Solar Field Logger with Lossless Compression

Context

A remote site with no mains power and no network. The brief: record enough environmental data to characterise a microclimate across a full winter without visiting the site more than twice. Off-the-shelf loggers either assume mains power, assume connectivity, or store CSV at a density that fills flash in weeks.

Constraints

• One 8 MB flash die, no SD card. The card slot is the first thing to corrode.
• Six months unattended, including weeks without meaningful sun.
• No network. Data leaves by physical extraction over USB-C.
• Sensor calibration must be applied at capture time, not at export, so a corrupt export cannot silently change the meaning of past samples.

Architecture

• Sampling. The RTC wakes the core every 60 s. Duty cycle is ~210 ms: four I²C reads, calibration polynomial, frame append, sleep. Everything else is stop mode at 2 µA.
• Framing. Each record is [u32 millis][u8 channel_map][i16 × n][u16 crc]. A double-buffered header page tracks write cursor and rollover count, so a torn write loses one record, never the log.
• Compression. Channel values are delta-encoded against the previous sample with a 1–2 byte zig-zag varint. Temperature and humidity move slowly; most samples compress to 31 bytes from an uncompressed 40.
• Power. MPPT is a simple perturb-and-observe on the panel voltage with a hard floor at 3.2 V cell voltage. Below the floor the logger stops sampling and records the gap explicitly rather than pretending continuity.

Trade-offs

• No retransmission. The log is append-only and extraction is destructive (read-then-truncate). Simpler, but a failed extraction can lose data — hence the physical rule of two extractions per visit to different media.
• Calibration baked in. Changing a sensor means a new firmware revision and a header record marking the change. It makes retroactive recalibration impossible, which was the point.
• Stop mode over low-power run. Wake latency is 1.2 ms; the alternative would keep I²C peripherals alive and roughly double idle draw.

Failure modes handled

• Panel disconnect: header flags a POWER_GAP; sampling resumes on restore.
• Sensor NAK: three retries, then the channel is marked invalid in that frame rather than writing a fabricated value.
• Flash block wear: log rotates across blocks in a fixed ring; the first estimated wear-out is beyond 30 years at this record rate.

Status

Prototype deployed. One full winter cycle completed with 187 consecutive sunless days (the panel was snow-covered) and no data loss. Enclosure revision 2 is in progress: the current gasket admits condensation at the cable gland during freeze-thaw, which the next revision addresses with a desiccant chamber and a vertical gland entry.