Future Fault Classification Using Sequential Neural Subnetworks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning techniques struggle to efficiently predict and classify future faults in complex industrial systems due to the high dimensionality and imbalance of multidimensional time-series data, leading to poor generalization and increased complexity in diagnosing rare or simultaneous faults.
Innovation Solution
A data-driven approach using a neural network with a specific structure that transforms multidimensional time-series data into an embedded space, allowing for feature learning and representation, which captures the dynamics of the system, enabling accurate prediction and classification of future faults by reducing dimensionality and improving fault separation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional machine learning techniques are used to analyze multidimensional time-series data, then the system can process data at high pace and volume, but the prediction accuracy deteriorates due to high dimensionality and data imbalance
Solution Approach 1:
The patent applies dimensionality reduction techniques (PCA, t-SNE, autoencoders) to transform high-dimensional multidimensional time-series data into lower-dimensional representations. This resolves the contradiction by reducing the data dimensionality while preserving essential fault-related patterns, thereby maintaining prediction accuracy without sacrificing data processing capability.
Solution Approach 2:
The patent implements data augmentation and synthetic sample generation (SMOTE, GANs) before the main prediction task. This preliminary action addresses class imbalance by creating additional minority class samples, improving prediction accuracy for rare faults while maintaining the system's ability to process high-volume data.
2Loss of time
If the prediction horizon is extended to predict future faults further in advance, then the useful prediction time increases, but the prediction accuracy deteriorates due to compounded errors and更难 to learn precursor dynamics
Solution Approach 1:
The patent divides the prediction horizon into multiple sequential time steps or stages. Instead of predicting all the way to the future fault directly, the model breaks down the prediction into intermediate steps, reducing error accumulation and making it easier to learn precursor dynamics at each stage, thus maintaining accuracy over extended prediction horizons.
Solution Approach 2:
The patent introduces intermediate latent representations or embedding layers that capture system state evolution over time. These intermediaries serve as mediators between current observations and future fault predictions, preserving critical dynamics information while extending the prediction horizon without significant accuracy loss.
3Adaptability or versatility
If multiple faults are considered simultaneously in the prediction model, then the comprehensive diagnostic capability improves, but the model complexity increases due to combinatorial number of fault combinations
Solution Approach 1:
The patent employs multi-output neural networks or ensemble models that handle multiple fault predictions separately but simultaneously. Each fault type is predicted through dedicated output nodes or sub-models, avoiding the need to model all possible fault combinations explicitly, thus reducing model complexity while maintaining comprehensive diagnostic capability.
Solution Approach 2:
The patent designs a universal prediction framework that can handle multiple fault types through a single model architecture. The model learns shared patterns across different fault types while maintaining separate prediction capabilities, achieving high adaptability without proportional increases in model complexity.
4Extent of automation
If the system operates autonomously with minimum human intervention, then the operational efficiency improves, but the difficulty of detecting and measuring faults increases due to subtle precursor dynamics in high-dimensional data
Solution Approach 1:
The patent replaces manual fault detection and analysis with automated machine learning models that process multidimensional time-series data. These models automatically learn and detect subtle precursor dynamics patterns that would be difficult for humans to identify, enabling autonomous operation while maintaining high fault detection capability through computational rather than human analysis.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
A system evaluates a plurality of faults in an operation of a machine at a set of future instances of time. The system uses a neural network including a first subnetwork sequentially connected with a sequence of second subnetworks for each of the future instance of time such that an output of one subnetwork is an input to a subsequent subnetwork. The first subnetwork accepts the current time-series data and the current setpoints of operation of the machine. Each of the second subnetworks accepts the output of a preceding subnetwork, an internal state of the preceding subnetwork, and a future setpoint for a corresponding future instance of time. Each of the second subnetworks outputs an individual prediction of each fault of a plurality of faults at the corresponding future instance of time.