Comparative analysis of the effectiveness of TinyML models for sensor data classification tasks on the ESP32 platform

Abstract

The article addresses the problem of efficient analysis of continuous streams of sensor data in embedded systems with limited hardware resources. Traditionally, the processing of large volumes of data obtained from inertial sensors, such as gyroscopes and accelerometers, has been performed using powerful cloud computing platforms. However, this approach has several disadvantages, including data transmission latency, dependence on network infrastructure, and increased energy consumption. In this context, the use of Edge Artificial Intelligence technologies, particularly the TinyML concept, is becoming increasingly relevant, as it enables machine learning inference to be executed directly on microcontrollers.

The aim of this study is to conduct a comparative analysis of the efficiency of different neural network architectures for the task of spatial sensor data classification under resource-constrained conditions. The experimental part of the research was implemented on the ESP32-C3 microcontroller, which represents a typical modern energy-efficient platform widely used in Internet of Things systems. Within the study, four machine learning models were developed and evaluated: Shallow Dense, Deep Dense with Dropout regularization, a one-dimensional convolutional neural network (1D CNN), and a hybrid architecture combining different layer types.

The results of hardware profiling demonstrated that the 1D CNN model provides the best balance between classification accuracy, inference speed, and memory usage. Due to the ability of convolutional filters to effectively extract spatial-temporal features from sensor data sequences, this architecture achieved the highest validation accuracy. At the same time, the hybrid model also demonstrated high performance; however, its practical use appeared less efficient due to increased computational complexity and longer inference time.

Special attention in the study was paid to the analysis of the impact of model quantization on classification accuracy. Experimental results showed that applying 8-bit quantization (Int8) leads to significant accuracy degradation because spectral features obtained through the Fast Fourier Transform contain important fractional components. Their forced compression into an integer representation distorts the feature space and consequently reduces classification performance.

Additional analysis of hardware resource consumption confirmed that the use of models with 32-bit floating-point precision (Float32) is feasible even for microcontrollers such as the ESP32-C3. The required RAM and Flash memory remain relatively small, allowing the system to be integrated into more complex embedded solutions. The overall system response time, including signal processing and neural network inference, is approximately 64 ms, which satisfies the requirements of real-time systems.

The obtained results confirm the effectiveness of TinyML technologies for developing autonomous intelligent devices capable of performing complex sensor data analysis without relying on cloud computing resources. The proposed approach can be applied in motion recognition systems, wearable devices, robotics platforms, and other Internet of Things applications.