Online gas classification at the edge

Ton Juny Pina; Hugh Greatorex; Elisabetta Chicca

cs.ETOct 27, 2026

Online gas classification at the edge

Authors: Ton Juny Pina, Hugh Greatorex, Elisabetta Chicca

Organizations: 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.

Abstract

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.

Figures & tables

		Train accuracy		Test accuracy
Method	Readout	Single ( $n{=}600$ )	Binary ( $n{=}600$ )	Single ( $n{=}300$ )	Binary ( $n{=}300$ )
Expand & sparsify	Hebbian – binary weights	$0.859\pm 0.009$	$0.602\pm 0.014$	$0.881\pm 0.014$	$0.645\pm 0.017$
Linear SVM	NA	$0.902\pm 0.000$	$0.551\pm 0.000$	$0.923\pm 0.000$	$0.563\pm 0.000$
Expand & sparsify	Linear classifier	$0.929\pm 0.004$	$0.849\pm 0.004$	$0.952\pm 0.003$	$0.791\pm 0.008$
RBF SVM	NA	$0.927\pm 0.000$	$0.853\pm 0.000$	$0.940\pm 0.000$	$0.800\pm 0.000$

Table 1: Train and test accuracy for the expand-and-sparsify architecture compared to linear and non-linear SVM methods. Network read-outs are reported as mean

\pm

std over 10 repetitions with independent random projections; the linear and RBF SVM baselines are deterministic fits on the same sequential train/test split as the network and are therefore reported from a single run (std

=0

Experiment	$M$	$p_{HD}$	$k$	$s$	Read-out	$L$	$p$
Classification (Hebbian)	$10\,000$	$0.025$	$5000$	$0.5$	Hebbian, 1-bit	3 / 6	$0.01$
Classification (GD)	$10\,000$	$0.025$	$5000$	$0.5$	GD, 64-bit	3 / 6	–
Few-shot	$10\,000$	$0.025$	$5000$	$0.5$	Hebbian, 1-bit	3 / 6	$0.01,\,0.1$
Exposure time	$10\,000$	$0.025$	$5000$	$0.5$	Hebbian, 1-bit	3 / 6	$0.01$
Count sketch	$10\,000$	$0.025$	$5000$	$0.5$	Hebbian, 1-bit	1	$10^{-3}$ – $4\!\times\!10^{-5}$
On-chip (TEXEL)	$2000$	$0.025$	$200$	$0.9$	STDP , binary	2	–

Table 2: Network parametrisation per experiment.

M

: dimensionality of the HD projection;

p_{\mathrm{HD}}

: density of the random binary projection matrix;

k

: number of

k

-WTA winners;

s=1-k/M

: sparsity of the HD representation;

L

: read-out dimensionality (single gases / binary mixtures);

p

: potentiation probability of the read-out weights. All software experiments used the

N=136

input expansion.

		Test accuracy ( $x_{\mathrm{dense}}$ )		Test accuracy ( $z_{\mathrm{sparse}}$ )
Expansion	$n_{\mathrm{dense}}$	Single gases	Binary mixtures	Single gases	Binary mixtures
raw	8	$0.793\pm 0.083$	$0.390\pm 0.054$	$0.840\pm 0.006$	$0.533\pm 0.015$
$\partial^{1}$	16	$0.919\pm 0.069$	$0.598\pm 0.074$	$0.969\pm 0.001$	$0.704\pm 0.008$
$8\text{D}$	16	$0.795\pm 0.070$	$0.368\pm 0.057$	$0.888\pm 0.014$	$0.613\pm 0.015$
$8\text{R}$	16	$0.807\pm 0.068$	$0.395\pm 0.061$	$0.837\pm 0.011$	$0.537\pm 0.006$
$\partial^{1}+\partial^{2}$	24	$0.933\pm 0.057$	$0.622\pm 0.074$	$0.969\pm 0.002$	$0.721\pm 0.015$

Table 3: Linear SVM accuracy on

\mathbf{x}_{\text{dense}}

and

\mathbf{z}_{\text{sparse}}

for different expansions of the sensor traces.

\partial^{1}

and

\partial^{2}

denote the first- and second-order temporal derivatives approximated with backward differences.

n\text{D}

and

n\text{R}

denote

n

pairwise differences and ratios between sensor traces respectively.

n_{\text{dense}}

is the resulting dimensionality. Test accuracy is reported as mean

\pm

standard deviation over 100 repetitions.

Figure 1: System overview. The sensor array is exposed to a randomized sequence of gas mixtures. Its response is then processed by the expand-and-sparsify network producing online category estimates.

Figure 2: The expand-and-sparsify architecture. As the sensor array responds to a randomized sequence of gas mixtures, the first and second order derivatives of each sensor signal are approximated using backward differences, and the 56 pairwise channel ratios and differences are appended to form a 136-dimensional feature vector. This signal is then projected onto a higher-dimensional space via multiplication with a random sparse binary matrix, and subsequently sparsified through a $k$ -WTA mechanism. A linear read-out layer produces online class-identity estimates.

Figure 3: Class identity estimates from the read-out neurons activity. The activity of the read-out layer neurons as the sensor array signals are processed through the expand-and-sparsify architecture. The sensor array is exposed to a random sequence of odours (CO, ethylene, methane). During the training phase, learning signals trigger the potentiation of the output synapses, enforcing the output neurons to learn the sparse representations of its assigned category. The weights are then frozen during the testing phase.

Figure 4: Few-shot learning. Top row: Train and test accuracy (mean

\pm

s.d. over 10 repetitions) versus the training set size for the single-gas dataset. With p = 0.01 (left), accuracy rises gradually and peaks at 0.887 ± 0.008 with 375 exposures. A higher rate (p = 0.1, right) reaches a comparable peak (0.886 ± 0.033) with only 105 exposures, at the cost of larger variance and a decline at larger training set sizes. Bottom: zoomed-in view of the

p=0.1

panel. The network reaches near-peak accuracy within the first 50 exposures.

Figure 5: Performance over exposure time. Test accuracy (mean

\pm

s.d. over 10 repetitions) versus time elapsed since the start of the exposure for the single-gas dataset (left) and the binary mixtures (right). The model is trained on the full training set and evaluated on individual test frames grouped by position. Dashed gray lines indicate the overall mean test accuracy.

Figure 6: Count sketch. The mean activity of the read-out neuron for each mixture in a random sequence of 200 presentations per gas of single gases.

Figure 7: On-chip hazardous emission detection a) The custom mixed-signal neuromorphic processor TEXEL. It hosts 180 neurons distributed across two cores, with each neuron equipped with analogue synaptic circuits that emulate local plasticity mechanisms. b) The network architecture mapped onto the chip. HD input vectors are distributed across neurons and their synaptic inputs, then converted into single spikes at nonzero vector entries with random delays relative to a teacher spike during training. c) Raster plot of the input spike pattern for an inference sample, and the spiking responses of the read-out neurons. d) The ROC curve and the confusion matrix for the operating point with the highest Youden index (0: hazardous, 1: non-hazardous). e) Histogram of the dynamic power consumption per inference across the test sequence. The table shows the energy cost per spike and Synaptic Operation (SynOp) in both analogue and digital domains, as reported in [ 21 ] .

Figure 8: Experimental setup. a) Photography taken during a recording session. b) Schematic view of the experimental setup. c) Sensor array signals. Traces recorded from the 8 active sensors during 600s of continuous sampling. d) Sensor read-out circuit. The sensor acts as a variable resistor. The transistor, controlled by the digital enable signal (EN), switches the heater on/off. The sensor and the load resistor form a voltage divider. The voltage $\text{V}_{\text{sensor}}$ is fed into a non-inverting amplifier with a gain set by $\text{R}_{\text{g1}}$ and $\text{R}_{\text{g2}}$ .

Figure 9: The kernel induced by the expand-and-sparsify architecture. a) The measured kernel for random pairs of expanded sensor read-outs ( $n_{pairs}=10000$ , $d_{input}=136$ , $p_{HD}=0.2$ and $1-s=0.2$ ) compared against the theoretical kernel. b) The theoretical kernel as a function of the sparsity of the final representation, determined by $k$ in the $k$ -WTA. c) Histograms representing the inner products between random pairs of sensor read-outs (left) and its sparsified representations (right).

Figure 10: Effect of the input dimensionality on the projection statistics. Top: raw 8-channel sensor traces. Bottom: 136-dimensional expanded traces. Histograms show the empirical density of the $\mathbf{z}_{\text{dense}}$ components for 3 random samples, with the MSE to the standard normal annotated ( $n=10000$ ). Heat-maps show the joint density for 3 random pairs of samples ( $n=10000$ ), with the Pearson correlation between the samples annotated. Scatter plots show the measured kernel versus the theoretical one described by Eq. 7 for random pairs of samples ( $n=10000$ ).

#	Model	Manufacturer	Target gas(es)	Detection range
1	MG812	Winsen	CO 2	350–10,000 ppm
2	WSP2110	Winsen	VOCs (toluene, benzene, formaldehyde, alcohol, acetone)	1–50 ppm
3	TGS1820	Figaro	Acetone	1–20 ppm
4	TGS2600	Figaro	Air contaminants (e.g., H 2 , ethanol)	1–30 ppm (as H 2 )
5	TGS2602	Figaro	Air contaminants (VOCs, NH 3 , H 2 S, etc.)	1–30 ppm (as EtOH)
6	TGS2603	Figaro	Odorants/air contaminants (amines, sulfur compounds)	1–10 ppm (as EtOH)

Table S1 : Sensor target gases and detection ranges (typical values from manufacturers).

Figure S1 : Sensor array characterisation. Response of the sensor array to single-gas baseline–exposure–recovery cycles for the three analytes (columns: methane, ethylene, CO). Top row: spatial maps of the mean response amplitude over the physical $4\times 4$ sensor layout, averaged over five repetitions and min–max normalised per analyte to $[0,1]$ (arbitrary units; colour bar). Marker outlines indicate whether a sensor is used in the classification experiments — black for the eight most-responsive sensors retained by the reduced data loader, grey for the excluded ones. Middle and bottom rows: distributions of the exposure ( $\tau_{\mathrm{exp}}$ ) and recovery ( $\tau_{\mathrm{rec}}$ ) time constants, obtained from single-exponential fits to the rising and decaying transients and pooled over the five repetitions and the eight retained sensors. The time constants lie predominantly in the $20$ – $40,\mathrm{s}$ range, consistent with the sensor response times quoted in the data-acquisition protocol.

Classification

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

See all 410 papers

Apr 29, 2026

EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures

We propose EdgeSpike, a co-designed spiking neural network (SNN) framework for autonomous low-power sensing in edge Internet of Things (IoT) architectures. EdgeSpike unifies (i) a hybrid surrogate-gradient and direct-encoding training pipeline, (ii) a hardware-aware neural architecture search (NAS) bounded by per-inference energy and memory budgets, (iii) an event-driven runtime targeting Intel Loihi 2, SpiNNaker 2, and commodity ARM Cortex-M microcontrollers with custom spike-sparse SIMD kernels, and (iv) a lightweight local plasticity rule enabling continual on-device adaptation without backpropagation. The framework is evaluated across five sensing tasks (keyword spotting, vibration-based machine fault detection, surface electromyography gesture recognition, 77 GHz radar human-activity classification, and structural-health acoustic-emission monitoring) on three hardware targets. EdgeSpike achieves a mean classification accuracy of 91.4%, within 1.2 percentage points (pp) of strong INT8 convolutional neural network (CNN) baselines (mean 92.6%), while reducing energy per inference by 18x to 47x on neuromorphic hardware (mean 31x) and by 4.6x to 7.9x on Cortex-M (mean 6.1x). End-to-end latency remains at or below 9.4 ms across all 15 task-hardware configurations. A seven-month, 64-node wireless field deployment confirms a 6.3x extension in projected battery lifetime (from 312 to 1978 days at 2 Wh per node) and bounded accuracy degradation under seasonal drift (0.7 pp with on-device adaptation versus 2.1 pp without). Hardware-aware NAS evaluates 8400 candidates and yields a 12-point Pareto front. EdgeSpike will be released as open source with reproducible training pipelines, hardware-portable runtimes, and benchmark suites.

Gustav Olaf Yunus Laitinen-Fredriksson Lundstrom-Imanov, Taner Yilmaz

Department of Economics, Stockholm University, Universitetsvägen 10 A, SE-106 91 Stockholm, Sweden · Department of Computer Engineering, Afyon Kocatepe University, 03200 Afyonkarahisar, Türkiye

Energy-Efficient ML Edge Inference Internet of Things Spiking Neural Networks Neuromorphic Computing

Ask PDF

Aug 26, 2026

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

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.

Xin Zhang, Liangxiu Han, Yue Shi +1

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

Classification

Ask PDF

Jul 22, 2026

Self-organizing Architecture of Receptron Units: a Hardware-Aware Framework for Edge Intelligence

The growing demand for intelligent processing at the edge of IoT networks is constrained by the severe computational and memory limitations of microcontroller units, which render impractical conventional deep learning approaches. We propose a neuromorphicinspired classifier based on the Receptron model, a single-unit architecture capable of implementing non-linearly separable decision boundaries, without resorting to multi-layer networks. The model is designed for direct deployment on mid-range MCUs, while supporting continuous on-device adaptation. Experimental evaluation on basic dataset benchmarks yields cross-validated accuracies compatible with standard machine learning method baselines. These results position the Receptron as a viable and interpretable alternative for resource-constrained neuromorphic edge systems operating in dynamic, non-stationary environments.

Stefano Radice, Ludovico Casaccia, Riccaro Emanuele Beccalli +2

Department of Physics “Aldo Pontremoli” and CIMaINa University of Milano Milano, Italy

Shallow Neural Networks Neuromorphic Computing Edge ML

Ask PDF

Abstract

Classification

Explore similar work

EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures

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

Self-organizing Architecture of Receptron Units: a Hardware-Aware Framework for Edge Intelligence