Resolver Signal Processing With ML Demodulation for In-Band Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Microcontrollers lack sufficient memory and processing capacity to implement complex machine learning and artificial intelligence algorithms for accurately and timely resolving the angle and velocity of rotating elements, particularly in the presence of in-band noise, leading to imprecise control and significant delay in resolver systems.
Innovation Solution
Implementing a hardware delta-sigma analog-to-digital converter, filters, and rectifiers to preprocess resolver signals, followed by a trained machine learning model executed by a digital signal processor, which reduces signal complexity and enables efficient noise removal and angle determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If tracking algorithms (e.g., Luenberger-Observer) are used to resolve the angle from sine and cosine signals, then the microcontroller can operate with limited memory and processing capacity, but the system experiences significant delay in resolving the angle during rapid frequency changes and cannot effectively remove in-band noise
Solution Approach 1:
The system performs preliminary filtering and rectification of the sine and cosine signals using dedicated hardware filters and rectifiers before feeding them to the machine learning model. This preprocessing reduces the complexity of the input signals, enabling the microcontroller to execute the ML model efficiently while achieving rapid angle resolution without the delays associated with traditional tracking algorithms.
Solution Approach 2:
The patent replaces traditional mechanical tracking algorithms (software-based iterative approaches) with a machine learning model that has been trained offline. This substitution enables the system to resolve angles rapidly during runtime without the computational burden of iterative algorithms, while the hardware filters and rectifiers handle signal conditioning, dividing the workload effectively.
2Measurement precision
If complex machine learning models are implemented to remove in-band noise and rapidly resolve the angle, then measurement precision and speed improve, but the microcontroller lacks sufficient memory and processing capacity to execute these models
Solution Approach 1:
The system segments the signal processing task into distinct functional blocks: hardware filters for noise filtering, rectifiers for signal conditioning, and a machine learning model for angle and velocity resolution. This segmentation allows each component to be optimized independently, with the ML model receiving pre-processed, lower-complexity inputs that reduce its computational requirements.
Solution Approach 2:
The patent changes the parameters of the input signals to the machine learning model by applying filtering and rectification that reduce their complexity. The filtered and rectified signals have reduced dynamic range and lower frequency content, which allows the ML model to operate with fewer computational resources while maintaining high measurement precision.
3Object-affected harmful factors
If traditional hardware filters are used to remove noise from resolver signals, then noise filtering is achieved, but the system cannot effectively handle in-band noise and the hardware complexity and cost increase
Solution Approach 1:
The system introduces rectifiers as intermediary components between the filters and the machine learning model. The rectifiers convert the filtered sine and cosine signals into a form that is more suitable for ML processing, enabling the model to effectively distinguish between noise and valid signal components. This intermediary step enhances noise rejection capabilities without requiring complex hardware filters.
Solution Approach 2:
The patent replaces complex hardware filtering solutions with a combination of simple hardware filters and a machine learning-based noise removal approach. The ML model, trained to recognize and remove in-band noise, substitutes for the need for complex analog filters, reducing hardware complexity and cost while improving noise rejection effectiveness.
Data Source
AI summary
Processing circuitry comprising an analog-to-digital converter configured to perform the techniques of this disclosure. The analog-to-digital converter may obtain, based on electrical interactions with a resolver sensor, a noisy modulated digital cosine signal (NMDCS) and a noisy modulated digital sine signal (NMDSS). The processing circuitry includes a filter configured to process the NMDCS and the NMDSS to output a filtered NMDCS and a filtered NMDSS, and a rectifier configured to process the filtered NMDCS and the filtered NMDSS to output a rectified NMDCS and a rectified NMDSS. The processing circuitry includes a digital signal processor that executes a first trained machine learning model to process the rectified NMDCS and the rectified NMDSS to obtain a demodulated DCS and a demodulated DSS, and compute, based on the demodulated DCS and the demodulated DSS, an approximate angle of the rotating element.