Renewable Asset Health Scoring for Multi-Class Failure Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems face challenges in accurately predicting component failures in renewable energy assets, such as wind turbines and solar panels, due to overwhelming sensor data and the need for skilled personnel to extract meaningful features, leading to reactive maintenance and inefficiencies.
Innovation Solution
A method using deep neural networks with fully connected, convolutional, and recurrent layers to train failure prediction models from historical sensor data, allowing for multi-class classifications and improved accuracy by selecting models based on confusion matrices and positive prediction values, and generating alerts for potential failures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional sensor data analysis methods are used, then skilled personnel can extract meaningful features, but the process is tedious and high in errors
Solution Approach 1:
The patent replaces manual feature extraction (mechanical human analysis) with deep neural network-based automated analysis. The system uses fully connected layers, convolutional layers, and recurrent layers to automatically extract meaningful features from sensor data, eliminating the need for skilled personnel to manually process data while improving accuracy and reducing errors.
Solution Approach 2:
The system enables self-service by allowing the deep neural network to autonomously extract features and predict failures without human intervention. The model automatically processes sensor data, identifies patterns, and generates failure predictions, making the system independent of skilled personnel for the core analytical tasks.
2Measurement precision
If deep neural networks with multiple layers are used, then predictive accuracy is enhanced, but computational complexity increases
Solution Approach 1:
The patent segments the deep neural network into distinct functional layers: fully connected layers for feature integration, convolutional layers for spatial pattern recognition, and recurrent layers for temporal sequence analysis. This segmentation allows each layer to specialize in specific aspects of data processing, improving overall accuracy while making the complex model more manageable through modular architecture.
Solution Approach 2:
The system processes sensor data across multiple dimensions by incorporating both spatial relationships (through convolutional layers) and temporal sequences (through recurrent layers). This multi-dimensional approach enables the model to capture complex patterns in the data that single-dimensional methods would miss, significantly enhancing predictive accuracy.
3Loss of time
If multi-class classifications with different lead times are used, then proactive maintenance planning is enabled, but model training complexity increases
Solution Approach 1:
The patent implements preliminary action by training the model to predict failures at multiple lead times (different classes representing different time horizons). This allows maintenance to be planned proactively before failures occur, with the model providing predictions at various time intervals to enable appropriate scheduling of maintenance activities.
Solution Approach 2:
The system changes the parameter of lead time by creating multiple classification categories representing different time horizons (e.g., short-term, medium-term, long-term predictions). This parameter transformation enables the model to provide flexible predictions at various time scales, supporting different maintenance planning scenarios while managing complexity through structured parameterization.
Data Source
AI summary
An example method comprises receiving historical wind turbine failure data and asset data from SCADA systems, receiving first historical sensor data, determining healthy assets of the renewable energy assets by comparing signals to known healthy operating signals, training at least one machine learning model to indicate assets that may potentially fail and to a second set of assets that are operating within a healthy threshold, receiving first current sensor data of a second time period, applying a machine learning model to the current sensor data to generate a first failure prediction a failure and generate a list of assets that are operating within a healthy threshold, comparing the first failure prediction to a trigger criteria, generating and transmitting a first alert if comparing the first failure prediction to the trigger criteria indicates a failure prediction, and updating a list of assets to perform surveillance if within a healthy threshold.