Towards Green Wearable Computing: A Physics-Aware Spiking Neural Network for Energy-Efficient IMU-based Human Activity Recognition

Naichuan Zheng; Hailun Xia; Zepeng Sun; Weiyi Li; Yinzhe Zhou

cs.LGApr 12, 2026

Towards Green Wearable Computing: A Physics-Aware Spiking Neural Network for Energy-Efficient IMU-based Human Activity Recognition

Authors: Naichuan Zheng, Hailun Xia, Zepeng Sun, Weiyi Li, Yinzhe Zhou

Organizations: School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing, China · International School, Beijing University of Posts and Telecommunications, Beijing, China

Abstract

Wearable IMU-based Human Activity Recognition (HAR) relies heavily on Deep Neural Networks (DNNs), which are burdened by immense computational and buffering demands. Their power-hungry floating-point operations and rigid requirement to process complete temporal windows severely cripple battery-constrained edge devices. While Spiking Neural Networks (SNNs) offer extreme event-driven energy efficiency, standard architectures struggle with complex biomechanical topologies and temporal gradient degradation. To bridge this gap, we propose the Physics-Aware Spiking Neural Network (PAS-Net), a fully multiplier-free architecture explicitly tailored for Green HAR. Spatially, an adaptive symmetric topology mixer enforces human-joint physical constraints. Temporally, an $O(1)$ -memory causal neuromodulator yields context-aware dynamic threshold neurons, adapting actively to non-stationary movement rhythms. Furthermore, we leverage a temporal spike error objective to unlock a flexible early-exit mechanism for continuous IMU streams. Evaluated across seven diverse datasets, PAS-Net achieves state-of-the-art accuracy while replacing dense operations with sparse 0.1 pJ integer accumulations. Crucially, its confidence-driven early-exit capability drastically reduces dynamic energy consumption by up to 98%. PAS-Net establishes a robust, ultra-low-power neuromorphic standard for always-on wearable sensing. The source code and pre-trained models are publicly available at https://github.com/zhengnaichuan2022/PAS-Net.git.

Figures & tables

Figure 1. System overview demonstrating the paradigm shift from traditional, high-latency continuous DNNs to our proposed event-driven PAS-Net. PAS-Net enables Green Wearable Computing by fundamentally eliminating multipliers in favor of sparse spiking dynamics and unlocking a sub-second streaming early-exit mechanism for continuous IMU streams.

Figure 2. Overall architecture of our proposed PAS-Net framework. A detailed description of PAS-Net is presented in Section 3

Dataset	Subjects	Classes	Sensor Nodes ( $V$ )	Sampling Rate	Window Size ( $T$ )	Stride	Physical Duration
TNDA-HAR ( Yan et al., 2021 )	50	8	6	$\sim$ 100 Hz	200	100	$\sim$ 2.0 s
PAMAP2 ( Reiss and Stricker, 2012 )	9	18	3	100 Hz	200	100	$\sim$ 2.0 s
Daily-Sports ( Altun et al., 2010 )	8	19	5	25 Hz	125	N/A	5.0 s (Full Segment)
HuGaDB ( Chereshnev and Kertész-Farkas, 2017 )	18	12	6	$\sim$ 50 Hz	100	50	$\sim$ 2.0 s
USC-HAD ( Zhang and Sawchuk, 2012 )	14	12	1	100 Hz	200	100	$\sim$ 2.0 s
HAR70+ ( Ustad et al., 2023 )	18	7	2	50 Hz	100	50	$\sim$ 2.0 s

Table 1. Summary of the selected IMU datasets. Window sizes (

T

) and strides are adaptively configured based on the native sampling rates to maintain a consistent physical observation duration (approx. 2 seconds), with the exception of Daily-Sports, which uses full 5-second episodic segments.

Paradigm	Model Architecture	PAMAP2	Daily-Sports	TNDA	HuGaDB	USC-HAD	HAR70+	Parkinson
Continuous DNNs	DeepConvLSTM ( Ordóñez and Roggen, 2016 )	70.23	84.31	92.16	92.80	64.68	88.31	100.00
	ResNet-SE ( Hu et al., 2018 ; He et al., 2016 )	76.31	86.86	92.60	92.48	64.06	88.36	100.00
	MCHCNN-GRU ( Lu et al., 2022 )	75.35	83.82	91.69	93.40	63.73	87.63	99.97
	Self-HAR ( Tang et al., 2021 )	71.55	88.73	93.00	92.77	58.89	88.09	100.00
	Uni-HAR ( Xu et al., 2023 )	46.22	53.53	36.37	72.37	30.40	78.20	100.00
	rTsfNet ( Enokibori, 2024 )	60.98	48.63	74.44	77.91	40.82	77.70	100.00

Table 2. Overall recognition performance (Accuracy %) on seven diverse IMU datasets under the subject-independent evaluation protocol. The best results among all models are highlighted in bold , and the second best are underlined .

Panel of Figure 3 — Figure 3. Confusion matrices comparing the predictions of the baseline continuous DNN (DeepConvLSTM) against our proposed event-driven PAS-Net on the PAMAP2 and TNDA datasets. By explicitly routing physical dependencies, PAS-Net effectively suppresses off-diagonal noise and mitigates severe inter-class confusion among geometrically similar activities.

Model Architecture	Complexity (PAMAP2)			Energy Consumption ( $\mu$ J) across Datasets
Model Architecture	Params (M)	FLOPs (M)	SOPs (M)	PAMAP2	TNDA	Daily-Sports	HuGaDB	HAR70+	USC-HAD	Parkinson
DeepConvLSTM ( Ordóñez and Roggen, 2016 )	0.30	19.31	–	94.62	94.32	94.32	192.80	93.41	93.11	186.20
ResNet-SE ( Hu et al., 2018 ; He et al., 2016 )	0.82	18.01	–	88.24	87.81	87.82	181.40	86.54	86.13	172.10
MCHCNN-GRU ( Lu et al., 2022 )	0.21	13.20	–	64.70	64.24	64.25	134.80	62.88	62.44	124.90
Self-HAR ( Tang et al., 2021 )	0.84	53.79	–	263.60	263.00	263.00	534.40	261.20	260.60	521.10
Uni-HAR ( Xu et al., 2023 )	0.84	53.77	–	263.40	262.80	262.80	534.10	261.00	260.40	520.80

Table 3. Comprehensive computational cost and system-level energy efficiency analysis. Network Parameters (M), theoretical FLOPs (M), and actual Synaptic Operations (SOPs, M) are profiled based on the PAMAP2 benchmark as the structural representative. The dynamic energy consumption (

\mu

J) is systematically evaluated across all seven diverse datasets to demonstrate consistent efficiency. Missing computational tracking logs from original baselines have been conservatively imputed based on average network firing rates (inclusive of the standardized FP32 continuous encoding overhead prior to the spike domain).

\dagger

Note: The energy profiling for rTsfNet represents only its shallow MLP backend and strictly excludes the massive CPU-based computational overhead required to extract its 45 prerequisite handcrafted time-series features (e.g., FFT), making its true end-to-end system energy significantly higher.

Panel of Figure 4 — Figure 4. Visualization of raw IMU signals (top), dense token features (middle), and deep sparse spike rasters (bottom), demonstrating the extreme sparsity and distinct spatiotemporal firing patterns learned by PAS-Net.

Panel of Figure 7 — Figure 7. t-SNE visualization of the learned latent spaces on the PAMAP2 dataset. The progression from (a) and (b) to (d) illustrates how our explicit spatial topology routing and temporal causal neuromodulation systematically disentangle overlapping kinematic signatures, culminating in a highly structured and separable feature space.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Module Component	Output Tensor Shape	Configuration Details
Input	Raw Continuous IMU Stream	$[200,6,3]$	$C_{in}=6$ (Acc + Gyro), $V=3$ spatial nodes
Perceptual Frontend	Invariant Tokenizer	$[50,24,3]$	Extract mean, max, var, $L_{2}$ norm (stride=4)
Perceptual Frontend	Permute & Spiking Embed	$[50,256,3]$	Conv1d ( $24\to 256$ ), T-BN, Dynamic LIF
Deep Spiking Core	PAS-Block $\times L$	$[50,256,3]$	Stage 1: Symmetric Mask $\tilde{\mathbf{A}}$ , Spatiotemporal Conv ( $k=5$ )
Deep Spiking Core	PAS-Block $\times L$	$[50,256,3]$	Stage 2: Spiking Dilated TCN, $D=256$ , Depth $L=5$
Readout & Exit	Spatial Pooling	$[50,256]$	Fused Mean + Max Pooling across the $V$ dimension

Appendix

Table 6. Detailed architectural configuration and internal tensor flow of PAS-Net (exemplified on the PAMAP2 dataset). The tensor shapes follow the

[T,C,V]

layout, where

T

is the temporal length,

C

is the feature dimension, and

V

is the number of physical sensor nodes. For PAMAP2, the initial temporal window

T=200

is robustly downsampled to

T^{\prime}=50

, and

V=3

Dataset	Nodes ( $V$ )	Input Ch. ( $C_{in}$ )	Classes ( $K$ )	Embed Dim ( $D$ )	Depth ( $L$ )	MLP Ratio
PAMAP2	3	6	12 †	128	3	2.0
Daily-Sports	5	3	19	192	7	4.0
TNDA-HAR	5	3	8	128	5	3.0
HuGaDB	5	6	12	128	5	3.0
USC-HAD	1	6	12	128	5	3.0
HAR70+	2	3	7	96	4	2.5

Appendix

Table 7. Dataset-specific core hyperparameter configurations for PAS-Net.

Layer Name	$T^{\prime}_{1}$	$T^{\prime}_{2}$	$T^{\prime}_{3}$	$T^{\prime}_{4}$	$T^{\prime}_{5}$	$T^{\prime}_{6}$	$T^{\prime}_{7}$	$T^{\prime}_{8}$
Stem	7.40%	9.45%	7.42%	7.42%	8.07%	5.86%	8.55%	8.16%
Block1_Topo	7.22%	10.60%	7.34%	6.81%	8.72%	7.86%	8.03%	7.55%
Block1_Dynamic	8.33%	9.97%	8.29%	6.90%	7.55%	7.73%	8.03%	7.99%
Block1_MLP1	6.03%	9.51%	7.07%	7.31%	6.81%	6.81%	7.83%	7.62%
Block1_MLP2	8.30%	12.83%	8.03%	8.77%	9.51%	7.94%	9.94%	11.20%
Block2_Topo	6.40%	12.05%	8.59%	7.38%	10.63%	8.29%	9.59%	9.38%

Appendix

Table 8. Spatiotemporal Firing Matrix on the PAMAP2 dataset. The table tracks the layer-wise average firing rate (%) across 8 uniform temporal segments (

T^{\prime}_{1}

T^{\prime}_{8}

) of a downsampled sequence. Bold values highlight the “Semantic Spike Enrichment” phenomenon in the deepest temporal layers during critical kinematic phases.

Explore similar work

Aug 10, 2026

LITEWAY: LIghtweight HAR via Temporal Efficient highWAY

Wearable human activity recognition (HAR) remains challenging due to the computational and energy constraints of deep learning models on resource-limited devices. Existing lightweight approaches often rely on recurrent architectures (e.g., GRU and LSTM), limiting parallelism and increasing inference latency. We propose LITEWAY, a modality-agnostic, fully convolutional framework for multichannel sensor time series that replaces recurrent temporal modeling with structured convolutional decomposition. LITEWAY combines lightweight convolutional blocks, strided temporal processing, and convolution-attention pooling to efficiently capture temporal dependencies while reducing computational complexity. We evaluate LITEWAY on 16 HAR datasets against TinyHAR, TinierHAR, and MLP-HAR. LITEWAY achieves competitive macro F1 while reducing model size by 4.06x-9.52x (Light) and 3.87x-9.07x (Full) compared with TinyHAR and TinierHAR. Deployment experiments further show energy reductions of 2.29x-3.14x (Light) and 1.46x-2.01x (Full) compared with TinierHAR and MLP-HAR, highlighting efficient fully convolutional temporal modeling for wearable HAR. The source code is publicly available at https://github.com/dominique-nshimyimana/liteway.

Dominique Nshimyimana, Vitor Fortes Rey, Mengxi Liu +2

RPTU, DFKI Kaiserslautern, Germany

Efficient Neural Network Inference Human Activity Recognition Temporal Convolutional Networks Wearable Sensing Lightweight CNNs

Ask PDF

Mar 15, 2026

Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks

WiFi-based human action recognition (HAR) has gained significant attention due to its non-intrusive and privacy-preserving nature. However, most existing WiFi sensing models predominantly focus on improving recognition accuracy, while issues of power consumption and energy efficiency remain insufficiently discussed. In this work, we present Wi-Spike, a bio-inspired spiking neural network (SNN) framework for efficient and accurate action recognition using WiFi channel state information (CSI) signals. Specifically, leveraging the event-driven and low-power characteristics of SNNs, Wi-Spike introduces spiking convolutional layers for spatio-temporal feature extraction and a novel temporal attention mechanism to enhance discriminative representation. The extracted features are subsequently encoded and classified through spiking fully connected layers and a voting layer. Comprehensive experiments on three benchmark datasets (NTU-Fi-HAR, NTU-Fi-HumanID, and UT-HAR) demonstrate that Wi-Spike achieves competitive accuracy in single-action recognition and superior performance in multi-action recognition tasks. As for energy consumption, Wi-Spike reduces the energy cost by at least half compared with other methods, while still achieving 95.83% recognition accuracy in human activity recognition. More importantly, Wi-Spike establishes a new state-of-the-art in WiFi-based multi-action HAR, offering a promising solution for real-time, energy-efficient edge sensing applications.

Nengbo Zhang, Yao Ying, Lu Wang +3

School of Computing and Information Technology, Great Bay University, Dongguan, China · School of Computer Science, Guangdong University of Petrochemical Technology, Maoming, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China +2

Human Activity Recognition Wi-Fi Sensing Energy-Efficient ML Spiking Neural Networks

Ask PDF

Mar 27, 2026

SPECTRA: An Efficient Spectral-Informed Neural Network for Sensor-Based Activity Recognition

Real time sensor based applications in pervasive computing require edge deployable models to ensure low latency privacy and efficient interaction. A prime example is sensor based human activity recognition where models must balance accuracy with stringent resource constraints. Yet many deep learning approaches treat temporal sensor signals as black box sequences overlooking spectral temporal structure while demanding excessive computation. We present SPECTRA a deployment first co designed spectral temporal architecture that integrates short time Fourier transform STFT feature extraction depthwise separable convolutions and channel wise self attention to capture spectral temporal dependencies under real edge runtime and memory constraints. A compact bidirectional GRU with attention pooling summarizes within window dynamics at low cost reducing downstream model burden while preserving accuracy. Across five public HAR datasets SPECTRA matches or approaches larger CNN LSTM and Transformer baselines while substantially reducing parameters latency and energy. Deployments on a Google Pixel 9 smartphone and an STM32L4 microcontroller further demonstrate end to end deployable realtime private and efficient HAR.

Deepika Gurung, Lala Shakti Swarup Ray, Mengxi Liu +2

RPTU & DFKI, Kaiserslautern, Germany

Human Activity Recognition Efficient Neural Network Inference Edge Computing Frequency-Domain Feature Learning

Ask PDF