Radar Object Classification via Synthetic Signal Rotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems for radars face challenges in achieving accurate object classification due to angular biases and varying signal quality, which are exacerbated by the need for large and varied training samples that are not always practically scalable, leading to inferior prediction capabilities.
Innovation Solution
The method involves generating synthetic training samples by manipulating real data signals to simulate rotations and creating composite samples that eliminate angular biases and signal strength variations, allowing for improved prediction accuracy by training the system with a larger, more diverse dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training samples are collected from all possible angles to remove angular bias, then prediction accuracy is improved, but the complexity and cost of data collection becomes prohibitively high
Solution Approach 1:
The patent uses signal processing techniques to create virtual copies of radar targets at different angular positions by manipulating the phase and amplitude of received signals. Instead of physically moving the radar or target to collect data from all angles, the system generates synthetic training samples that simulate multi-angle observations, thereby achieving angular diversity without the complexity of physical data collection from all possible angles.
2Measurement precision
If a large number of diverse training samples are collected to improve classification accuracy, then prediction capability is enhanced, but the time and resources required for data collection and processing increase significantly
Solution Approach 1:
The patent performs preliminary signal processing and feature extraction during the data collection phase, pre-computing angular information and signal characteristics that will be needed for training. By preparing and organizing training samples with pre-extracted features and metadata before the actual training process, the system reduces the computational burden during training and accelerates the overall development timeline.
Solution Approach 2:
The system generates synthetic training samples by copying and transforming existing radar signals through phase rotation and amplitude modulation techniques. This allows the creation of large numbers of diverse training samples from a relatively small set of actual measurements, significantly reducing the time and resources needed for data collection while maintaining sample diversity for effective training.
3Ease of manufacture
If training is performed with limited angular coverage to reduce data collection complexity, then data collection becomes more practical, but angular bias degrades prediction accuracy
Solution Approach 1:
The patent applies phase rotation techniques to copy the signal characteristics observed at limited angles and synthesize what the signals would look like from other angular positions. By manipulating the phase of received signals according to geometric relationships, the system generates virtual training samples that represent targets at angles not physically observed, thereby eliminating angular bias while maintaining practical data collection constraints.
Solution Approach 2:
The system changes the angular parameter of training samples through signal processing transformations. By applying phase shifts and amplitude adjustments to the received signals, the system effectively transforms the angular parameter of the training data, allowing samples collected at one angle to be converted into equivalent samples for other angles, thus achieving angular coverage without physical repositioning.
Data Source
AI summary
Techniques of machine learning of a radar are disclosed, where the radar has a plurality of antennas that are arranged on an antenna array. In an example, a method of machine learning includes obtaining a real training sample from a first real target in field of view of the radar, where the real training sample includes a plurality of first real data signals, where each of the first real data signals are obtained from a corresponding antenna from amongst the plurality of antennas. The method further includes deriving a synthetic training sample by manipulating the plurality of first real data signals to simulate a rotation of the first real target about a pre-determined axis of the antenna array.