Resources

Classifying human physical activities such as walking, running, sitting, cycling, and climbing stairs from wearable IMU data. Uses time-series classification on sliding windows of accelerometer and gyroscope readings. Core technology for fitness trackers, health monitoring wearables, and rehabilitation devices. Models are typically small CNNs or dense networks trained on labeled activity datasets like UCI HAR.

Detecting unusual patterns in gas sensor readings to identify air quality hazards, chemical leaks, or ventilation failures. Learns baseline environmental patterns from gas sensor arrays measuring VOCs, CO2, and particulates, then flags deviations using lightweight autoencoder models. The BME688 sensor with built-in AI capabilities is a common choice for edge deployment. Used in building management, industrial safety, and agricultural monitoring.

Detecting unusual patterns in sensor data using lightweight autoencoders or statistical models. Learns normal operating behavior and flags deviations. Extremely resource-efficient — runs on the most constrained MCUs. Applicable to vibration, current, temperature, or any repetitive signal.

Detecting infant crying from ambient audio for smart nursery monitors. Privacy-preserving since audio is classified on-device without cloud streaming — parents are alerted locally without recordings leaving the device. Distinguishes cry from other household sounds like TV, music, or pet noise using mel-spectrogram features and lightweight CNN classifiers. Low power consumption enables battery-operated monitor deployments.

Detecting and decoding barcodes or QR codes from camera images on-device. Modern ML-based approaches handle damaged, partially occluded, or angled codes that traditional decoders reject. Combines localization with decoding logic running entirely on the MCU. Used in warehouse inventory management, asset tracking, point-of-sale terminals, and healthcare specimen tracking where network-independent scanning is required.

Specialized predictive maintenance focused on detecting early-stage bearing failures in rotating machinery. Analyzes high-frequency vibration signatures from accelerometers to classify fault types including inner race defects, outer race defects, ball defects, and cage failures before catastrophic failure occurs. Models process spectral features extracted from vibration data windows. The CWRU Bearing Dataset is the standard benchmark for training and evaluating bearing fault classifiers.

Classifying plant leaf images to identify diseases, nutrient deficiencies, or pest damage. Deployed on battery-powered devices in agricultural fields where connectivity is limited. Models classify leaf images into healthy vs disease categories using quantized MobileNet or custom CNN architectures. Enables early intervention before diseases spread across crops. Training datasets like PlantVillage provide labeled images for common crops and disease types.

Detecting fall events from IMU accelerometer and gyroscope data in real time. Classifies motion patterns to distinguish genuine falls from normal activities like sitting down quickly, stumbling and recovering, or placing the device on a surface. Critical for elderly care wearables and workplace safety devices in construction or manufacturing. Models process short windows of 6-axis inertial data and trigger alerts within seconds of a fall event.

Classifying hand or body gestures from IMU (accelerometer + gyroscope) data. Models process short windows of motion data to recognize predefined gestures. Used for touchless control, wearable interfaces, and accessibility devices. Requires 6-axis or 9-axis inertial measurement unit.

Detecting the specific acoustic signature of breaking glass for security applications. Uses mel-spectrogram features fed into a small CNN to distinguish glass-break events from other loud sounds like slamming doors, dropping objects, or thunder. Extremely low resource requirements make this viable on the most constrained MCUs. Deployed in home security sensors, commercial alarm systems, and window monitors.

Classifying entire images into predefined categories without localization. The model outputs a single class label per frame — no bounding boxes or object positions. Uses quantized MobileNet, EfficientNet-Lite, or custom CNN architectures optimized for microcontrollers. Simpler than object detection with lower resource requirements. Common applications include quality inspection, scene recognition, and presence detection.

Detecting and reading license plates from camera images at the edge. Combines object detection for plate localization with character recognition for reading plate numbers. Requires higher resolution input and more compute than simple classification tasks. Deployed in parking management, toll collection, and access control systems where cloud connectivity is unreliable or latency-critical.

Classifying machinery sounds to detect abnormal operation such as bearing noise, belt squeal, pump cavitation, or motor imbalance without physical contact sensors. Uses audio spectral analysis instead of accelerometer-based vibration monitoring. Deployed as a non-invasive retrofit on existing equipment — mount a microphone near the machine, no mechanical attachment required. Particularly useful for legacy equipment where installing contact sensors is impractical.

Identifying and localizing objects in camera images on-device. Typical models include quantized MobileNet-SSD or YOLO-based architectures optimized for microcontrollers. Used for presence detection, people counting, quality inspection, and simple classification tasks.

Counting the number of people entering or occupying a space using on-device vision. Uses lightweight detection or classification models to track head and body presence without cloud connectivity. Applicable in retail footfall analytics, occupancy-based HVAC control, and building access management. Typically runs quantized MobileNet-SSD or FOMO (Faster Objects, More Objects) architectures optimized for low-memory MCUs.

Monitoring electrical current and voltage waveforms to detect anomalies in power distribution systems. Identifies power surges, harmonics, phase imbalances, and degradation patterns using lightweight time-series models on current sensor data sampled via ADC. Deployed on MCUs connected to current transformers or Hall-effect sensors on electrical panels. Enables early detection of electrical faults before they cause equipment damage or safety hazards.

Monitoring machinery vibration, temperature, and current patterns to detect anomalies before failures occur. Models analyze time-series sensor data to predict remaining useful life or flag abnormal behavior. Runs continuously on low-power MCUs attached to industrial equipment.

Detecting cavitation in pumps and hydraulic systems by analyzing vibration and acoustic patterns simultaneously. Cavitation — the formation and collapse of vapor bubbles in liquid — causes rapid erosion damage if undetected. Models classify vibration frequency patterns and acoustic signatures characteristic of cavitation onset, distinguishing them from normal flow noise. Dual-sensor approach using both accelerometer and microphone improves detection accuracy over single-sensor methods.

Classifying entire camera frames into scene categories such as indoor vs outdoor, day vs night, crowded vs empty, or room type without detecting individual objects. Used for context-aware IoT devices that adapt behavior based on their environment — adjusting lighting, HVAC, or security modes automatically. Lower compute than object detection since the model outputs a single class label per frame with no localization required.

Classifying environmental sounds into categories such as glass breaking, dog barking, machinery noise, sirens, or alarms. Uses mel-spectrogram feature extraction fed into lightweight CNN or dense networks. Distinct from voice recognition — no language model or speech-to-text involved, focuses on non-speech audio events and acoustic patterns.

Identifying who is speaking from voice characteristics without recognizing what is said. Extracts speaker embeddings from short audio segments and matches against enrolled voice profiles stored on-device. Used for personalization, multi-user device access, and voice-based authentication. Privacy-preserving since raw audio never leaves the device. Requires more memory than keyword spotting due to embedding model complexity.

Detecting abnormal vibration patterns in rotating machinery such as motors, pumps, fans, and compressors. Learns normal vibration signatures using autoencoders or statistical models and flags deviations indicating bearing wear, shaft imbalance, or misalignment. Runs on constrained MCUs attached directly to equipment housings. Extremely resource-efficient — the model processes accelerometer data windows and outputs a binary normal/anomaly classification.

Identifying manufacturing defects such as cracks, scratches, missing components, or surface irregularities on production lines using on-device image classification or object detection. Models are trained on good vs defective part images and deployed on MCUs with camera interfaces. Enables real-time quality inspection at the point of manufacture without network dependency or cloud latency.

Recognizing a vocabulary of 10 to 50 spoken commands on-device without cloud connectivity. Extends beyond single-keyword spotting to support a broader command set such as directional controls, numeric inputs, or appliance operations. Uses DS-CNN or convolutional architectures trained on speech command datasets. Requires more model capacity than wake-word detection but runs only after activation, not continuously.

On-device keyword spotting and voice command recognition without cloud connectivity. Typical models use DS-CNN or depthwise separable convolutions to classify short audio segments into predefined command categories. Privacy-preserving since audio never leaves the device.

Always-on keyword spotting that listens for a specific activation phrase. Ultra-low-power operation is critical since the model runs continuously waiting for the trigger word. Uses DS-CNN or depthwise separable convolutions on mel-spectrograms of 1-second audio windows. The model classifies short audio segments as either the target keyword or background noise. Common in smart speakers, voice-activated appliances, and hands-free industrial interfaces.

Detecting and classifying wildlife species from camera trap images on-device. Reduces data transmission by only sending frames containing animals of interest. Uses image classification or lightweight object detection optimized for battery-powered field deployment with solar charging. Models trained on species-specific datasets process images locally, enabling deployment in remote areas without cellular coverage.