Sensor Signal Filtering

Design the signal chain from sensor output to reported value. Keep raw samples, filtered values, timing, saturation, and fault flags observable; a filter must not conceal a broken sensor or unsafe voltage.

Intake gate

Record or label unknowns before selecting a component or algorithm:

• exact board, MCU, core/framework, ADC channel, resolution, reference, input range, acquisition time, and toolchain versions
• sensor part/module, output type and drive impedance, supply, cable length, expected signal bandwidth, valid range, warm-up, and failure behavior
• sample period/jitter budget, acceptable latency, step-response requirement, threshold policy, calibration model, and available measurement tools

Resolve the board with board-support before board-specific ADC advice. For physical or multi-session work, load embedded-project-loop first and keep its measurement gate open.

Workflow

1. Capture raw samples at a stated, repeatable rate before filtering. Separate sensor dynamics from spikes, aliasing, quantization, ADC settling, rail/ ground noise, EMI, and calibration error.
2. Set the sampling rate from the signal bandwidth and latency budget. Check Nyquist and alias attenuation; do not use a software filter to recover information already aliased into the band.
3. Choose one filter with an explicit window/alpha/model and calculate its startup behavior, delay, memory, CPU cost, and effect on thresholds. Common starting choices are moving-average for bounded windows, median for isolated spikes, and EMA/IIR for low-cost smoothing; model-based filters require a stated model and noise assumptions.
4. Check the analog path: sensor loading, RC cutoff, ADC acquisition/settling, reference and ground, cable/noise coupling, input protection, voltage limits, decoupling, and motor/radio current return paths.
5. Implement raw-plus-filtered telemetry and fault/saturation indicators. Validate deterministic vectors on the host before target compilation.
6. Change one causal variable per hardware iteration. Measure the ADC pin and supply/ground with the same configuration used by the firmware.

Load on demand

• Read references/filter-selection.md for sampling math, algorithm tradeoffs, calibration order, and latency choices.
• Read references/analog-front-end.md for RC/input networks, ADC drive/settling, wiring, protection, and measurements.
• Read references/verification-and-tdd.md for host, simulation, build, target, and system gates.
• Use the diagnostic-only examples/adc-filter-pipeline.ino as a board-neutral fixed-rate EMA starting point; fill in board-specific ADC assumptions before converting counts to volts. Reusable generated filtering patterns remain owned by arduino-code-generator.
• Run the deterministic helper with python3 scripts/filter_benchmark.py --help or a captured vector. It is a design/TDD aid, not a replacement for target timing or electrical measurements.

The existing arduino-code-generator filtering pattern owns reusable code snippets. This skill owns signal-chain diagnosis, filter selection, hardware interaction, and evidence boundaries. Use circuit-debugger for fault isolation, wiring-safety-check for logic and rail limits, and sensor-calibration-workbench only after detection and signal integrity are established.

Output contract

Use the shared Arduino skill contract: state assumptions, required tools and versions, implementation steps, tests/evidence, known limitations, and recovery/security notes.

• Build proof: host tests and an exact-board compile, if run.
• Upload proof: an uploader result for a named port/image, if run.
• Hardware proof: measured ADC pin, rail/ground, waveform, or wiring result.
• System proof: observed sensor behavior under the real load/environment.
• Deployment proof: field rollout, rollback, and maintenance evidence.

Never promote a passing host test, HTTP response, serial banner, or CI job to hardware or system proof. If the board, sensor, circuit, or measurement is unknown, stop at a bounded plan and name the next measurement.