AI-enhanced flexible wearable sensor system using a 1D convolutional neural network for real-time environmental heat-risk classification

Abstract

Wearable sensing has become central to continuous health and safety monitoring, yet most systems either stream raw data to an external server or focus on a single sensing modality. This work reports the design, fabrication, and evaluation of a flexible, wearable sensor system that performs environmental heat-risk classification directly on the device. The platform integrates a BME280 environmental sensor and an ADXL345 triaxial accelerometer with an ESP32-S3 microcontroller on a flexible printed circuit board, sampling ambient temperature, relative humidity, barometric pressure, and triaxial acceleration at 2 Hz during deployment. Each 16-sample window is labeled by combining a temperature and humidity heat index with an activity intensity derived from the variability of the acceleration magnitude across the window, yielding Stable, Caution, and Danger categories. A lightweight one-dimensional convolutional neural network (1D CNN) with 6291 parameters classifies these windows; after INT8 quantization and TensorFlow Lite conversion, it fits the 80 KB tensor arena allocated on the microcontroller. Across ten independent runs using block-wise partitioning, the network achieved 93.1 ± 4.6% accuracy, outperforming a Random Forest baseline (84.9 ± 5.4%, corrected resampled paired t-test, p = 0.001) and matching a larger conventional CNN with 23 843 parameters (90.2 ± 7.1%, p = 0.36) at a quarter of the parameter count. Pooling the ten held-out partitions yields 92.9% accuracy across 1807 predictions drawn from approximately 1000 distinct windows, with no window misclassified between the Stable and Danger categories in either direction. A feature ablation shows that the motion channels contribute 12.4 percentage points, whereas adding barometric pressure reduces accuracy by 3.2 percentage points, indicating redundancy with the other environmental channels. The platform senses only environmental and motion variables, so its output is an environmental risk category rather than a physiological measurement. All reported metrics were collected offline; on-device operation was verified functionally but not benchmarked, so the results establish the engineering feasibility of embedded multimodal edge inference for workplace environmental monitoring rather than validating real-time performance.