Sliding Window

The sliding window operates with a predefined, constant length (e.g., 256 samples, 1 second of audio). This constraint is critical for deterministic memory allocation on microcontrollers, as it bounds the maximum working memory required for feature extraction or inference. The size is a key hyperparameter, balancing temporal resolution (smaller windows) with feature stability (larger windows).

The window advances over the data stream by a hop size or stride. A hop size smaller than the window length creates overlap between consecutive windows, ensuring no transient events are missed at the window boundaries. This is essential for continuous monitoring. A hop equal to the window size provides discrete, non-overlapping segments. The overlap percentage directly impacts computational load and temporal sensitivity.

The algorithm processes data incrementally as new samples arrive, maintaining a first-in, first-out (FIFO) buffer. This enables low-latency analysis suitable for real-time control systems, such as detecting anomalies in vibration data or classifying audio keywords. It avoids the need to store the entire historical data stream, which is infeasible on devices with kilobytes of RAM.

By design, the window isolates the most recent segment of the time-series. This focuses analysis on current conditions, which is vital for tasks like:

• Activity Recognition: Classifying the current motion (walking, running).
• Peak Detection: Identifying events like heartbeats or impacts.
• Trend Analysis: Computing a moving average to smooth sensor noise. The window acts as a short-term, contextual memory for the system.

Sliding windows enable efficient algorithms through incremental computation. Instead of recalculating features from scratch for each new window position, results can be updated using the exiting and entering data points. For example:

• A moving average can be updated by subtracting the oldest value and adding the newest.
• FFT-based features can use the Overlap-Add or Overlap-Save methods. This reduces CPU cycles and power consumption on microcontrollers.

The window provides the isolated data block upon which feature extraction algorithms operate. Common operations performed on each window include:

• Spectral Analysis: Applying an FFT to compute frequency-domain features.
• Statistical Moments: Calculating mean, variance, skewness.
• Domain-Specific Features: Extracting MFCCs for audio or RMS for vibration. These extracted features form the input vector for a downstream TinyML model.

The Short-Time Fourier Transform (STFT) is a Fourier-related transform used to determine the sinusoidal frequency and phase content of local sections of a signal as it changes over time. It is the mathematical foundation for applying a sliding window to frequency analysis.

• Core Mechanism: It multiplies the signal by a window function (e.g., Hann, Hamming) that is zero-valued outside of a short interval, then computes the Fourier transform of that windowed segment. The window is then slid along the signal.
• Output: Produces a spectrogram, a time-frequency representation where one axis is time, the other is frequency, and the color/intensity represents magnitude.
• TinyML Application: Critical for audio and vibration analysis on edge devices, enabling features like keyword spotting or bearing fault detection by analyzing how frequency components evolve.

A moving average filter is a simple, ubiquitous application of a sliding window for temporal smoothing. It is a specific type of Finite Impulse Response (FIR) filter.

• Operation: Replaces each data point with the mean (simple moving average) or a weighted mean of the neighboring points within the sliding window. This attenuates high-frequency noise.
• Types: Simple Moving Average (SMA) uses equal weights. Exponential Moving Average (EMA) uses exponentially decreasing weights, approximating an infinite window with recursive computation, saving memory.
• TinyML Consideration: Extremely lightweight and common as a pre-processing step for sensor data (e.g., smoothing accelerometer readings). The window size directly trades off smoothing lag versus noise reduction.

Feature extraction is the process of transforming raw data within a sliding window into a condensed set of informative, non-redundant values suitable for machine learning models.

• Role of the Window: The sliding window defines the temporal context for feature calculation. Features are computed per window to create a stream of feature vectors for classification or regression.
• Common Window-Based Features:
  • Statistical: Mean, variance, standard deviation, min, max, percentiles.
  • Temporal: Zero-Crossing Rate (ZCR), Root Mean Square (RMS).
  • Spectral: Spectral centroid, bandwidth, flux (derived from an STFT applied to the window).
• TinyML Impact: Efficient feature extraction within the window is often more compute-friendly on microcontrollers than running raw data through a large neural network.

Activity recognition is a primary use case for sliding windows in embedded systems, classifying physical actions (e.g., walking, sitting, cycling) from sensor streams.

• Pipeline: A sliding window (typically 1-5 seconds) moves over raw inertial data (accelerometer, gyroscope). Features are extracted from each window. A compact classifier (e.g., decision tree, tiny neural network) makes a prediction per window.
• Challenges: Window size is critical—too short lacks context, too long introduces latency and may blend activities. Overlap between consecutive windows smooths predictions but increases compute.
• Deployment: The entire pipeline—windowing, feature extraction, inference—must run in real-time on the microcontroller, often with strict power budgets.

Anomaly detection uses sliding windows to model normal system behavior and flag deviations in real-time, crucial for predictive maintenance and security.

• Approach: A model is trained on features from windows of "normal" sensor operation. During inference, features from the current sliding window are compared to this baseline.
• Techniques: Can use statistical thresholds (e.g., if window variance exceeds a limit), reconstruction error from a tiny autoencoder, or distance from a cluster centroid.
• Edge Advantage: Detecting anomalies on-device eliminates cloud latency, enabling immediate local alerts (e.g., for machinery vibration faults) even without connectivity.

Peak detection algorithms often employ a sliding window as a local context to identify significant maxima (peaks) or minima in a signal.

• Sliding Window Logic: The algorithm examines points within a window to determine if the central point is a local maximum relative to its neighbors, adjusting for noise with thresholds.
• Applications: Detecting heartbeats in PPG signals, identifying impacts in vibration monitoring, or finding vocal onsets in audio.
• Optimization for MCUs: Algorithms like Savitzky-Golay filtering (which fits a polynomial within the sliding window) can smooth and find peaks simultaneously, but require careful implementation to manage computation on constrained hardware.