LDAC-Net: A Learnable Multi-Lag Differencing Attention-Convolution Network for Drift-Robust Recognition with Low-Cost MOX Gas Sensors

Xin Zhang; Liangxiu Han; Yue Shi; Tam Sobeih

cs.LGAug 26, 2026

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

Organizations: Department of Computing and Mathematics, Manchester Metropolitan University, Manchester, M1 5GD, U.K.

Abstract

Portable electronic-nose systems based on low-cost metal-oxide (MOX) gas sensors offer a practical solution for gas and odour recognition, but their signals are affected by slow chemical transients, drifting sensor offsets, scale variation, and cross-channel correlations. Existing pipelines commonly use fixed first-order temporal differencing (FOTD), which requires a manually selected lag and may discard useful response information. We propose LDAC-Net, an end-to-end learnable multi-lag differencing attention-convolution network that operates directly on multi-channel MOX signals. Its learnable differential feature enhancement front-end combines window-conditioned statistical affine normalisation, which compensates for window-specific offset and scale variation, with learnable multi-lag differencing, which weights and combines temporal differences across multiple lags. A compact attention-convolution backbone subsequently models local transients and longer-range temporal dependencies. On the 50-class SmellNet-Base task, LDAC-Net achieves 68.2% top-1 accuracy, exceeding the best FOTD-preprocessed comparison model by approximately 14 percentage points and the raw-input Transformer by more than 30 points. Ablation studies confirm the contributions of both proposed components. The representation also transfers to SmellNet-Mixtures, improving accuracy from 45.4% to 50.5%, and generalises to the 62-channel eNose-Drift benchmark under strong long-term drift, achieving 70.6% top-1 accuracy and 69.6% macro-F1. These results outperform the best comparison model with dataset-retuned FOTD preprocessing by 8.0 and 3.0 points, respectively, demonstrating that learnable, sensor-aware preprocessing is more effective than fixed handcrafted differencing for low-cost MOX gas-sensor recognition.

Figures & tables

Figure 1: The LDAC-Net architecture. A raw window $\mathbf{x}$ first passes through the LDFE front module (WSAN normalisation and LMLD differencing), is then processed by three AC blocks, and is finally read out by the Attention-Pooling head into the 50-way logits.

Figure 2: The WSAN module. For each multi-channel window, the per-channel mean and standard deviation feed a small MLP that emits per-channel affine offsets $(1{+}\Delta\boldsymbol{\gamma},\Delta\boldsymbol{\beta})$ , which are applied to the standardised window to produce the normalised output.

Figure 3: The LMLD module. The WSAN-normalised window is differenced at a bank of lags $P=\{5,10,25,50\}$ (only $5,25,50$ drawn); each difference is scaled by a learnable per-channel weight, concatenated with the identity stream, mixed by a $1{\times}1$ convolution, layer-normalised, and gated by a squeeze-excitation block.

Figure 4: The AC block: two half-step macaron feed-forward layers around a multi-head self-attention and a ConvModule, with per-step residual connections (dashed) and a final LayerNorm. A whole-block identity bypass realises stochastic depth, and the inset expands the ConvModule.

Figure 5: Example sensor time curves from the three tasks. (a) SmellNet-Base ( Feng et al., 2026 ) : one representative substance per category across the six MOX channels. (b) SmellNet-Mixtures ( Feng et al., 2026 ) : a pure substance and two- and three-component blends across the four mixture channels. (c) eNose-Drift ( Wörner et al., 2025 ) : raw log-resistance of one channel (R46) for the three analytes on six measurement days spanning the year.

Setting	Value
Data / windowing
Channels $C$ / classes	6 / 50
Window length $T$ / stride	300 (swept 100 to 400) / 50
Train / test windows	about 1 650 / about 330
Architecture
Model width $d$	128

Table 1: LDAC-Net configuration and training recipe used throughout, unless a row states otherwise. All values are fixed across seeds.

	Raw input			FOTD input
Model	Acc@1	Acc@5	F1	Acc@1	Acc@5	F1
MLP	$21.09\pm 2.64$	$56.05\pm 1.85$	$16.79\pm 2.67$	$18.92\pm 2.97$	$49.80\pm 5.85$	$16.17\pm 3.37$
CNN	$26.82\pm 1.44$	$67.31\pm 0.65$	$22.80\pm 1.99$	$49.56\pm 7.80$	$83.19\pm 4.03$	$48.12\pm 7.20$
LSTM	$27.72\pm 2.40$	$66.74\pm 1.17$	$25.31\pm 1.73$	$53.78\pm 1.81$	$85.90\pm 1.14$	$52.40\pm 1.99$
Transformer	$37.29\pm 2.34$	$73.22\pm 1.73$	$33.49\pm 1.99$	$53.55\pm 0.83$	$85.58\pm 1.86$	$52.08\pm 0.77$
LDAC-Net (ours)	$\mathbf{68.2\pm 1.6}$	$\mathbf{89.4\pm 0.3}$	$\mathbf{65.2\pm 4.5}$	N/A	N/A	N/A

Table 2: Model performance on SmellNet-Base (50-way; Top-1 / Top-5 accuracy and macro-F1, %, mean

\pm

std). Mainstream baselines (MLP, CNN, LSTM, Transformer) are trained on raw and lag-25 FOTD-preprocessed inputs. LDAC-Net operates end-to-end on raw windows with no external preprocessing. All models use the same analysis window length

T=300

at stride

50

. Best per metric in bold.

Panel of Figure 6 — Figure 6: Model performance across analysis window lengths on SmellNet-Base (50-way, Top-1 %, mean $\pm$ std). Solid: LDAC-Net; dashed: LDAC-Net w/o LMLD. LDAC-Net rises by more than ten points from the default $T=100$ to a peak at $T=300$ , then falls at $T=400$ as each session yields too few windows. LMLD is neutral at $T=100$ , where a lag- $50$ difference is half boundary, and its benefit grows with the window, reaching $+5.9$ points at $T=300$ .

Configuration	Acc. (%)	$\Delta$
LDAC-Net	$\mathbf{68.2\pm 1.6}$	(ref.)
$-$ WSAN (plain InstanceNorm)	$54.2\pm 0.9$	$-14.0^{\ast}$
$-$ LMLD	$62.3\pm 1.2$	$-5.8^{\ast}$
$-$ squeeze-excitation gate	$61.9\pm 1.3$	$-6.2^{\ast}$
$-$ channel mixing	$63.2\pm 0.4$	$-4.9^{\ast}$
$-$ multi-lag bank (single lag $25$ )	$63.6\pm 2.5$	$-4.6^{\ast}$

Table 3: Component ablation of LDAC-Net at

T=300

(Top-1 %, mean

\pm

std over five seeds). Each row removes one part of LDAC-Net. All accuracy drops are statistically significant against the full model (

{}^{\ast}p<0.05

, two-sided paired

t

-test over the five shared seeds).

Model	Top-1 Acc. (%)
Transformer (raw)	$37.29\pm 2.34$
Transformer (FOTD)	$53.55\pm 0.83$
Non-stationary Transformer ( Liu et al., 2022 )	$52.61\pm 2.50$
Autoformer ( Wu et al., 2021 )	$50.28\pm 1.22$
TCN ( Bai et al., 2018 )	$43.63\pm 2.62$
Neural ODE ( Chen et al., 2018 )	$50.47\pm 1.51$

Table 4: Comparison of LDAC-Net against the Transformer backbone (on raw and lag-25 FOTD input) and four published time-series methods (the Non-stationary Transformer, Autoformer, TCN, and Neural-ODE) on SmellNet-Base (Top-1 %, mean

\pm

std).

Model	[email protected] $\uparrow$	MAE $\downarrow$	Top- $K$ $\uparrow$
MLP	$43.0\pm 1.5$	$0.056\pm 0.001$	$76.2\pm 1.2$
CNN	$43.1\pm 1.3$	$0.059\pm 0.000$	$75.1\pm 0.8$
LSTM	$41.0\pm 2.3$	$0.056\pm 0.001$	$80.8\pm 1.5$
Transformer	$45.4\pm 3.1$	$0.054\pm 0.002$	$78.3\pm 0.3$
LDAC-Net (ours)	$\mathbf{50.5\pm 2.0}$	$\mathbf{0.050\pm 0.001}$	$\mathbf{81.1\pm 0.2}$

Table 5: Model performance on the SmellNet-Mixtures task ( Feng et al., 2026 ) (12-component ratio prediction). [email protected] and Top-

K

in %, higher is better; MAE lower is better. Best per metric in bold.

	Raw		FOTD (lag $100$ )
Model	Acc@1	F1	Acc@1	F1
MLP	$56.3{\pm}4.3$	$55.6{\pm}4.5$	$53.4{\pm}1.4$	$52.4{\pm}1.2$
CNN	$59.2{\pm}5.3$	$58.1{\pm}6.7$	$62.6{\pm}3.1$	$66.6{\pm}3.2$
LSTM	$43.4{\pm}3.4$	$38.9{\pm}3.1$	$52.7{\pm}1.8$	$47.8{\pm}2.3$
Transformer	$56.3{\pm}3.5$	$55.3{\pm}3.9$	$61.2{\pm}3.8$	$58.5{\pm}5.4$
LDAC-Net (ours)	$\mathbf{70.6{\pm}4.8}$	$\mathbf{69.6{\pm}3.9}$	N/A	N/A

Table 6: Cross-dataset generalisation on the eNose-Drift benchmark ( Wörner et al., 2025 ) (

62

-channel MOX array, chronological drift split, raw drift-affected input). Acc@1 / macro-F1 (%, mean

\pm

std; ten seeds for LDAC-Net, the Transformer, and the FOTD CNN, five seeds otherwise); FOTD uses the dataset-retuned lag-

100

difference. Best per column in bold.

Lag bank $P$	Top-1 Acc. (%)
$\{25\}$	$63.6\pm 2.5$
$\{25,50\}$	$64.4\pm 5.0$
$\{5,10,25\}$	$64.9\pm 1.5$
$\{10,25,50\}$	$64.6\pm 3.3$
$\{15,25,50\}$	$64.0\pm 0.6$
$\{25,50,100\}$	$64.5\pm 3.0$

Table 7: Effect of the LMLD lag bank on SmellNet-Base (50-way,

T=300

, Top-1 %, mean

\pm

std). The bank used by LDAC-Net is in bold.

Classification

Classification maps observations to discrete labels, including familiar, unknown, hierarchical, or multiple categories.

See all 410 papers

Online gas classification at the edge
Oct 27, 2026Ton Juny Pina, Hugh Greatorex, Elisabetta ChiccaClassification Binary Neural Networks
Bio-Inspired Circuits and Systems (BICS) Lab, Zernike Institute for Advanced Materials, University of Groningen, Netherlands. · Groningen Cognitive Systems and Materials Center (CogniGron), University of Groningen, Netherlands. · Institute of Neuroinformatics, University of Zurich and ETH Zurich.
Are Time-Series Foundation Models Ready for E-Nose Data? An Empirical Assessment of Their Embeddings
Jun 26, 2026Taeyeong Choi, Mohammed KamruzzamanTime Series Classification Time Series Representation Learning
Department of Information Technology Kennesaw State University Marietta, GA 30060, USA · Department of Agricultural and Biological Engineering University of Illinois Urbana-Champaign Urbana, IL 61801, USA
Physics Closure Matters for Machine Olfaction: A Maxwell-Stefan Graph Solver for Identifiable Dynamic Gas Unmixing
Jul 20, 2026Yue Shi, Liangxiu Han, Xin Zhang +1Graph Neural Networks Neural PDE Solvers
Department of Computing and Mathematics, Faculty of Science and Engineering, Manchester Metropolitan University, Manchester M1 5GD, U.K.

Oct 27, 2026J/K move · Enter open · S save

Online gas classification at the edge

Ton Juny Pina, Hugh Greatorex, Elisabetta Chicca

Bio-Inspired Circuits and Systems (BICS) Lab, Zernike Institute for Advanced Materials, University of Groningen, Netherlands. · Groningen Cognitive Systems and Materials Center (CogniGron), University of Groningen, Netherlands. · Institute of Neuroinformatics, University of Zurich and ETH Zurich.

Classification Binary Neural Networks Neuromorphic Computing Online Learning

Open paper AskPDF arXiv

Field applications in chemical sensing, from portable safety monitors to mobile robots, require models that are lightweight, fast to respond, and can be recalibrated from limited labelled data. Current gas classification pipelines assume controlled sampling cycles and offline feature extraction, and rely on computationally expensive models. With a few thousand neurons and sub-milliwatt power consumption, insects reliably classify odours in open environments. Inspired by the insect olfactory system, we present a neural network architecture that classifies raw sensor array readings using a sparse, high-dimensional binary representation. We evaluate it on random sequences of single gases and binary mixtures, introducing new samples dynamically and classifying gas identity from the transient sensor response. In this setting, the architecture out-performs a linear \gls{svm} on gas mixtures and matches an optimised non-linear \gls{svm}, while maintaining online class-level frequency estimates that distinguish familiar from novel stimuli. It relies exclusively on binary weights and Hebbian-based local learning, remaining computationally lightweight and well suited to ultra-low-power neuromorphic processors. We demonstrated the network on-chip with a hazardous gas detection experiment running on the mixed-signal neuromorphic chip TEXEL. The on-chip classifier yielded an accuracy of 74.6% while running at 482

\unit{\nano\joule}

per inference and 27 \unit{\micro\watt} of static-power.

Abstract

Classification

Explore similar work

Online gas classification at the edge

Are Time-Series Foundation Models Ready for E-Nose Data? An Empirical Assessment of Their Embeddings

Physics Closure Matters for Machine Olfaction: A Maxwell-Stefan Graph Solver for Identifiable Dynamic Gas Unmixing