Asynchronous Feature Extraction Circuit for Low-Power IoT Signal Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional feature extraction methods in intelligent chips require converting time domain signals to frequency domain, resulting in high power consumption, especially when dealing with asynchronous pulse signals from LC-ADCs, which complicates direct information extraction from these signals.
Innovation Solution
A circuit with feature extracting units that directly extract and classify the instant range of change (IROC) feature from asynchronous pulse coding in the time domain, using pulse request and direction signals, thereby avoiding the need for frequency domain conversion and reducing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional synchronous data processing method with frequency domain conversion is used for feature extraction, then feature extraction can be performed, but power consumption increases
Solution Approach 1:
The patent inverts the traditional feature extraction approach by performing feature extraction directly in the time domain using asynchronous pulse signals, rather than converting to frequency domain. The asynchronous neural network processes events as they occur in time, eliminating the need for FFT and synchronous sampling, thereby achieving feature extraction with significantly reduced power consumption while maintaining accuracy
Solution Approach 2:
The patent replaces the mechanical synchronous data processing system with an asynchronous event-driven system. Instead of using traditional synchronous circuits that require clock signals and sequential processing, the system uses asynchronous pulse coding and event-based neural networks that process data only when changes occur, substituting the rigid mechanical processing paradigm with a more flexible and energy-efficient approach
2Loss of information
If frequency domain conversion is performed for feature extraction, then feature information can be obtained, but additional power consumption overhead is introduced
Solution Approach 1:
The patent extracts the essential feature information directly from the time domain signal using asynchronous pulse coding, taking out only the necessary information (instant range of change) without performing full frequency domain conversion. This selective extraction approach maintains information completeness while avoiding the energy overhead of complete FFT processing
Solution Approach 2:
The patent applies partial action by performing only the necessary time domain feature extraction (instant range of change calculation) without executing the complete frequency domain conversion process. This partial processing approach achieves sufficient feature extraction for neural network input while consuming significantly less power than full frequency domain analysis
3Use of energy by moving object
If LC-ADC asynchronous pulse coding is used for signal sampling, then ultra-low power consumption is achieved, but direct feature extraction from the pulse signals becomes complex
Solution Approach 1:
The asynchronous neural network is designed to naturally process LC-ADC pulse outputs without requiring complex preprocessing circuits. The network's event-driven architecture inherently handles the asynchronous pulse coding format, with neurons firing based on incoming pulse timing and direction signals, thereby simplifying the overall feature extraction system while maintaining ultra-low power operation
Data Source
AI summary
The present application discloses a circuit for extracting a feature, a network and a signal processing system. The circuit includes: one or more instant range of change feature extracting units, which are connected in parallel and configured to extract and classify an instant range of change (IROC) feature from an input of an asynchronous pulse coding in the time domain, the input of the asynchronous pulse coding is a pulse request signal and a pulse direction signal obtained by performing time-domain quantization and coding on an analog signal. By the circuit for extracting a feature according to the present application, the instant range of change of an analog input of the asynchronous pulse coding can be directly extracted and classified in the time domain, converting from the time domain to the frequency domain required by the traditional feature extraction process is avoided, and the power consumption overhead is reduced.