edge-impulse-data-collection

name	edge-impulse-data-collection
description	Add a new sensor source (phone sensor, BLE peripheral, audio, image, location, etc.) to the Edge Impulse Data Collector Android app and wire it through to Edge Impulse ingestion. Use this skill whenever the user asks to "add a <sensor> source", "capture <signal> and upload to Edge Impulse", "wire up a new BLE characteristic", or similar in this codebase.
metadata	{"version":"1.0","author":"edgeimpulse"}

Adding a new data source to the EI Android Data Collector

This codebase already implements phone IMU/PPG, GPS, microphone, camera image, Wear OS relay, and Zephyr BLE central. New sources follow the same pattern.

Architecture (read this first)

• SensorCollector.kt — wraps the Android SensorManager and emits SensorData on a MutableSharedFlow. Other collectors (LocationCollector, AudioFileRecorder, ZephyrBLEClient, WearOSClient) follow the same collector-emits-flow shape.
• SensorViewModel.kt — single ViewModel that owns every collector, exposes collectSourceOptions: List<String> for the UI dropdown, and routes the selected option in startSensorForDuration(sensorType, durationMs).
• DataRepository.kt — the only place that talks to the EI ingestion API. Offline samples are written as CSV via saveSensorData and later flushed with uploadStoredCsvFiles. Images use uploadImage, audio uses uploadAudio. Do not add a new HTTP client.
• MainActivity.kt — Compose UI. The Collect screen reads viewModel.collectSourceOptions for its dropdown and calls viewModel.startSensorForDuration(...). Runtime permissions are declared in optionalPermissions().

Step-by-step

1. Build the collector. Either add a sensor type to SensorCollector.kt (preferred for anything exposed by SensorManager — barometer, magnetometer, gyroscope, etc.) or create a sibling class (MySourceCollector.kt) that exposes data: SharedFlow<SensorData> and start()/stop() methods. Match the shape of the existing collectors.

2. Expose it in the UI. Add a human-readable label to SensorViewModel.collectSourceOptions so it appears in the Collect-screen "Sensor" dropdown. Keep labels short — they're shown in a small chip.

3. Route it. In SensorViewModel.startSensorForDuration(sensorType, …), add a branch for the new option that starts the collector, sets the active label, and stops after durationMs. Use the same offline-vs-online split as the existing branches (write CSV via DataRepository.saveSensorData when offline, or stream via the existing ingestion path).

4. Upload. Reuse DataRepository:

   • Numeric samples → saveSensorData(...) + later uploadStoredCsvFiles().
   • Images → dataRepository.uploadImage(bytes, label) (suspend, returns Boolean).
   • Audio → dataRepository.uploadAudio(wavBytes, label). Do not introduce a new OkHttpClient or hand-rolled multipart code — the existing helpers already handle x-api-key, x-label, retries, and the ingestion-vs-remote-management split.

5. Permissions. Declare any new runtime permission in app/src/main/AndroidManifest.xml and add it to MainActivity.optionalPermissions() so it gets requested at launch. BLE permissions vary by Android version — copy the existing Build.VERSION.SDK_INT gating from the BLUETOOTH_* block.

6. Test. Add a unit test next to the existing ones, e.g. app/src/test/java/com/edgeimpulse/gattsensors/MySourceCollectorTest.kt. Follow SensorDataTest.kt / voice/VoiceCommandParserTest.kt as the pattern (JUnit 4, no Robolectric unless absolutely needed).

7. Wear OS / GATT relay (optional). If the source also runs on the watch or the EI-Monitor firmware, mirror the addition in wearosdatalogger/WearSensorBus.kt (watch) or ZephyrBLEClient.kt + GattProfile.kt (BLE). The phone aggregates both via GattServerManager — you usually do not need to touch that.

Verification

After your change, this must still hold:

• ./gradlew :app:compileDebugKotlin passes.
• ./gradlew :app:testDebugUnitTest passes.
• The new source appears in the Collect-screen "Sensor" dropdown and, when selected, produces either a new CSV under the app's offline directory (visible from the Datasets screen) or a successful ingestion upload (status row turns green with "Uploaded N samples").

Pitfalls

• Do not call ingestion HTTP from the ViewModel or UI — always go through DataRepository.
• Do not block the main thread in a collector — emit on a background dispatcher and let SensorViewModel collect on viewModelScope.
• SensorData.values is a FloatArray; preserve channel order across all samples in a session or Studio will reject the upload.
• For new BLE characteristics, add the UUID to GattProfile.kt so the Zephyr firmware repo (ei-zephyr-ble-gatt-client/) can stay in sync.