LDAC-Net, a machine-learning model for low-cost MOX sensors, recorded 68.2% Top-1 accuracy on SmellNet-Base, a 50-way task, and the paper reports a 14.4-percentage-point advantage over the strongest baseline using fixed differencing. It also recorded 89.4% Top-5 accuracy and a macro-F1 of 65.2.
The result comes from a methods study that treats low-cost MOX recognition as multivariate time-series classification from raw windows. In plain terms, it assigns labels to readings from several sensor channels collected over time. The study asks whether the model can learn a drift adjustment without external drift correction or handcrafted temporal differencing, then tests it on a mixture task and a different drifting sensor array.
Inside the model
LDAC-Net learns drift compensation and temporal dynamics end to end through an LDFE front end. Its output then passes through attention-convolution blocks and an attention-pooling head before reaching a linear classifier.
Inside that front end, WSAN uses each window's per-channel mean and standard deviation to emit affine offsets. LMLD uses lags of 5, 10, 25 and 50 samples. The design also includes learnable per-channel scaling, identity concatenation, 1-by-1 mixing, normalization and squeeze-excitation gating.
What the main benchmark recorded
SmellNet-Base contains 50 base substances measured with six MOX channels, with six 10-minute sessions for each substance and about 180,000 time steps in total. The protocol used six-fold session-level cross-validation. At the default 300-sample window, each fold contained about 1,650 training windows and 330 test windows.
On SmellNet-Base, the reported figures were 68.2 ± 1.6% Top-1, 89.4 ± 0.3% Top-5 and 65.2 ± 4.5 macro-F1. The main configurations were summarized as mean ± standard deviation over five seeds, and confidence intervals were not reported.
In a separate window-length comparison, the tested LDAC-Net curve scored 57.8% at 100 samples, 68.2% at 300 and 64.8% at 400. The paper selected the analysis window through an offline grid search, so the results do not establish that 300 samples is best for every substance, channel or response speed.
Testing the model's parts
Component tests at 300 samples reported lower Top-1 scores for each tested variant than for the full configuration. The full model scored 68.2 ± 1.6%; versions without WSAN, LMLD, squeeze-excitation or channel mixing scored 54.2 ± 0.9%, 62.3 ± 1.2%, 61.9 ± 1.3% and 63.2 ± 0.4%, respectively. A version using only the 25-sample lag scored 63.6 ± 2.5%.
All reported drops met the paper's p < 0.05 threshold in two-sided paired t-tests across five shared seeds, although exact p-values were not reported. A separate lag-bank sweep compared the single 25-sample lag with the adopted 5, 10, 25 and 50-sample bank: the reported Top-1 scores were 63.6 ± 2.5% and 68.2 ± 1.6%, respectively.
The same model on other tasks
Beyond the classification benchmark, SmellNet-Mixtures asked the model to predict ratios for 12 base odorants from four channels, with a 12-component target. LDAC-Net recorded 50.5 ± 2.0% Top-1@0.1, 0.050 ± 0.001 mean absolute error and 81.1 ± 0.2% Top-K. The reported Transformer comparison was 45.4 ± 3.1%, 0.054 ± 0.002 and 78.3 ± 0.3%, respectively.
The cross-array test used eNose-Drift, a 62-channel array measuring three analytes, with about 700 recordings across 40 measurement days spanning 12 months. The authors trained on the earliest days and tested on the latest days. They used raw drift-affected log-resistance after initial calibration, 400-sample windows and LMLD lags of 25, 50, 100 and 150 samples.
On that chronological split, LDAC-Net reached 70.6 ± 4.8% Acc@1 and 69.6 ± 3.9% macro-F1. The paper reports advantages of 11.4 and 11.5 percentage points over the strongest raw baseline on those two measures. Against the strongest dataset-retuned fixed-differencing baseline, the reported margins were 8.0 points in Acc@1 and 3.0 in macro-F1.
What remains untested
The findings are computational benchmark results from sensor recordings, not human or clinical outcomes. The evaluation covered SmellNet-Base, SmellNet-Mixtures and one additional eNose-Drift array under controlled offline recording conditions. It does not establish performance in uncontrolled settings or beyond the tested datasets and hardware.
The study did not measure on-device latency or power consumption, including the extra cost of WSAN. The analysis window was a global setting selected offline rather than an input-adaptive one. Open questions include replication on larger and more varied multi-session datasets, testing in uncontrolled settings, input-adaptive or multi-scale windows, and deployment measurements for latency and power.
The document is an arXiv version-1 preprint dated 26 August 2026. Its acknowledgements list BB/R019983/1, BB/Y513763/1, BB/S020969/1, EP/X013707/1 and UKRI3606.
Paper data and sources
Original title: LDAC-Net: A Learnable Multi-Lag Differencing Attention-Convolution Network for Drift-Robust Recognition with Low-Cost MOX Gas Sensors
Authors: Xin Zhang, Liangxiu Han, Yue Shi, Tam Sobeih
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text