Industrial Abnormality Detection via Deep Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional deep learning methods for abnormal detection in industrial systems require large amounts of labeled data, which are difficult to collect, leading to over-fitting and poor performance when applied to different machines due to data offset issues, especially with time series sensor data.
Innovation Solution
A method using deep transfer learning that trains a model with labeled source domain sensor data and unlabeled target domain sensor data, employing preprocessing, feature extraction layers, and discriminators to predict abnormalities across multiple machines with minimal data, allowing for real-time detection and redeployment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional deep learning methods are used for abnormal detection, then the model can be trained with sufficient labeled data, but the model becomes over-fitted and cannot be applied to different machines due to data offset
Solution Approach 1:
The patent introduces an intermediary domain adaptation layer that transforms source domain features into target domain features. This intermediary mechanism bridges the gap between different machine data distributions, allowing the model trained on source domain labeled data to be effectively applied to target domain machines without requiring extensive retraining or collection of target domain labeled data.
Solution Approach 2:
The patent dynamically adjusts domain-specific parameters and feature representations during the transfer learning process. By changing the parameter space to accommodate domain differences while preserving the core detection logic, the model maintains reliability across different machines despite data offset variations in sensor readings, operating conditions, and system characteristics.
2Measurement precision
If labeled data is collected from different machines, then the model can be trained accurately, but it takes a long time and high cost due to infrequent abnormality occurrence
Solution Approach 1:
The patent performs preliminary actions by collecting and storing unlabeled sensor data from target domain machines in advance. When abnormality detection is needed, the pre-collected data can be quickly processed using the transfer learning model, eliminating the need for time-consuming data collection during actual abnormal events. This preliminary data preparation significantly reduces the time loss associated with rare abnormality occurrences.
Solution Approach 2:
The patent creates a copied version of the source domain model and adapts it to the target domain through transfer learning. Instead of collecting extensive labeled data from each target machine, the model knowledge is copied and transferred from the source domain, where labeled data is abundant, to the target domain, achieving high detection accuracy without the time and cost of collecting equivalent labeled data from scratch.
3Adaptability or versatility
If extensive labeled data is collected from multiple machines, then the model can be trained to be universally applicable, but the complexity and cost of data collection and labeling increases significantly
Solution Approach 1:
The patent extracts only the essential domain-invariant features and detection patterns from the source domain labeled data, separating them from machine-specific variations. By extracting and transferring only the core abnormality detection knowledge rather than the entire dataset, the system achieves multi-machine applicability without requiring complex data collection infrastructure or extensive labeling efforts across multiple machines.
Solution Approach 2:
The patent creates a universal transfer learning framework that can be applied to multiple target machines using a single source domain model. This multi-functional approach allows the same trained model to detect abnormalities across different machine types and configurations, eliminating the need for separate data collection and training systems for each machine, thereby reducing overall system complexity and cost.
Data Source
AI summary
The present invention discloses a method for detecting abnormity in an unsupervised industrial system based on deep transfer learning. Labeled machine sensor sequence data from a source domain and unlabeled sensor sequence data from a target domain are used in the present invention to train an industrial system abnormal detection model with good generalization ability, and the industrial system abnormal detection model is trained and tested to finally generate a trained industrial system abnormity discrimination model. Using the model, received machine sensor sequence data can be analyzed and whether a machine is abnormal is discriminated.