Training Data Generator Using Synthetic System Plans for Image Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The lack of a standardized form of exchange for technical system plans and the absence of sufficient training data hinder the digitization and maintenance of technical systems, necessitating a high detection accuracy in machine learning methods.
Innovation Solution
A training data generator that extracts symbols from digital system plans, applies positioning rules, and generates synthetic system plans to create a sufficient amount of training data for image detection modules, using a random selection and augmentation techniques to enhance realism.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised machine learning methods are used for digitization of system plans, then detection accuracy can be improved, but a very large amount of annotated training data is required which is not guaranteed to be available
Solution Approach 1:
The patent creates synthetic copies of system plan symbols through a generator that produces artificial training data. Instead of relying solely on scarce real annotated plans, the system generates synthetic symbol images with corresponding annotations by programmatically creating symbol representations, adding noise and transformations, and extracting features to build training datasets without requiring additional manual annotation effort
Solution Approach 2:
The patent performs preliminary data preparation by pre-defining symbol libraries, annotation schemas, and generation rules before training begins. The system pre-generates large quantities of synthetic training data with embedded annotations in advance, so that when actual training occurs, sufficient labeled data is already available to train the detection model effectively without needing to manually annotate real plans
2Measurement precision
If a large number of annotated plan examples are used for training, then high detection accuracy is achieved, but the complexity and effort of manual annotation increases
Solution Approach 1:
The system makes the training data self-annotating by automatically generating both the symbol images and their corresponding ground truth annotations through the synthesis process. The generator inherently knows the correct annotations since it creates the symbols programmatically, eliminating the need for human annotators to manually label each training example while still providing accurate annotation data for training
3Quantity of substance
If synthetic system plans are generated to create training data, then the amount of training data increases, but ensuring realism and variety of generated examples becomes challenging
Solution Approach 1:
The patent introduces dynamic variability into the synthetic data generation process by implementing random transformations including geometric transformations (rotation, scaling, flipping), noise addition, distortion effects, and random cropping. These dynamic modifications ensure that generated training examples exhibit diverse real-world variations rather than identical repetitive patterns, improving the model's ability to handle varied actual plans
Solution Approach 2:
The system changes multiple parameters during synthesis including symbol sizes, positions, orientations, noise levels, transformation angles, and other visual characteristics. By varying these parameters randomly across different generated examples, the system creates a diverse set of training data that maintains realism while providing sufficient quantity for effective model training
Data Source
AI summary
A training data generator includes an interface to read symbols extracted from digital system plans for technical systems, wherein each of the system plans represents the design and/or the functionality of a technical system by of symbols, and the system plans are identical. The training data generator also includes a storage module which is designed to store the extracted symbols, a selection module which is designed to randomly select a symbol sub-quantity of the stored symbols using a random generator, a generator which is designed to generate at least one synthetic system plan on the basis of the selected symbol sub-quantity, and an output module to output the at least one synthetic system plan as training data in order to train a trainable image detection module, wherein the trainable image detection module is designed to generate a digital system plan using an analog system plan of a technical system.