Live Intelligence Engine

SomaticAI Engine

Watch raw IMU sensor data flow through the Temporal Convolutional Network in real time — from physical motion to labelled and annotated intelligence.

Raw IMU Data

IMU Sensor Stream — 100Hz

0{ax:-1.522, ay:-9.098, az:-0.400, gx:-0.080, gy:0.249, gz:0.152}

1{ax:-1.512, ay:-9.088, az:-0.390, gx:-0.079, gy:0.250, gz:0.153}

2{ax:-1.502, ay:-9.078, az:-0.380, gx:-0.078, gy:0.251, gz:0.154}

3{ax:-1.492, ay:-9.068, az:-0.370, gx:-0.077, gy:0.252, gz:0.155}

4{ax:-1.482, ay:-9.058, az:-0.360, gx:-0.076, gy:0.253, gz:0.156}

5{ax:-1.472, ay:-9.048, az:-0.350, gx:-0.075, gy:0.254, gz:0.157}

↑ live packets @ 795ms latency

→

SomaticAI
Engine

Temporal Convolutional Network (TCN)

Input

TCN Layer 1

TCN Layer 2

TCN Layer 3

Output

Receptive field: 512 timesteps · Dilation: 1, 2, 4, 8

→

Labelled Output

Current Activity

sweeping

Confidence94.0%

sweeping

mopping

lifting

carrying

Annotated Output

joint_angle_shoulder

29.5

joint_angle_elbow

97.6

joint_angle_hip

62.7

force_vector_z

71.6

ergonomic_risk

36.0

temporal_segment

10.3

Output format: JSON · 6 annotation fields · 100Hz resolution

Architecture Overview

Input Layer

· 6-axis IMU (accel + gyro)
· 100Hz sampling rate
· Sliding window: 512 samples
· Normalisation + detrending

TCN Core

· 4 dilated causal conv layers
· Dilation factors: 1, 2, 4, 8
· Residual connections
· Dropout: 0.2 (training only)

Output Head

· Softmax classification (8 classes)
· Confidence score per label
· Joint angle estimation
· Force vector annotation

Trade Secret Protection

The SomaticAI Engine model weights, training pipeline, and annotation methodology are protected as trade secrets. The architecture shown above is a simplified representation for demonstration purposes.