Time-Series & Signals

Why this matters at Archetype

The Archetype JD names this skill explicitly: "Hands-on experience with raw time-series sensor data and/or video," "Iterative data preprocessing cycle execution with baseline comparison," "Time-series and signal visualization." The work is taking raw sensor traces (often messy, noisy, asynchronous across sources) and producing features Newton can reason about.

Preprocessing & cleaning

Common defects in raw sensor data

• Missing timestamps: sensor dropped a packet.
• Duplicate timestamps: sensor sent the same reading twice.
• Clock drift: sensor's clock vs server clock differ by hundreds of ms.
• Saturation: values clipped at the sensor's max.
• Stuck values: same reading for many samples (sensor failure).
• Spikes: single-sample outliers from electrical noise.
• Drift: gradual sensor calibration loss.

Cleaning recipe

1. Deduplicate on timestamp + sensor_id.
2. Sort by timestamp.
3. Identify and fill gaps explicitly (with NaN or interpolation, with a flag).
4. Cap or remove outliers (winsorize, clip, or robust z-score).
5. Detect "stuck sensor" runs and flag them.
6. Resample to a uniform rate if downstream models require it.

import pandas as pd
import numpy as np

def clean_sensor(df: pd.DataFrame, value_col: str, ts_col: str) -> pd.DataFrame:
    # 1. dedupe + sort
    df = df.drop_duplicates([ts_col]).sort_values(ts_col).reset_index(drop=True)
    # 2. flag stuck runs
    df['stuck'] = (df[value_col].diff() == 0).rolling(window=10).sum() >= 9
    # 3. winsorize spikes (robust)
    q_low, q_high = df[value_col].quantile([0.001, 0.999])
    df[value_col] = df[value_col].clip(q_low, q_high)
    # 4. uniform resample to 1Hz, forward-fill short gaps
    df = df.set_index(ts_col).asfreq('1S').ffill(limit=3).reset_index()
    return df

Time-series features

For supervised modeling over sensor/time-series data, manual feature extraction is often more interpretable and reliable than end-to-end deep learning. The standard set:

• Aggregates over windows: mean, median, std, min, max, quantiles. Over fixed-length windows or trailing windows.
• Counts and rates: number of events above threshold per minute; rate of change.
• Lag features: value at t−1, t−5, t−60.
• Differences and ratios: short-window vs long-window mean ratio (a tactical version of trend detection).
• Cross-sensor: correlation between two sensors over a window, lag of one against another.
• Frequency-domain: dominant frequency, spectral energy in specific bands.
• Statistical: skewness, kurtosis, autocorrelation at specified lags.

tsfresh, sktime, and similar libraries auto-generate hundreds of features. Useful for baselines; the staff-bar move is then identifying which 10–20 features actually drive the signal and discarding the rest.

Alignment & resampling

When you have multiple sensors with different rates and clocks, alignment is the first hard problem:

• Pick a reference clock (one sensor or the server clock) and align everything to it.
• Resample to a common rate. Upsample with interpolation (linear, cubic, or "as-of merge"); downsample with aggregation (mean, max, last).
• Handle clock drift: if clocks drift, align by event landmarks (cross-correlate to find best offset) rather than naive timestamp.
• Watch interpolation choice: interpolating a categorical sensor reading is meaningless; use forward-fill or last-observation-carried-forward instead.

Filtering & denoising

Three families to know:

• Moving average (boxcar): simple, introduces lag. Median filter is more robust to outliers.
• Exponentially weighted moving average: weights recent samples more. Familiar from finance and IoT.
• Butterworth low-pass / high-pass / band-pass: frequency-domain filtering with sharper cutoffs. scipy.signal.butter + filtfilt is the standard.

from scipy.signal import butter, filtfilt

def low_pass(signal, fs, cutoff_hz, order=4):
    nyq = 0.5 * fs
    b, a = butter(order, cutoff_hz / nyq, btype='low')
    return filtfilt(b, a, signal)  # zero-phase

Use filtfilt (forward-backward filtering) to avoid introducing phase delay. The cost is non-causality — you're filtering with future samples. Fine for offline; for online prediction, use a regular causal filter.

Frequency-domain

For periodic or quasi-periodic signals, the frequency domain often surfaces structure that's invisible in time. scipy.signal.welch for power spectral density; scipy.signal.spectrogram for time-frequency views.

When this matters:

• Motor / equipment health (specific frequencies indicate specific wear).
• Activity recognition from accelerometer / gyroscope.
• Network traffic patterns.
• Anything with seasonality at multiple timescales.

Video data prep

For video, the equivalent of "clean the time series" is:

• Frame extraction at the right rate. Newton or downstream models may not use every frame; pick a sample rate that captures the dynamics without wasting compute (3–10 Hz is common for slowly changing scenes; 30 Hz for fast motion).
• Resolution / cropping. Most pretrained vision models expect specific input sizes. Crop to regions of interest if the customer cares about a specific zone.
• Annotation handling. Bounding boxes, segmentation masks, action labels — track formats and timestamps consistently. Encord and similar tools manage this.
• Synchronization with sensor streams. If video has a timestamp and sensors do too, align both to a single clock; verify with a known event visible in both streams.

Interview probes