Self-Powered Wearable Sensor Tracks Exercise Posture and Burns Calories With 95% Accuracy

A new study published in Discover Artificial Intelligence describes a self-powered wearable sensing system that can recognize sports exercise postures and estimate how many calories a person is burning, all in real time. The research, conducted by Haiwen Deng of Hunan University of Arts and Science, tackles one of the most persistent problems in fitness technology: consumer wearables are convenient, but they often cannot tell what movement you are actually doing, and their calorie estimates can be wildly inaccurate. By combining two small inertial sensor nodes with a carefully engineered machine-learning pipeline, the study reports posture recognition accuracy of 95.4 percent and energy-expenditure predictions substantially better than both standard baselines and a commercial wrist tracker.

The hardware at the heart of the study, a prototype designated EPS-01, is what the author calls an energy power sensor, although the paper is careful to clarify that this term is shorthand rather than a new sensing principle. Each node contains a triaxial accelerometer with a range of plus or minus 16 g, a triaxial gyroscope rated to 2,000 degrees per second, a microcontroller, a wireless communication unit, an energy-harvesting interface, and a supercapacitor-based power buffer. Two nodes were worn during testing, one on the dominant wrist and one on the lower back near the fifth lumbar vertebra, a placement chosen to capture both the fine dynamics of the limbs and the gross motion of the trunk. All twelve inertial channels were sampled at 100 Hz with 16-bit resolution, and wireless clocks were synchronized before each trial so that residual timing deviation between nodes stayed below two milliseconds.

The energy-harvesting unit is an important but frequently misunderstood part of the design. It combines a motion-energy-harvesting element, a rectification and regulation circuit, and a supercapacitor buffer, and its purpose is to extend battery life during repeated exercise, not to measure human metabolism. The paper even includes a detailed solar-irradiance model, using Cooper’s equation for solar declination and the Hottel clear-sky transmittance model, to estimate how much auxiliary power a node could theoretically harvest in outdoor conditions. Crucially, the author draws a firm boundary between this electrical power management and the biological energy-expenditure model: harvested power affects sensor availability, never the calorie predictions. When harvesting falls short, for example indoors or under clothing, the controller shifts through three power states, reducing radio duty cycles or transmitting only calibrated inertial summaries until the buffer recovers.

On the data side, sixty adults aged 18 to 35, balanced between 30 women and 30 men with body mass index values from 18.5 to 27.9, performed eight representative movements: standing, walking, jogging, squatting, lunging, jumping jacks, sit-ups, and push-ups. Five valid trials per movement produced 2,400 labeled motion sequences. For energy validation, four dynamic exercises were performed at three controlled intensities, generating 720 five-minute exercise bouts, each followed by adequate recovery. Breath-by-breath indirect calorimetry provided the reference energy-expenditure values, calculated from the stable final two minutes of each bout. Importantly, participants rather than individual data windows were split into training, validation, and test subsets at a ratio of 70:10:20, a subject-independent protocol that prevents the model from simply memorizing individual movement quirks.

Signal preprocessing proved essential to the system’s performance. Raw acceleration and angular-velocity signals were calibrated using six-position static calibration for accelerometers and zero-rate and constant-rate calibration for gyroscopes, bringing calibration errors down to within 0.04 g and 1.2 degrees per second respectively. A fourth-order zero-phase Butterworth band-pass filter from 0.25 to 20 Hz removed baseline drift and high-frequency noise, and a five-sample median filter suppressed impulsive artifacts. These steps raised the average signal-to-noise ratio from 17.8 dB to 26.4 dB. The synchronized signals were then divided into 2.56-second windows with 50 percent overlap, preserving transition information between movements while keeping latency low.

The recognition architecture pairs a one-dimensional convolutional neural network with a support vector machine. The CNN contains three convolutional blocks with 64, 128, and 256 filters and kernel sizes of 7, 5, and 3, each followed by batch normalization and rectified-linear activation, with max pooling after the first two blocks. Global average pooling produces a 256-dimensional representation that feeds a fully connected layer of 128 units with 0.40 dropout, and an SVM with a radial-basis-function kernel performs the final classification. The rationale is pragmatic: the CNN learns local temporal features directly from raw inertial signals, while the SVM provides a stable margin-based decision boundary suited to a moderate-sized dataset. The entire feature extractor contains only about 0.42 million trainable parameters and requires 5.9 million floating-point operations per window, running on a mobile edge processor with a mean inference latency of 6.3 milliseconds, far shorter than the 1.28-second window update interval.

The results were strong across the board. The CNN-SVM pipeline achieved 95.4 percent test accuracy, 95.5 percent macro-precision, 95.4 percent macro-recall, and a macro-F1 score of 95.2 percent, outperforming a conventional handcrafted-feature SVM at 88.7 percent, a standalone CNN at 93.2 percent, an LSTM at 94.1 percent, and a compact Transformer at 94.6 percent. The Transformer matched the leaders on accuracy but needed 14.8 milliseconds per window, more than double the proposed method’s latency. Across five participant-level cross-validation folds, mean accuracy was 95.1 percent with a standard deviation of 1.8 percent. The most common confusion was between squatting and lunging, which both involve repeated knee and hip flexion, with smaller error clusters appearing between sit-ups and push-ups during transition windows.

Energy-expenditure estimation followed a different strategy built on the insight that posture context matters. In the dataset, acceleration root mean square showed the strongest univariate correlation with reference expenditure at r equal to 0.74, followed by heart rate at 0.69, body mass at 0.42, cadence at 0.39, and age at negative 0.18. Posture class was retained as a categorical input because identical acceleration magnitudes can reflect very different muscular demands; squatting repeatedly displaces the body’s center of mass, while push-ups recruit entirely different muscle groups despite similar trunk acceleration. A compact regression network was then tuned by Bayesian optimization using a Gaussian-process surrogate with a Matérn 5/2 covariance function and an expected-improvement acquisition function. After 50 evaluations, the optimized model reduced root mean square error from 1.12 to 0.82 kcal/min and mean absolute error from 0.84 to 0.61 kcal/min, while raising the coefficient of determination from 0.84 to 0.91. Against a standard metabolic-equivalent baseline, which produced an RMSE of 1.47 kcal/min, the optimized model cut error by 44.2 percent. On a 20-participant subset, a commercially available wrist tracker produced an RMSE of 1.21 kcal/min and offered no posture labels, whereas the two-node system achieved 0.82 kcal/min while distinguishing all eight movement classes.

The study is candid about its limits. Prediction error rose with exercise intensity, with RMSE values of 0.55, 0.78, and 1.08 kcal/min for low-, moderate-, and high-intensity bouts respectively, reflecting stronger motion artifacts, transient breath dynamics, and greater physiological variability at high cadence. All participants were young adults tested under supervised indoor conditions, and sensor displacement during free-living behavior could degrade generalization. The energy harvester extended operating time but could not continuously supply peak wireless and processing power, and the calculated solar irradiance should be read as a theoretical upper bound rather than guaranteed operating power. Future work, the author notes, should include older and clinical populations, outdoor free-living validation, automatic sensor-placement correction, domain adaptation, and model quantization. Even so, the combination of subject-independent accuracy, millisecond-scale latency, and posture-aware calorie estimation points toward a generation of fitness wearables that understand not just how hard you are moving, but exactly how you are moving.

Subject of Research: Wearable sensor-based sports posture recognition and energy expenditure estimation using machine learning

Article Title: Research on sports exercise posture monitoring and energy consumption law based on energy power sensors

Article References: Deng, H. (2026). Research on sports exercise posture monitoring and energy consumption law based on energy power sensors. Discover Artificial Intelligence, 6(1), Article 1384. https://doi.org/10.1007/s44163-026-02029-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02029-x

Keywords: wearable sensors, posture recognition, energy expenditure, CNN-SVM, Bayesian optimization, inertial measurement unit, energy harvesting, sports technology, machine learning, indirect calorimetry, fitness tracking, edge computing