IoT Vibration Fault Detection Using Shared Encoder and Local Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Detecting faults in machines and systems using vibration measurement data is challenging due to the difficulty in distinguishing between normal and anomalous behaviors, and vibration sensors often produce inconsistent results across different devices, making it hard to create a model applicable to multiple systems.
Innovation Solution
A network architecture that leverages machine learning by training a device with a supervisory service to use encoder and decoder models for anomaly detection, where the encoder is trained using vibration data from multiple devices and the decoder evaluates structural and behavioral anomalies, with additional training at the fog layer to tailor models to specific systems, employing techniques like LSTM-based autoencoders and transfer learning for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If vibration measurement data is collected from multiple devices to train a model, then the model's applicability across different systems improves, but the computational load and training time increase significantly
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) training a general encoder on aggregated vibration data from multiple devices to learn universal fault patterns, and (2) training a device-specific decoder using only that device's own vibration data to adapt to individual characteristics. This segmentation allows the system to achieve broad applicability while minimizing per-device training time and computational requirements.
Solution Approach 2:
The patent performs preliminary action by pre-training the encoder on aggregated data from multiple devices before deployment. This pre-trained encoder captures general vibration patterns and fault characteristics across different device types, so that when a new device is deployed, only the decoder needs to be trained on that device's specific data, dramatically reducing the training time and computational load at deployment.
2Reliability
If vibration sensors are used to detect faults, then early fault detection is possible, but the sensors produce inconsistent results across different devices making model creation difficult
Solution Approach 1:
The patent merges vibration data from multiple different devices during the encoder training phase. By aggregating data across devices with potentially inconsistent sensor characteristics, the encoder learns to recognize fault patterns that are invariant to device-specific variations. This merging process enables the model to achieve reliable fault detection across diverse devices despite measurement inconsistencies.
Solution Approach 2:
The patent transforms the raw vibration measurements into a different parameter space through the encoder's feature extraction process. Instead of working directly with raw sensor data that has device-specific inconsistencies, the encoder converts these measurements into abstract feature representations that capture essential fault characteristics while being invariant to sensor variations, thereby improving measurement consistency across devices.
3Measurement precision
If a model is trained to assess vibration measurements from a specific system, then accuracy for that system improves, but the model is unlikely to be applicable to other systems
Solution Approach 1:
The patent segments the model into two functional components with different scopes: an encoder trained on aggregated data from multiple devices that provides generalizability, and a decoder trained on device-specific data that provides accuracy for that particular system. This segmentation allows each component to specialize in its intended function while working together to achieve both accuracy and applicability.
Solution Approach 2:
The patent creates a universal encoder that can process vibration data from any device type, making the core feature extraction capability applicable across multiple systems. This universal encoder is then combined with device-specific decoders, creating a multi-functional system that maintains high accuracy for each specific system while being applicable to others through the shared encoder component.
Data Source
AI summary
In one embodiment, a device in a network receives a machine learning encoder and decoder trained by a supervisory service. The service trains the encoder and decoder using vibration measurement data sent to the service by a plurality of devices. The device trains, based on the received encoder, a classifier to determine whether vibration measurement data is indicative of a behavioral anomaly. The device receives vibration measurement data captured by a particular set of one or more vibration sensors of a monitored system. The device evaluates, using the trained decoder, the received vibration measurement data to determine whether the data is indicative of a structural anomaly in the monitored system. The device evaluates, using the trained classifier, the received vibration measurement data to determine whether the data is indicative of a behavioral anomaly in the monitored system.