Test Measurement Neural Networks With De-Noised Training and Noise Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning algorithms for test and measurement systems, such as oscilloscopes, require large and costly data sets for training, which can lead to overfitting due to noise inclusion, resulting in poor generalization and increased training time.
Innovation Solution
Separate noise components from training waveforms to create noiseless data sets, train the ML algorithm with these, and compensate for noise in the operational environment, using techniques like waveform averaging and noise characterization to improve generalization and avoid overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If noise is included in training waveforms to make the model more realistic, then the model's ability to handle real-world data improves, but the model risks overfitting to noise characteristics and loses generalization capability
Solution Approach 1:
The patent extracts and removes noise components from training waveforms to create clean training data. By separating the signal from noise and eliminating the noise portion, the model trains on pure signal characteristics without learning noise patterns, thus preventing overfitting while maintaining generalization ability.
Solution Approach 2:
The patent performs noise removal as a preliminary step before training the model. By pre-processing the training data to eliminate noise components beforehand, the model is exposed only to clean signal patterns from the start, preventing noise-related overfitting before it can occur during training.
2Measurement precision
If large training data sets with multiple sweep parameters are used to improve model accuracy, then measurement precision improves, but training time and development cost increase significantly
Solution Approach 1:
The patent extracts only the essential signal features from training waveforms by removing noise components. This creates a more compact and efficient training dataset that maintains predictive power without requiring excessive data volume, thereby reducing training time while preserving accuracy.
Solution Approach 2:
The patent changes the quality parameter of training data from noisy to noiseless by removing noise components. This transformation creates higher-quality training examples that require fewer samples to achieve the same level of model accuracy, thus reducing the total training time and computational resources needed.
3Adaptability or versatility
If multiple DUTs and instruments are used to generate comprehensive training data covering different noise levels, then model adaptability improves, but device complexity and training cost increase
Solution Approach 1:
The patent extracts and removes noise components from training data regardless of their source. By eliminating noise rather than attempting to capture all possible noise variations through multiple DUTs and instruments, the system achieves robustness to different noise conditions using a simpler, more scalable approach.
Solution Approach 2:
Instead of the conventional approach of adding noise variations to training data to improve adaptability, the patent inverts the approach by removing noise completely. This counterintuitive strategy achieves better generalization across different noise conditions by training on clean signals, allowing the model to focus on learning fundamental signal characteristics rather than memorizing noise patterns.
Data Source
AI summary
A test and measurement system has one or more inputs connectable to a device under test (DUT), and one or more processors configured to execute code that causes the one or more processors to: gather a set of training waveforms by acquiring one or more waveforms from one or more DUTs or from simulated waveforms, remove noise from the set of training waveforms to produce a set of noiseless training waveforms, and use the set of noiseless training waveforms as a training set to train a neural network to predict a measurement value for a DUT, producing a trained neural network. A method of training a neural network having receiving one or more waveforms from one or more DUTs, or generating one or more waveforms from a waveform simulator, removing noise from a set of training waveforms gathered from the one or more waveforms to produce a set of noiseless training waveforms, and use the set of noiseless training waveforms as a training set to train a neural network to predict a measurement value for a DUT, producing a trained neural network.