Read and interpret a saved recording

Muttakin Rahman · AD7606 / two EMG / ICM-42688-P · 3 October 2026

The integrated recorder saves interpretable data. Two captured Wokwi runs were independently checked against their original binary records, known inputs and exported CSV files. Every one of their 21,116 records passes CRC32 and matches its CSV row. This verifies storage and interpretation within the simulator; the recordings are synthetic.

Download the example recordings and CSVs · Independent audit results · Digital timing verification · Complete recorder

What is actually saved

Each example is 679,808 bytes: a 4,096-byte versioned metadata header followed by 10,558 fixed-size, 64-byte records. Each header and record has its own CRC32. The firmware's FatFS writes, syncs and closes the file on a PSRAM-backed substitute medium. No physical microSD card or SDMMC transfer is tested here.

The foot sensor starts earlier than the shank sensor. Its first estimated time is 875,001 µs; the shank starts at 930,002 µs. Eleven additional foot packets belong to startup. Both have 252 interrupt events within the ADC acquisition window in the complete-scene pin capture. The differing file counts are explained, rather than silently filled or truncated.

Open the files

Extract the example download. Each of examples/known-dc/ and examples/synthetic-waveform/ contains:

Open CSVs in Excel, MATLAB, Python or R. Read quality.json and metadata.json first. The convenient analysis tables retain fault flags and negative startup times; they do not repair or synchronize data.

The normal firmware header says synthetic=false, even when that image executes in Wokwi. The evidence exporter changes only synthetic and adds simulation in the analysis copy, recalculating the header CRC. It never edits a sample. The independently checked sidecars and manifests identify both files. Do not use the original header alone to decide whether data were measured on hardware.

EMG voltage and channel identity

ch0 means EMG1; ch1 means EMG2, as recorded in metadata. Both codes belong to the same AD7606 conversion. The driver reads all eight words, but the remaining channels are not exported as muscle measurements. The two unused slots in the common four-channel record are zero and the enabled mask is 3.

For the selected ±5 V range:

ADC input voltage (V) = signed ADC code × 5 / 32768
Nominal resolution = 0.000152587890625 V per code

The known-voltage example stores code 6554 → 1.000061 V on EMG1 and 13107 → 1.999969 V on EMG2. These round-trip correctly through the actual recorder and converter. The CSV is in volts at the ADC input, not input-referred muscle microvolts, force, activation percentage or normalized EMG.

The selected analog circuit has unity nominal scaling, so *_myoware_raw_v_estimate currently equals the ADC-input voltage. input_midpoint_mv=0 is the reconstruction offset. The unplugged-input 1.5 V bias is recorded separately; do not subtract it as a calibrated electrode baseline. ADC reference, analog gain/offset and MyoWare gain have not been measured.

Blank adc_status and adc_crc columns are intentional: the original AD7606 has neither a sensor-frame status word nor a sensor CRC. File CRC32 detects damaged stored records; it does not prove that an unprotected ADC serial word was electrically correct.

Motion axes and units

The metadata selects ±16 g acceleration and ±2000 degrees/s gyroscope ranges. Conversion preserves the sensor's FIFO axis order:

Acceleration (g) = signed count × 16 / 32768
Angular velocity (degrees/s) = signed count × 2000 / 32768

Both models produce az_raw=2048 → 1 g and zero angular velocity. Tiny x-axis markers distinguish foot from shank: one count and two counts respectively. They are deliberate virtual identities, not measured motion. Temperature remains a raw count in this export. Sensor axes are not yet a calibrated anatomical or earth reference frame; mounting orientation must be measured before gait analysis.

Which time should be used

EMG conversion_trigger_host_us uses the ESP32's monotonic microseconds since boot, associated with the CONVST software write. It is not UTC or a measured aperture time. For relative seconds, subtract the first EMG trigger and divide by 1,000,000. The example spans 1.257537 s between its first and last EMG triggers, with a timestamp-derived rate of 7,983.86 Hz, inside the declared 8 kHz ±1% virtual gate. This short run does not establish long-term timing accuracy.

IMU host_time_estimate_us is reconstructed from a FIFO timestamp and interrupt timing; it is explicitly an estimate. sensor_timestamp_raw is the original 16-bit counter. sensor_time_unwrapped_us retains continuous sensor order across the 20 foot and 19 shank counter wraps in this short example. The two independent counters do not share an absolute clock origin.

All 515 IMU records retain flag 2, meaning timing uncertainty. A FIFO read interval describes a burst: later packets may be produced after its read starts or after its first IRQ anchor. The recorded read end bounds all nominal packet estimates here. Do not substitute read_start_us or irq_anchor_us for every packet's sample time. Drift and filter delay have not been calibrated; precision cross-sensor alignment still requires physical validation.

The convenience motion tables use the same first-EMG origin. Early foot samples therefore have negative relative time. This is intentional; no synchronization, interpolation or gap filling is performed.

The saved waveform is visible

The second run feeds 8,000 precomputed ngspice samples into the behavioral ADC and repeats that one-second input sequence. Every saved EMG code matches the corresponding source sample, including the repetition. Maximum difference between the input voltage and decoded quantized voltage is 76.293 µV, below half the nominal ADC step. This is ideal quantization evidence; it is not measured ENOB, sensor noise or ADC accuracy. The substitute op-amp model and offline analog coupling limits remain in the stimulus description.

Saved data and live preview differ deliberately

In both examples the live archive omits EMG sequence 5330 and reports one device preview drop. Every received measurement record is byte-identical to the corresponding saved record. The saved file contains sequence 5330 and has no detected sample loss. Use the saved file as the primary research archive and retain preview gap reports when using live data.

The existing fault runs also make incomplete data visible:

A passing fault test means that the fault is reported correctly. It does not mean its recording is complete. research_ready remains false throughout these examples.

Reproduce the comparison

The independent audit uses Python's standard library and does not import the existing converter. From the extracted example download:

python3 verify_saved_data.py --examples examples --output checked-results.json
python3 analyze_example.py examples/synthetic-waveform/converted my-analysis

The audit checks all original record CRCs, exact sample-body preservation, enabled-channel mask, unused slots, input code identity, every CSV field, units, sensor counter wraps, uncertain timing flags, final counters and live-to-saved byte parity. The eleven damaged-data cases reject CRC damage, truncation, swapped channels, wrong units, hidden disabled columns, missing rows, altered IMU flags and counter mismatches. Source evidence remains unchanged.

For a new physical recording, use the package's existing converter and recording/recovery guide. This example auditor intentionally expects these two specific synthetic sessions; it is not a general claim that arbitrary future recordings are valid.

Physical SD behavior, actual reference/scaling, MyoWare waveform quality, timing alignment, RF performance and long recordings remain unmeasured. The format is interpretable; research conclusions still depend on measured calibration, experimental labels and hardware validation.