Sensor Time-Series Augmentation for Imbalanced ML Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to effectively classify time series of sensor data due to class imbalances, particularly in manufacturing processes where defective products are underrepresented, leading to poor training results and missed anomalies.
Innovation Solution
A method for generating additional training examples by distorting time and/or value axes of existing time series using randomly parameterized weights, allowing for more balanced training datasets without requiring complex generative models like GANs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If GANs are used to generate synthetic training time series, then the model can learn from generated data, but the computational complexity increases and the interpretability decreases
Solution Approach 1:
The patent creates synthetic training time series by copying and transforming existing real training examples through geometric transformations (scaling, shifting, rotating in time-frequency space). This approach generates additional training data without requiring complex generative models like GANs, thereby maintaining computational efficiency and interpretability while still providing diverse training samples for the machine learning model
Solution Approach 2:
Instead of using expensive and complex GAN architectures, the patent employs simple, computationally inexpensive transformations of existing data. These transformations can be applied rapidly and discarded after use, providing a cost-effective alternative that generates sufficient training diversity without the overhead of training and maintaining complex generative models
2Quantity of substance
If GANs are used to generate synthetic time series, then data generation is possible, but the smoothness and realism of generated data becomes difficult to achieve
Solution Approach 1:
The patent applies geometric transformations to existing training time series, including scaling in time and amplitude domains, shifting, and rotating in time-frequency representation. These parameter changes create varied synthetic examples that maintain the underlying physical characteristics and smoothness of the original data while providing sufficient diversity for robust model training
3Manufacturing precision
If extensive hyperparameter tuning is performed on GANs to achieve realism, then data quality improves, but training time and computational resources increase
Solution Approach 1:
The patent performs data augmentation by applying transformations to existing training data before the main model training begins. This preliminary preparation of diverse training examples eliminates the need for iterative hyperparameter tuning during model training, saving significant computational time and resources while ensuring high-quality, realistic training data is available from the start
Data Source
Figure 1
Figure 2
Figure 3
AI summary
According to various embodiments, a method for training a machine learning model to classify time series of sensor data is described, comprising, for each training example of a set of training examples, generating at least one additional training example by generating, for each additional training example, a sequence of weights which has a weight for each time index of a sequence of time indices of the training example, weights for each time index except the last time index of the sequence of time indices, a time increment to the time index following the time index with the weight that the sequence of weights has for the time index, and/or weights for each time index of a value of the training example for the time index, and training the machine learning model using a training data set which contains at least the additional training examples.