part", Train … opens the training dialog. Photo delete with Undo.
• Inspect: verdict card marked Live preview, counters below as Logged inspections. Keyboard help with "?".

New: fits into a machine

• Digital I/O (vorsa/eaio.py): input trigger, outputs ready / OK / NOK / fault, pulse or hold. Drivers: simulation (default), GPIO via gpiozero (optocoupler/relay modules), Modbus TCP I/O module. Without the hardware a driver falls back to simulation and says why. GPIO 18 (belt relay) and 0–3 are blocked. New inspection trigger External.
• PLC interface (vorsa/sps.py): SE Inspect as a Modbus TCP server — status bits, verdict, object, confidence, counters, job number, heartbeat, result sequence; commands inspect / reset counters / load job. Off by default, port 1502, optional read-only.
• User roles (vorsa/rollen.py): operator / setter / admin with PIN (hashed), enforced in the server, automatic sign-out, change log. Off until an admin PIN is set.

Configure your cell

spikingedge.com/configure turns six answers into a parts list with test status, rules of thumb for camera and light, a wiring sketch, guide chapters and the configuration files (start.local.sh, eaio.json, sps.json, pruef.json) for this release.

Known limitations

• Alpha: interfaces, file layouts and the vorsa_daten/ format may change between alphas without migration.
• Digital I/O, the Modbus I/O-module driver and the PLC interface are benched against a mocked gpiozero and an in-process Modbus peer — not yet on real I/O hardware. Modbus has no authentication: isolated machine networks only.
• Recognition rates on a real test set are not yet published (Q1–Q10 in SE_INSPECT_1_0_SPEC.md); the test-set tool is in the software.
• The trained SE-M3 detector is not included; live recognition runs geometrically (OpenCV) plus on-chip silhouette matching.
• Tested on Pi 5 8 GB with 1 and 4 cards and Camera Module 3. Nothing in this software is a safety function (SAFETY.md).

Upgrading from alpha.1

Pull, then sudo systemctl restart vorsa. Learned objects and settings in vorsa_daten/ stay. Roles, I/O and Modbus are off until you switch them on.

Changes

CHANGELOG.md lists every build 1.9.56 → 1.0.0-alpha.2; each has a reproducible patch script in tools/patch_19xx_*.py.

SE Inspect 1.0.0-alpha.1 — first public alpha

2026-10-02 · first public alpha · licence PolyForm Noncommercial 1.0.0 (LICENSE, commercial use: COMMERCIAL.md)

What this is

A neuromorphic inspection cell for one camera: Raspberry Pi 5, one to four BrainChip AKD1500 cards, Camera Module 3. It learns a part on the chip from a handful of photos, recognises it by aligned silhouettes, measures geometry on the Pi and says unknown instead of guessing. The web UI is bilingual (German/English) and runs without any hardware in synthetic mode.

What works in this alpha

• Stationary profile (default, VORSA_PROFIL=stationaer): inspection cell without belt or arm. Trigger by hand (button / space bar) or automatically when the scene settles; per-part inspection log; multi-frame averaging before segmentation.
• Line profile (VORSA_PROFIL=linie): belt via relay or L298N, Arduino stepper arm, delivery and pickup zones from one camera. A second camera, "view 2" fusion and per-camera learning zones are switched off in this release (VORSA_ZWEITKAMERA=1 brings them back, untested here).
• Recognition chain: empty-image difference → geometry → hypotheses for overlapping parts (two rounds) → card distributor across 1–4 AKD1500 → verdict with confidence bars. CPU stand-in (VORSA_CPU_LERNEN=1, automatic in synthetic mode) gives the same numbers in software, slower, with no hardware claim.
• Camera dock under the image (image, colour, window, background), right-click menu for overlays, capture-window resize in the learning view, three looks (SpikingEdge, industrial, Magna).
• Test-set tool: record scenes with expected results, tools/eval_testsatz.py scores a build against them.
• Benches without hardware (tools/test_*.py, tools/run_selftest.py) and UI smokes (tools/smoke_*.js, jsdom) — run in CI on every push.
• Installer install.sh (Pi 5, Bookworm 64-bit, kernel 6.12, patched Akida PCIe driver) and install.sh --laptop.

Known limitations (read before relying on it)

• Alpha: interfaces, file layouts and the vorsa_daten/ format may change between alphas without migration.
• The published quality numbers (Q1–Q10 in SE_INSPECT_1_0_SPEC.md) are not yet measured on a real bench test set; what exists are synthetic benches and lab observations. Evidence pages on spikingedge.com label every figure with its provenance.
• Hypotheses for ring-shaped fragments reach ~80 % on the synthetic bench (target 75 %); real overlapping parts are untested at scale.
• The trained SE-M3 detector is not included; live recognition runs geometrically (OpenCV) plus on-chip silhouette matching.
• One Pi, one camera, tested on Pi 5 8 GB with 1 and 4 cards. Pi 4 and other cameras are untested.
• Safety: a belt and an arm are machinery. SAFETY.md is mandatory reading; nothing in this software is a safety function.
• bereich_info can show a partially updated state for one frame at low frame rates (B9 in AUDIT_FREIGABE.md), cosmetic.

How to try it

See README.md — five minutes on a laptop without hardware, thirty minutes on a Pi 5 with one card. Website: https://spikingedge.com/get-vorsa/

Reporting

Bugs and bench results: GitHub issues. Attach install.log and the build string from the UI footer. Contributions: CONTRIBUTING.md (DCO + one-line CLA).

Changes since the lab prototype (VORSA INSPECT 1.9.37)

CHANGELOG.md lists every build 1.9.38 → 1.0.0-alpha.1; each one has a reproducible patch script in tools/patch_19xx_*.py.