⚡ electrical engineering

A Rule-Based IoT Wearable Prototype for Companion Animal Vocalization Monitoring: Design, Implementation, and Documented Limitations

This paper presents the design and evaluation of a rule-based IoT wearable prototype using an ESP32, MEMS microphone, and IMU to monitor companion animal behavioral states through real-time multimodal sensor fusion, while critically documenting the shift from semantic emotion recognition to feasible state estimation due to embedded computational constraints.

Original authors: Muhammad Ali Siddiqui

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Ali Siddiqui

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a pet collar that acts like a tiny, super-smart detective. Its job is to listen to your pet's voice and watch how they move, then tell you on your phone what your pet is doing right now. This is the story of a prototype device built by Muhammad Ali Siddiqui, designed to help people (especially those who might have trouble watching their pets closely all day) understand their furry friends better.

Here is how the paper breaks down this invention, using simple analogies:

The Detective's Toolkit (The Hardware)

Think of the device as a small backpack for a pet. Inside this backpack are three main tools:

The Ears (Microphone): A tiny digital microphone (INMP441) that listens to every sound the pet makes.
The Inner Ear (Motion Sensor): A motion sensor (MPU6050) that acts like a human's inner ear, feeling every shake, jump, and step the pet takes.
The Brain (Microcontroller): An ESP32 chip that acts as the brain. It's small, cheap, and powerful enough to do math on the fly without needing to call a giant computer in the cloud for help.

The Original Dream vs. The Reality Check

The Dream: The creator originally wanted to build a "Pet Translator." The idea was that if the dog barked a certain way, the device would say, "He is hungry," or "She is sad." It was like trying to translate a foreign language instantly.

The Reality Check: After building the prototype, the creator realized this dream was too big for the tiny brain inside the device.

The Problem: A loud bark could mean "I'm excited!" or "I'm scared!" or "I want a treat!" Just listening to how loud the bark is (volume) isn't enough to know what the pet is feeling. It's like trying to guess if a person is happy or angry just by how loud they are shouting; you can't tell the difference without seeing their face or knowing the situation.
The Pivot: The creator had to change the goal. Instead of a "Pet Translator" (which guesses emotions), the device became a "Pet Activity Monitor" (which guesses actions). It stopped trying to guess feelings and started focusing on what the pet is actually doing.

How It Works Now (The Rules)

Since the device is too small to use complex "Artificial Intelligence" (like a brain that learns from thousands of examples), it uses a simple "If-Then" rulebook, like a traffic light system.

Listening (The Sound Rules):
- If the sound is very quiet $\rightarrow$ The pet is Silent/Resting.
- If the sound is medium $\rightarrow$ The pet is Low Activity (maybe just looking around).
- If the sound is loud $\rightarrow$ The pet is High Activity (maybe barking or meowing loudly).
Watching (The Motion Rules):
- If the sensor feels no movement $\rightarrow$ Resting.
- If the sensor feels a steady, slow shake $\rightarrow$ Pacing (walking back and forth).
- If the sensor feels fast shaking $\rightarrow$ Running.
- If the sensor feels a sudden, sharp jolt $\rightarrow$ Jumping.
The Magic Mix (Fusion):
The real magic happens when the device combines the sound and the movement.
- Scenario A: Loud sound + Running = Excitement/Play.
- Scenario B: Loud sound + Standing still = Distress or Attention Seeking (The pet is yelling but not moving, which might mean they are stuck or upset).
- Scenario C: Quiet sound + Resting = Calm/Sleeping.

The Test Drive (Results)

The creator tested this device in a quiet room with a cat (since cats were the main test subject).

Motion: The device was very good at telling the difference between sleeping, walking, running, and jumping. It got it right 90% of the time.
Sound: The device was okay at telling the difference between "quiet" and "loud," but it sometimes got tricked by background noise (like a TV turning on). It got the sound right about 82.5% of the time.
Combined: When it used both sound and motion together, it got the final answer right 92.5% of the time. It was fast, taking only 1 to 2 seconds to tell the phone what was happening.

The "But..." (Limitations)

The paper is very honest about what this device cannot do yet:

It's not a mind reader: It cannot tell you if your pet is "sad" or "happy." It can only tell you if they are "moving fast" or "making noise."
It needs tuning: The rules (like how loud is "loud") were set by hand. If you put this on a different animal (like a big dog vs. a small cat), the rules might need to be changed manually.
It's sensitive to noise: If your house is very noisy, the device might think your pet is barking when they aren't.
It's a prototype: This is a working model, not a finished product you can buy today. It works well in a lab, but the creator admits it needs more testing in real, messy homes.

The Bottom Line

This paper is a story of a smart engineer who tried to build a "Pet Translator," realized the technology wasn't ready for that specific job, and instead built a very useful "Pet Activity Tracker" that combines sound and movement. It proves that you can put a smart sensor on a pet's collar to tell you what they are doing in real-time, but for now, it's better at describing actions than emotions.

Technical Summary: Rule-Based IoT Wearable Prototype for Companion Animal Vocalization Monitoring

Problem Statement
The paper addresses the challenge of developing a low-cost, lightweight, and resource-constrained wearable system for real-time monitoring of companion animal behavior. While existing research has successfully utilized accelerometers for activity classification and deep learning for animal vocalization analysis, there is a significant gap in integrated, multi-modal systems designed specifically for domestic pets that operate entirely on embedded hardware without cloud offloading.

A critical problem identified is the difficulty of performing semantic emotional inference (e.g., distinguishing hunger from distress) from pet vocalizations on microcontroller-class devices. The original design objective aimed to create a "pet language translator" mapping vocalizations to emotional states. However, the authors note that this goal faces fundamental barriers: the lack of standardized, annotated datasets for pet emotions, high intra-species variability in vocal patterns, environmental noise interference, and the computational limitations of low-power microcontrollers which preclude the use of complex deep learning models required for reliable audio semantic decoding.

Methodology
The proposed solution is a four-layer modular IoT framework centered on an ESP32-WROOM-32 microcontroller. The system integrates two primary sensing modalities:

Acoustic Sensing: An INMP441 digital MEMS microphone captures audio at 16 kHz. The system computes the average absolute amplitude ( $A_{avg}$ ) over a 256-sample buffer.
Motion Sensing: An MPU6050 6-axis IMU (accelerometer and gyroscope) samples at 100 Hz. The system calculates scalar magnitudes for acceleration and gyroscope data, along with statistical features such as standard deviation ( $\sigma_a$ ) and peak-to-peak range ( $P_{pp}$ ) over a 50-sample sliding window.

Classification and Fusion Logic
Due to the inability to run deep learning models on the target hardware, the authors implemented a rule-based classification framework:

Acoustic Rules: Hierarchical thresholds ( $T_1=1200, T_2=2500, T_3=4000$ ) map average amplitude to heuristic states (Silent, Low, Moderate, High Activity). The authors explicitly acknowledge these are intensity proxies, not semantic emotional labels.
Motion Rules: Decision trees based on $\sigma_a$ , $g_{std}$ , $P_{pp}$ , and minimum acceleration ( $a_{min}$ ) classify activities into Resting, Pacing, Running, and Jumping.
Multi-Modal Fusion: A deterministic lookup table combines acoustic and motion outputs to generate a composite behavioral state. For instance, "High Activity" (acoustic) combined with "Running" (motion) is interpreted as "Excitement," whereas "High Activity" with "Resting" is interpreted as "Distress" or "Attention Seeking."

The system transmits data via Bluetooth Classic (SPP) to a companion mobile application (built with MIT App Inventor) with an average end-to-end latency of 1–2 seconds.

Key Contributions

Prototype Development: Creation of a self-contained, low-cost wearable device integrating acoustic and inertial sensing on an ESP32 platform, capable of local processing and wireless transmission.
Rule-Based Multi-Modal Framework: Implementation of a lightweight classification system that fuses acoustic intensity and motion statistics to improve contextual robustness over single-sensor approaches.
Critical Design Evolution: A transparent documentation of the design pivot from an unfeasible "semantic emotion recognition" goal to a practical "behavioral state estimation" paradigm. This includes a failure analysis of why amplitude-based acoustic features cannot reliably decode emotional valence on embedded hardware.
Experimental Validation: A systematic three-phase evaluation (acoustic-only, motion-only, and combined) demonstrating the system's performance under controlled conditions.

Experimental Results
The prototype was evaluated in a controlled indoor environment using a domestic cat as the primary subject for calibration.

Motion Subsystem: Demonstrated high consistency (90% overall), correctly classifying Resting, Pacing, Running, and Jumping. It showed no false triggers during baseline resting trials.
Acoustic Subsystem: Achieved 82.5% consistency. While "Silent" and "High Activity" states were detected with high accuracy (95–100%), "Low" and "Moderate" activity classes suffered from borderline threshold issues. Crucially, the acoustic subsystem exhibited a 15% false trigger rate during baseline trials due to ambient noise, confirming the limitation of amplitude-only analysis.
Multi-Modal Fusion: The combined system achieved a 92.5% composite state accuracy. The fusion logic successfully reduced ambiguity by cross-referencing acoustic events with motion context.
Latency: The system maintained a response latency of approximately 1–2 seconds, suitable for real-time monitoring.

Significance and Claims
The paper claims significance primarily as a proof-of-concept at Technology Readiness Level (TRL) 4, demonstrating that real-time, embedded, multi-modal behavioral monitoring is feasible on resource-constrained hardware.

The authors modestly frame the system's utility not as a definitive "pet language translator," but as an assistive monitoring tool for elderly individuals and users with limited mobility. By shifting the focus from semantic emotion decoding to behavioral state estimation (e.g., detecting distress, excitement, or inactivity), the system aims to reduce the burden of constant human observation and enable timely caregiving responses.

The study emphasizes that while the current rule-based approach is functional, it is limited by manual threshold calibration and noise sensitivity. The authors posit that future iterations should incorporate machine learning (e.g., TensorFlow Lite Micro) and larger, annotated datasets to improve acoustic classification robustness and generalization across different animal breeds and environments. The paper concludes that the transition from an ambitious semantic goal to a practical behavioral estimation system provides valuable empirical guidance for the embedded systems community regarding the boundaries of feasibility in acoustic emotional inference.