Dual-Feedback Adaptive Wavelet Packet Thresholding for Denoising IoT Gas Sensor Signals

To address the dual challenges of environmental temperature–humidity fluctuations and spurious impulse artifacts in real-time IoT monitoring, this paper proposes a novel digital sensor signal denoising method based on wavelet packet decomposition (WPD) with a dual-feedback adaptive threshold. A four-layer IoT monitoring architecture spanning the perception, control, transmission, and cloud service layers is constructed, with an STM32F103C8T6 microcontroller, an SGP30 multi-pixel metal oxide (MOX) gas sensor, and an ESP8266 WiFi module as the core components; the proposed denoising algorithm is executed entirely on the STM32 edge node at the perception layer. Real-time acquisition of volatile organic compound (VOC) concentration and environmental parameters was conducted across 20 independent experimental groups in a sealed, newly renovated laboratory, using a single SGP30 sensor in chronologically ordered sessions. Based on the non-stationary time–frequency characteristics of the SGP30 on-chip-processed output signals, the optimal four-level decomposition scale was determined through Parseval energy conservation analysis, and the optimal wavelet basis (sym5) was identified through fifth-order Fourier polynomial fitting. The core innovation lies in the dual-feedback adaptive threshold strategy, which integrates physical environmental compensation with statistical node energy weights through a multiplicative mechanism. An ablation study confirms that the multiplicative coupling outperforms the best single-feedback variant by 2.52 dB and the additive-fusion baseline by 1.97 dB. Using the ensemble median of 50 repeated measurements under stable conditions as the pseudo-reference signal, experiments on 66,888 samples demonstrate an SNR improvement of 9.86 dB, a normalized mean square error (NMSE) of 0.0009, and a weighted-average abnormal impulse suppression rate of 79.1%, with an embedded execution latency of only 1.4 ms per frame on the STM32F103 platform.