VAE Continual Training for Sensor Drift and Fault Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning systems require extensive historical data to detect faults and are unable to identify new faults in real-time, struggle with sensor drifts, and fail to distinguish operational changes due to faulty conditions or load variations, necessitating improved methods for identifying sensor drifts and varying operational conditions.
Innovation Solution
Implementing continual training with variational autoencoders (VAE) to identify sensor drifts and operational conditions using a processor-based method that involves training generative models on current and previous datasets, leveraging rehearsal techniques and latent sequences to adapt to changing environments without forgetting past knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning systems use extensive historical data for fault detection, then detection accuracy is improved, but system latency increases and real-time detection capability deteriorates
Solution Approach 1:
The system performs preliminary training of VAE models on historical data in advance, creating multiple pre-trained models corresponding to different time windows. When real-time detection is needed, the pre-trained models are immediately deployed without requiring extensive real-time data accumulation, thus reducing latency while maintaining detection accuracy.
Solution Approach 2:
The historical data is divided into multiple time windows, with each window training a separate VAE model. This segmentation allows the system to use appropriate models for different detection scenarios and reduces the computational burden of processing entire historical datasets in real-time.
2Measurement precision
If traditional systems accumulate large amounts of historical data for training, then model accuracy is improved, but adaptability to new fault types deteriorates
Solution Approach 1:
The system dynamically creates new VAE models for new fault types by training on recent time window data. Instead of retraining on entire historical datasets, the system adapts by training new models on relevant recent data, enabling quick adaptation to new fault types while maintaining accuracy through the specialized nature of each model.
Solution Approach 2:
The system changes the training parameters (time window size, data selection) based on the specific fault type being detected. For new fault types, the system adjusts training parameters to focus on recent relevant data, enabling adaptability while maintaining model accuracy through parameter optimization.
3Stability of the object's composition
If traditional methods use fixed training datasets, then training stability is improved, but ability to distinguish operational changes deteriorates
Solution Approach 1:
The system periodically retrains VAE models using rolling time window data, creating a cycle of training and deployment. This periodic retraining with updated data allows the system to maintain training stability through structured cycles while improving its ability to distinguish operational changes through exposure to recent data patterns.
Solution Approach 2:
The training dataset dynamically evolves as time windows roll forward, incorporating new data and excluding old data. This dynamic approach maintains stability through the structured windowing mechanism while improving operational change detection through continuous exposure to current operational patterns.
4Adaptability or versatility
If extensive historical data is used for training, then comprehensive fault coverage is improved, but data processing complexity increases
Solution Approach 1:
The system segments historical data into multiple time windows, with each window training a dedicated VAE model. This segmentation reduces the complexity of processing entire datasets at once while maintaining comprehensive fault coverage through the collective capability of multiple specialized models.
Solution Approach 2:
Each VAE model is trained on specific local data characteristics from its corresponding time window, making each model specialized for particular operational conditions. This local quality approach reduces overall system complexity by distributing processing requirements across multiple simpler, specialized models rather than one complex general model.
Data Source
AI summary
Existing machine learning systems require historical data to perform analytics to detect faults in a machine and are unable to detect new types of faults/changes occurring in real time. These systems further fail to identify operation changes due to sensor drift and forget past events that have occurred. Present application provides systems and methods for identifying and classifying sensor drifts and diverse varying operational conditions from continually received sensor data using continual training of variational autoencoders (VAE) following drift specific characteristics, wherein sensor drift is compensated based on identified changes in sensors and degradation in machine(s). Rehearsal technique is performed by either VAE based generative models trained in previous iterations that are configured to generate a dataset corresponding to a current iteration, or discriminative instances of original dataset in previous iterations that are configured to generate a dataset corresponding to a current iteration, thus preventing from catastrophic forgetting.