Does SMA always exclude gravity?

No; that depends on the stated preprocessing and definition.

Is SMA an energy measure?

The common absolute-sum form is not physical energy and differs from squared-signal energy features.

Is it invariant to rotating the sensor?

Generally no, because an L1 component sum changes under arbitrary rotations.

Can two windows have identical SMA but different motion?

Yes; the summary discards much of the timing and directional information present in the original sequence.

AI Glossary / Statistics and probability / Intermediate

Signal magnitude area

Definition

Signal magnitude area, commonly abbreviated SMA, is a feature that summarizes the absolute magnitude of a signal over a time window. For three-axis accelerometer data, a common definition averages the sum of the absolute values of the three components across the window. It is a time-domain summary used in activity recognition and related sensor analysis. Definitions and preprocessing vary, so an SMA value should always be accompanied by its formula, units, window and treatment of gravity or baseline components.

What it means

SMA turns a sequence of signed sensor readings into a nonnegative summary of their overall component-wise magnitude. Taking absolute values prevents positive and negative movements from canceling. For equally spaced three-axis samples, the common normalized form is the average of |x| + |y| + |z|. This is different from averaging the Euclidean vector length, which uses the square root of x squared plus y squared plus z squared. Neither quantity by itself identifies a specific activity or measures metabolic energy.

How it works

Calibrate the sensor, select a window and apply the preprocessing required by the intended feature definition. Some applications use raw acceleration, while others remove gravity, subtract a local mean or filter particular frequency components. Calculate the absolute value on each axis, sum across axes and average across samples for a uniformly sampled normalized feature. A continuous-time version divides the integral of the component-wise absolute sum by window duration. Uneven sampling requires appropriate time weighting rather than an unexamined average over records.

Three practical examples

Example 1: A wearable activity classifier uses SMA alongside other features to help distinguish low and high movement windows. Example 2: An industrial sensor pipeline summarizes component-wise vibration magnitude before applying an AI anomaly detector. Example 3: A research prototype compares movement summaries from two devices while controlling placement and preprocessing. These are potential feature uses, not proof that one SMA threshold can identify an activity reliably across people, devices and environments.

Think of it this way

Think of adding the size of every movement along three marked directions without allowing movements in opposite directions to cancel. The result describes how much component-wise signal appeared during the window. It does not retain the exact sequence of movements. Two very different motion patterns can therefore have the same SMA, just as two journeys can cover the same total component-wise distance while following different routes.

Worked example

Consider four equally weighted three-axis samples in acceleration units: (1,2,2), (-1,-2,-2), (0,0,0) and (0,0,0). Their absolute component sums are five, five, zero and zero. The normalized SMA is (5 + 5 + 0 + 0)/4 = 2.5 acceleration units. The corresponding Euclidean vector lengths are three, three, zero and zero, whose average is 1.5, showing that the two features differ. If each sample represents a constant half-second interval, the unnormalized absolute-sum integral is 0.5 times ten = five acceleration-unit seconds, and dividing by the two-second duration again gives 2.5. Orientation also matters: rotating the vector (1,0,0) to (1/square root of two,1/square root of two,0) preserves Euclidean length one but increases its absolute component sum from one to square root of two.

Illustrative case study

In this fictional case study, Brookfield Devices develops an AI classifier for machinery activity. Engineers initially compare SMA values from sensors mounted at different angles and discover unexpected differences. They standardize the mounting convention, document filtering and calculate several complementary features. Evaluation uses separate operating sessions rather than randomly mixing overlapping windows from the same recordings. The report describes SMA as one input to a validated task-specific model, without equating the feature directly with machine health or energy consumption.

Common misconceptions and mistakes

Common mistakes include confusing SMA with signal vector magnitude, forgetting normalization by window length and comparing raw and gravity-removed signals as if they were equivalent. Another error is assuming rotation invariance: the sum of absolute axis components is generally orientation-dependent. A mean-subtracted version can remove a constant component but does not reconstruct physical motion perfectly. Sensor noise, clipping, sampling rate and filter choices can all alter the feature. Equal SMA values do not imply equal activities or identical temporal patterns.

Business use

Businesses can use SMA as an inexpensive interpretable feature in sensor-based monitoring and AI pipelines. Its usefulness should be evaluated for the intended task and sensor arrangement, with baseline performance and independent test data. Document the units, axis conventions, sample interval, window overlap and preprocessing so the feature can be reproduced. When comparing devices or sites, assess calibration and placement effects. Operational decisions should depend on validated relationships to the target outcome, not an assumed universal meaning of a high or low value.

Check your understanding

What is the normalized SMA in the worked example?

Answer: 2.5 acceleration units.

Why do opposite signed readings not cancel?

Answer: Absolute values are taken before summation.

What is the mean Euclidean vector magnitude for those same samples?

Answer: 1.5 acceleration units.

Why must an unnormalized integral be labeled differently?

Answer: It depends on duration and has acceleration-times-time units rather than acceleration units.

Related terms

Root mean square
Related terms include accelerometer
time-domain feature
L1 norm
signal vector magnitude
signal energy
windowing and activity recognition
whereas vector magnitude uses the L2 norm

Learn next

Learn next how filtering and gravity estimation affect accelerometer features. Compare L1 and L2 summaries under rotations and changes in window duration. Then examine how overlapping windows can create dependence or evaluation leakage, and how combining time-domain and frequency-domain features can retain information that a single scalar summary loses.

Sources

Sources: A Comprehensive Analysis on Wearable Acceleration Sensors in Human Activity Recognition, https://pmc.ncbi.nlm.nih.gov/articles/PMC5375815/; Gubbi and colleagues, Motor recovery monitoring using acceleration measurements in post acute stroke patients, https://biomedical-engineering-online.biomedcentral.com/counter/pdf/10.1186/1475-925X-12-33; Activity Tracking Using Ear-Level Accelerometers, https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2021.724714/full. The vector calculations and fictional machinery case are original illustrations, not clinical guidance or measured detection performance.

Previous: Signal detection theory Next: Signaling compression