Neural Network Training Data Expansion for Transient Engine States
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models for estimating transient engine operating states face challenges in data expansion, as methods like rotating and reversing images are not applicable to time series data, limiting the amount of training data and potentially leading to bias or decreased precision.
Innovation Solution
A method involving acquiring short-period time series data, expanding it by extracting multiple sets with different phases and adding noise, and shifting measurement points to generate diverse training data, which increases the dataset size and improves generalization ability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional data expansion methods (rotating and reversing images) are used for neural network training, then the amount of training data can be expanded by several multiples, but these methods are not applicable to time series data of engine operating states
Solution Approach 1:
The patent changes the parameters of time series data by adding noise with different characteristics (white noise, pink noise, brown noise) and applying signal processing transformations. This allows expansion of training data while maintaining temporal relationships, resolving the contradiction between data quantity expansion and method applicability to time series data.
Solution Approach 2:
The patent introduces a new dimension of data expansion by generating multiple phases of time series data and combining them. Instead of spatial transformations like image rotation, it uses temporal phase shifting and noise addition in the time domain, making data expansion applicable to time series data while increasing training data quantity.
2Productivity
If a neural network model is trained with limited training data, then the training process can be completed faster, but the model may become overtrained and lose generalization ability
Solution Approach 1:
The patent performs preliminary data expansion by generating expanded training data with noise and phase variations before model training. This prepares a more diverse training dataset in advance, enabling the model to learn better generalization patterns without requiring excessively long training times, thus balancing training speed and generalization ability.
Solution Approach 2:
The patent creates multiple copies of the original time series data with different noise additions and phase shifts. These copied and transformed datasets serve as expanded training examples, allowing the model to generalize better from limited original data without proportionally increasing training time.
3Quantity of substance
If the period of training data is shorter than the operation period of the neural network model, then data expansion is possible, but the training data must be extracted and reconstructed to match the model's operation period
Solution Approach 1:
The patent segments the time series data into multiple phases and extracts training data at different time points within each phase. By dividing the data extraction process into discrete phases and time points, it enables systematic data expansion while managing the complexity of matching different periods through structured sampling.
Data Source
AI summary
A method for generating a neural network model, the method includes acquiring first time series data having a first period that is shorter than an operation period of the neural network model; extracting, from the first time series data, a plurality of sets of second time series data having a second period that is longer than the first period; and executing training on the neural network model using training data that include the plurality of sets of second time series data.