Wearable Smart Sensor With Incremental CNN and Early Exit
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing smart sensors face challenges in implementing Convolutional Neural Networks (CNNs) due to high computational requirements and energy consumption, leading to sample loss and conflict with the need for low power consumption in wearable devices.
Innovation Solution
Implementing an incremental first Neural Network within the smart sensor's first electronic unit, which applies convolutional filters to each new sample, and an early exit module to trigger transmission to a second electronic unit only when predefined conditions are met, reducing computational load and energy consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a CNN network is implemented in the first electronic unit for processing detected samples, then calculation accuracy is improved, but the processing time exceeds the sampling interval causing sample loss
Solution Approach 1:
The patent divides the CNN processing into two segments: a first electronic unit with limited resources that performs initial processing and filtering of detected samples, and a second electronic unit with higher computational capacity that performs the full CNN analysis only on selected samples. This segmentation allows accurate CNN processing without requiring every sample to be processed in real-time, preventing sample loss while maintaining calculation accuracy.
2Productivity
If the second electronic unit is kept active to process all transmitted samples, then processing capability is improved, but energy consumption increases significantly
Solution Approach 1:
The patent implements periodic action by having the second electronic unit remain in a low-power state and only activate when the first electronic unit transmits samples that meet predefined conditions. This allows the system to maintain high processing capability when needed while dramatically reducing energy consumption during normal operation by keeping the high-performance unit dormant.
Solution Approach 2:
The first electronic unit acts as an intermediary between the detector and the second electronic unit. It filters and pre-processes samples, transmitting only those meeting predefined conditions to the second electronic unit. This intermediary role reduces the burden on the second electronic unit, allowing it to remain less active and consume less energy while still maintaining processing capability.
3Use of energy by moving object
If the first electronic unit performs simple calculations only, then energy consumption is reduced, but computational capability becomes insufficient for complex AI algorithms
Solution Approach 1:
The patent segments computational tasks between two electronic units with different capabilities. The first electronic unit handles energy-efficient simple calculations and filtering, while the second electronic unit provides the computational power needed for complex AI algorithms. This segmentation allows the system to maintain low overall energy consumption while still possessing the adaptability to perform complex computations when required.
Data Source
Figure 1
Figure 2
AI summary
Smart sensor (200) comprising: - a detector (210) configured for continuously detecting a physical quantity, sampled with a predefined sampling time tc, thus obtaining a plurality of the detected physical quantity samples xt-i with i∈[0, T-1] of size Nin where • xt-i ∈ RNin with i∈[0, T-1] is the sample acquired at discrete time t-i and is a vector of size Nin, • T is the number of samples acquired in a predetermine time size T x tc of an observation window; - an embedded first electronic unit (220) configured for receiving the acquired samples xt-i with i∈ [0, T-1] and for executing, each time a new sample xt-i with i∈ [0, T-1] is received, an incremental Neural Network g(incr) (·) having such a received new sample xt-i with i∈ [0, T-1] as an input, wherein such an incremental Neural Network g(incr) (·) provides the application of a number of n convolutional filters to each sample xt-i obtaining a corresponding first output output matrix Y having elements yt−ijkwith j∈ [1, n] and k ∈ [1, Nin] according to the following equation: yt−ijk=yt−i−1jk+cij×xt−i ∀j ∈ [1, n] and ∀k ∈ [1, Nin] wherein cij is the weight of the j-th filter applied to the sample at time t-i, said first output matrix Y having size n x Nin and being initialized to 0, and wherein the embedded first electronic unit (220) is configured to transmit the first output matrix Y to a second electronic unit (300) if a predefined condition is satisfied.