Machine Condition Monitoring With Adaptive Multi-Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for monitoring machine operating states in industrial settings, such as hydraulic systems, face challenges due to varying dynamics and production cycles, leading to unreliable models when a fixed training period is used, which does not adapt to different operating states and can result in complex settings.
Innovation Solution
A method that dynamically adapts by detecting operating parameters during reference operations, forming data blocks across different time periods, training models based on these blocks, and establishing an error limit for each model, allowing for continuous expansion and updating of the model set without a fixed training period, using machine learning techniques like autoencoders or PCA for dimensional reduction and anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a fixed training period is used for machine learning models, then the model training process is simplified and has clear boundaries, but the model reliability deteriorates because it cannot adapt to varying machine dynamics and production cycles
Solution Approach 1:
The patent implements a dynamic training period determination mechanism that automatically adapts the training duration based on the actual operating conditions and data characteristics. Instead of using a fixed predetermined period, the system dynamically adjusts when to stop training and create a new model by evaluating data block characteristics and model performance metrics in real-time, allowing the training process to adapt to varying machine dynamics and production cycles
2Ease of operation
If a fixed training period is used, then the training process is easier to manage, but the adaptability to different operating states deteriorates
Solution Approach 1:
The system implements a feedback mechanism where the training process continuously monitors model performance and data characteristics, then adjusts the training period accordingly. The feedback loop evaluates whether the current data block represents a new operating state by comparing against existing models, and automatically determines when to terminate training and create a new model, enabling adaptive response to different operating states while maintaining automated management
3Reliability
If continuously new models are created from newly acquired data, then the model stays updated with current operations, but the computational resources and system complexity increase
Solution Approach 1:
Instead of continuously creating new models from all incoming data, the system applies partial action by selectively triggering model creation only when necessary. It evaluates each data block to determine if it represents a significant change in operating state, and only creates new models when the feedback mechanism indicates a need, thus avoiding unnecessary computational overhead while maintaining model currency
4Adaptability or versatility
If the training period is extended to cover more operating states, then the model coverage improves, but the training time and computational cost increase
Solution Approach 1:
The patent segments the training process into discrete data blocks that are evaluated independently. Instead of training on all available data in one long continuous period, the system divides the data stream into manageable blocks and processes them sequentially, creating multiple specialized models for different operating states rather than one large comprehensive model, thus reducing individual training times while improving overall coverage
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method for monitoring the operating states of a machine, comprising: acquiring at least one operating parameter of the machine to obtain reference operating data of at least two different reference states; forming several data blocks from the reference operating data, each extending over data within different first time intervals; training a first model for operating parameters based on a first data block and determining an error limit of the first model; applying the first model to a second data block and, if an error of the first model when applied to the second data block is greater than the error limit of the first model, training a second model for operating parameters based on the second data block and determining an error limit of the second model; defining a model set comprising the first and the second models.Apply all models of the model set to another data block and, if an error of at least one of the models from the model set is greater than the error limit of the respective model when applied to the other data block, train another model for operating parameters based on the other data block, determine an error limit of the other model and add the other model to the model set; repeat the previous step for all other data blocks.