Adaptive selection of machine learning/deep learning model with optimal hyper-parameters for anomaly detection of connected chillers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Chillers in HVAC systems often experience unexpected faults leading to unplanned shutdowns, causing efficiency losses and damage to other equipment, due to various influencing factors, making it challenging to predict chiller failures effectively.
Innovation Solution
A chiller model management system that uses machine learning and deep learning to analyze data from connected sensors, generating prediction models by evaluating performance using techniques like receiver operating characteristics and probability density functions, and selecting the most accurate models for fault prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple machine learning/deep learning models are trained with different hyperparameters to improve prediction accuracy, then the reliability of fault prediction improves, but the device complexity and computational resources required increase
Solution Approach 1:
The system pre-trains multiple prediction models with different hyperparameters before deployment. These models are evaluated using cross-validation techniques to determine their predictive performance in advance, allowing the system to select the optimal model before actual fault prediction is needed, thereby improving reliability while managing complexity through preliminary preparation
Solution Approach 2:
The system varies hyperparameters (such as learning rate, batch size, number of epochs, network architecture parameters) across multiple model training iterations. By systematically changing these parameters and evaluating each configuration's performance through cross-validation, the system identifies optimal parameter settings that maximize prediction accuracy without requiring excessive complexity in the final deployed model
2Measurement precision
If cross-validation techniques are used to evaluate model performance and select optimal models, then the measurement precision of model performance assessment improves, but the loss of time and computational resources increases
Solution Approach 1:
The system performs cross-validation with a controlled number of folds and iterations, using sufficient validation to achieve reliable performance measurement but limiting the extent to prevent excessive time consumption. The validation process is configured to balance between obtaining precise performance metrics and completing evaluation within acceptable timeframes for model selection
Solution Approach 2:
The system conducts preliminary hyperparameter tuning and model selection using cross-validation before deploying models to production. By performing this intensive evaluation process in advance during the model development phase, the system achieves precise performance measurement without incurring continuous time losses during operational fault prediction
3Adaptability or versatility
If the system continuously updates fault predictions using selected models for each epoch, then the adaptability of the prediction system improves, but the use of energy and computational resources increases
Solution Approach 1:
The system updates fault predictions at specific intervals (epochs) rather than continuously for every incoming data point. By periodic re-evaluation using the selected model, the system maintains adaptability to changing conditions while significantly reducing computational energy consumption compared to continuous real-time prediction updates
Solution Approach 2:
The system uses the selected prediction model to autonomously generate fault predictions based on incoming chiller data without requiring continuous external computational resources. Once the model is selected and deployed, it self-servingly processes data and generates predictions, reducing the energy burden on external computing infrastructure
Data Source
AI summary
A model management system for a building, including one or more memory devices and one or more processors. The one or more memory devices are configured to store instructions to be executed on the one or more processors. The one or more processors are configured to determine whether chiller fault data exists in chiller data used to generate a plurality of chiller shutdown prediction models. The one or more processors are further configured to generate a first performance evaluation value for each of the plurality of chiller shutdown prediction models using a first evaluation technique in response to a determination that chiller fault data exists in the chiller data, and generate a second performance evaluation value for each of the plurality of chiller shutdown prediction models using a second evaluation technique in response to a determination that chiller fault data does not exist in the chiller data. The one or more processors are configured to select one of the plurality of chiller shutdown prediction models based on the first performance evaluation in response to the determination that chiller fault data exists in the chiller data, and select one of the plurality of chiller shutdown prediction models based on the second performance evaluation in response to the determination that chiller fault data does not exist in the chiller data.