AI Control for Automation Systems With Few-Shot Fault Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern automation systems with AI for quality control face challenges in detecting new faults that were not included in the training data, as they require extensive data sets which are difficult to obtain due to low defect rates in production, especially in small-batch production scenarios.
Innovation Solution
A computer-implemented method for training an artificial neural network using a pre-provided database with a selection and training procedure that creates minibatches of classified data records, allowing for efficient learning and adaptation to new fault detection with minimal additional training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional AI training with extensive data sets is used, then fault detection reliability is improved, but training time and data acquisition effort increase significantly
Solution Approach 1:
The system performs preliminary action by pre-training the AI model with extensively labeled training data before deployment. This preliminary training establishes a strong baseline model that can quickly adapt to new faults with minimal additional data, resolving the contradiction by preparing the system in advance rather than requiring extensive training when faults occur
Solution Approach 2:
The training process is segmented into two distinct phases: (1) offline pre-training with extensively labeled data to build foundational knowledge, and (2) online fine-tuning with minimal new data to adapt to specific faults. This segmentation allows the system to achieve high reliability through comprehensive initial training while minimizing the time and data required for subsequent fault-specific adaptation
2Measurement precision
If extensively labeled training data is collected for all possible faults, then AI system accuracy is improved, but data availability decreases due to low defect rates
Solution Approach 1:
The system performs preliminary action by pre-collecting and labeling extensively training data for common faults during normal production operations. This preliminary data collection builds a comprehensive database that would otherwise require impossibly long time periods to accumulate, enabling high accuracy without requiring rare fault data to be available in real-time
Solution Approach 2:
The system creates synthetic copies of fault data through data augmentation techniques and uses transfer learning to copy knowledge from related domains. This allows the AI model to learn from abundant synthetic or transferred data rather than requiring scarce real fault data, maintaining high accuracy while working around data availability constraints
3Adaptability or versatility
If the AI system is retrained with new fault data, then detection capability for new faults is improved, but system downtime and production loss increase
Solution Approach 1:
The training process is segmented into offline pre-training (performed during non-production time) and online fine-tuning (performed rapidly with minimal data). This segmentation allows the bulk of adaptation work to be done when production is not affected, while limiting production disruption to only the brief fine-tuning phase
Solution Approach 2:
The system performs preliminary action by pre-training the model with extensively labeled data before deployment, so that when new faults occur, only minimal fine-tuning is needed. This preliminary preparation dramatically reduces the retraining time required in production, maintaining both adaptability and productivity
4Measurement precision
If manual labeling of training data is performed, then data quality is improved, but labeling effort and cost increase
Solution Approach 1:
The system applies self-service by using the AI model to automatically label its own training data through self-supervised learning and pseudo-labeling techniques. The model generates its own training labels from unlabeled production data, eliminating the need for extensive manual labeling while maintaining sufficient data quality for effective training
Solution Approach 2:
The system applies partial manual labeling only to critical or ambiguous cases, while using automated methods for the majority of data labeling. This partial approach maintains data quality for important examples while dramatically reducing overall labeling effort compared to complete manual labeling
Data Source
AI summary
A control device of an automation system, which is configured to control a plant, such as a production plant, including using an AI system, is provided. In an application of the control device, the device monitors the production with regard to the quality of the objects produced, for example, with regard to the presence of fault cases. The AI system is trained in advance based on a plurality of known states of the objects, so that the AI system may be trained for the occurrence of new, previously unknown states, where only a small number of example cases are required.