Sensor Data Stream

Sensors generate data at high, often fixed frequencies, creating a continuous, potentially unbounded flow. A single accelerometer sampling at 100 Hz produces 100 data points per second. For a fleet of devices, this scales to massive volumes, making centralized collection and storage impractical. This characteristic is a primary driver for on-device processing and federated learning, where raw data never leaves the sensor.

• Example: An industrial vibration sensor sampling at 1 kHz generates 86.4 million readings per day.

Data points in a stream are not independent; they are ordered sequences where the value at time t is often statistically dependent on previous values. This temporal correlation is critical for detecting patterns like anomalies, periodic signals, or trends. Models for sensor streams, such as Recurrent Neural Networks (RNNs) or Temporal Convolutional Networks (TCNs), must explicitly account for this sequential nature.

• Key Implication: Random shuffling of data, common in classical ML, destroys essential temporal information and is not applicable during on-device training on live streams.

The statistical properties of a sensor data stream—such as its mean, variance, and underlying patterns—can change over time. This concept drift occurs due to sensor degradation, environmental changes (e.g., seasonal temperature shifts), or evolving usage patterns. For federated learning, this means the data distribution is non-IID (Non-Independent and Identically Distributed) both across clients and over time for a single client.

• System Requirement: Algorithms must support continuous learning and personalization to adapt models locally without catastrophic forgetting of global knowledge.

A single device may have multiple sensors (an inertial measurement unit combines accelerometer, gyroscope, magnetometer), creating a multi-variate time series. While the raw dimensionality can be high, the information relevant to a specific task is often sparse within the signal. Effective on-device preprocessing (e.g., applying a Fast Fourier Transform to vibration data) is essential to extract compact, informative features, reducing the computational load for the subsequent model.

• Design Goal: Minimize the memory footprint and compute constraint by extracting only task-relevant features from the high-dimensional stream.

Sensor readings are inherently noisy due to electrical interference, quantization error, and physical sensor limitations. Streams may also contain missing values (e.g., due to transmission glitches) and outliers. Models deployed on the edge must be robust to these imperfections without relying on extensive cloud-based cleaning pipelines. Data augmentation techniques that simulate noise are often used during training to improve model resilience.

• Challenge: Noise characteristics can vary significantly between device hardware batches, contributing to heterogeneous client data distributions in a federation.

The value of sensor data is often highest immediately after it is generated, especially for control systems (e.g., autonomous robots) or safety alerts (e.g., fall detection). This demands real-time processing with strict latency bounds, often in the millisecond range. Transmitting raw streams to the cloud for inference introduces unacceptable delay, making on-device inference a necessity. This real-time nature also constrains the time available for on-device training within an availability window.

• Metric: End-to-end latency from sensor measurement to model prediction is a critical benchmark for TinyML systems.

An IMU combines an accelerometer, gyroscope, and often a magnetometer to provide a continuous stream of motion and orientation data.

• Key Metrics: Linear acceleration (g), angular velocity (deg/s), magnetic field strength (µT).
• Sample Rate: Typically 50-200 Hz for human activity.
• TinyML Applications: Human Activity Recognition (HAR), fall detection, gesture control, industrial equipment vibration monitoring.
• Data Characteristics: Multivariate time-series, often requiring filtering (e.g., for gravity removal) and windowing before feature extraction.

A MEMS microphone generates a stream of PCM (Pulse-Code Modulation) audio samples, representing sound pressure levels over time.

• Key Metric: Audio amplitude, typically as 16-bit integers.
• Sample Rate: 8 kHz (telephony) to 44.1 kHz (full audio), with 16 kHz common for keyword spotting.
• TinyML Applications: Keyword spotting ("Hey Google"), audio scene classification (e.g., glass breaking, baby crying), voice command recognition.
• Data Characteristics: High-frequency 1D signal. Requires spectral feature extraction (e.g., Mel-Frequency Cepstral Coefficients - MFCCs) or raw waveform processing with 1D CNNs.

This category includes sensors that stream measurements of ambient physical conditions.

• Temperature & Humidity: Provides a stream of °C/°F and %RH values for climate monitoring and predictive maintenance.
• Air Quality/Gas Sensors: Streams concentrations of specific gases (e.g., CO2 in ppm, VOC levels) for indoor air quality and safety monitoring.
• Barometric Pressure: A stream of pressure readings (hPa) used for weather forecasting and altitude tracking.
• TinyML Applications: Predictive maintenance (correlating temperature spikes with failure), smart agriculture, leak detection, occupancy sensing.

Ranging from simple light sensors to low-resolution image arrays, these generate streams of photometric data.

• Ambient Light Sensor: A simple stream of lux values for adaptive display brightness.
• Proximity/Infrared (IR): A stream of distance or reflectance values for presence detection.
• Low-Resolution Camera: Streams of small grayscale or RGB image frames (e.g., 96x96 pixels) from a CMOS sensor.
• TinyML Applications: Simple visual wake words, presence detection, rudimentary object classification (e.g., product sorting), anomaly detection in industrial settings.

Sensors that generate streams of physiological data from the human body.

• Photoplethysmogram (PPG): A stream of optical measurements from a wrist or finger used to derive heart rate and heart rate variability (HRV).
• Electrocardiogram (ECG): A stream of microvolt-level electrical potentials from the heart, used for arrhythmia detection.
• Electrodermal Activity (EDA): A stream of skin conductance measurements related to stress or arousal.
• TinyML Applications: Continuous heart rate monitoring, atrial fibrillation detection, sleep stage classification, stress level estimation.

Sensors deployed in industrial IoT, automotive, and agricultural settings with specific measurement targets.

• Current Sensors: Streams of current draw (mA/A) for Non-Intrusive Load Monitoring (NILM) to identify appliance usage.
• Ultrasonic Sensors: Streams of time-of-flight measurements for precise distance or liquid level monitoring.
• Capacitive Sensors: Streams of capacitance changes for touch, proximity, or material sensing.
• TinyML Applications: Predictive maintenance (motor current analysis), precision agriculture (soil moisture), smart inventory (weight/presence), flow rate monitoring.

The on-device dataset is the collection of raw sensor readings and derived features stored locally on an edge device. This is the private, non-exportable data pool used for local model training in a federated learning round.

• Formation: Built incrementally from the sensor data stream.
• Role in FL: Serves as the client's local training set; the model learns from this data without the data ever leaving the device.
• Key Challenge: Often non-IID (not independently and identically distributed) compared to other clients' data, which is a core challenge in federated optimization.

On-device preprocessing refers to the real-time algorithms that transform raw sensor data streams into a format suitable for model input, executed directly on the microcontroller.

• Purpose: Reduces data volume, extracts relevant features, and manages compute constraints.
• Common Techniques: Includes filtering, normalization, windowing, and computing spectral features (e.g., FFT) from time-series data.
• Impact: Critical for meeting the memory footprint and latency requirements of TinyML by reducing the dimensionality of the input fed to the neural network.

A resource-constrained device is an embedded system, typically a microcontroller unit (MCU), with severe limitations that define the TinyML operating environment. The sensor data stream originates on and is processed by these devices.

• Primary Constraints:
  • Memory Footprint (KB of RAM/Flash)
  • Compute Constraint (MHz clock speed, lack of FPU)
  • Energy Budget (milliwatt-hour capacity)
• Implication: Dictates the complexity of models that can process the data stream and perform on-device training.

An embedded Federated Learning (FL) runtime is the lightweight software library on a microcontroller that orchestrates the local client's role in the federated process, managing the lifecycle of the model trained on the sensor data stream.

• Core Functions:
  • Downloads the global model.
  • Executes local training using the on-device dataset.
  • Applies compression techniques (e.g., sparse updates, quantization) to the model update.
  • Securely transmits the update to the aggregator.
• Requirement: Must operate within the memory footprint and compute constraints of a resource-constrained device.

A sparse update is a communication-efficient technique in federated learning where a client transmits only the most significant changes (a small subset of non-zero values) to the model parameters after local training on its sensor data stream.

• Mechanism: Uses magnitude-based or random masking to create a sparse gradient or weight delta tensor.
• Benefit for TinyML: Drastically reduces the communication cost, which is a major contributor to battery drain on wireless IoT devices.
• Trade-off: Requires careful algorithm design to ensure global model convergence is not impaired.

The availability window is the limited, often intermittent, period when a federated learning client is able to participate in training. This is dictated by the device's duty cycle, power state, and connectivity—all factors influenced by continuous sensor data stream acquisition.

• Causes: Device sleeping to conserve energy budget, lack of network coverage, or being busy with primary sensing tasks.
• System Impact: Leads to partial participation in each federated round and can exacerbate the straggler problem.
• Design Consideration: Federated learning orchestrators must schedule training rounds and perform client selection around these unpredictable windows.