Semi-Servised Soft Sensor Ensemble for Limited Labeled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In industrial processes, it is challenging to measure key quality variables using hardware sensors due to harsh environments, leading to a scarcity of labeled data samples for soft sensor modeling, which limits the effectiveness of existing soft sensing methods.
Innovation Solution
An ensemble learning-based semi-supervised regression model is developed, utilizing both labeled and unlabeled data samples through the Bagging algorithm, Dissimilarity algorithm, and K-nearest neighborhood method to improve diversity and accuracy, selecting significant data samples for prediction, and employing Gaussian Process Regression for estimating missing quality variables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional soft sensing methods (PCR, PLS, ANN, SVM, GPR) are used, then prediction capability is provided, but performance is limited due to scarcity of labeled data samples
Solution Approach 1:
The dataset is segmented into labeled and unlabeled portions, with the unlabeled data further divided into multiple subsets through Bagging. This segmentation allows the model to utilize both labeled and unlabeled data effectively, improving prediction reliability without requiring more labeled samples.
Solution Approach 2:
The invention transitions from traditional supervised learning (using only labeled data) to semi-supervised learning by incorporating unlabeled data as an additional dimension. This dimensional expansion allows the model to leverage the structure and distribution information from unlabeled data, enhancing performance despite limited labeled samples.
2Measurement precision
If only labeled data samples are used for modeling, then model accuracy can be maintained, but the limited number of labeled samples restricts model performance
Solution Approach 1:
Before final model training, the invention performs preliminary processing on unlabeled data through Bagging to generate multiple subsets, and applies dissimilarity measurement to identify diverse representatives. This preliminary action prepares the unlabeled data to complement the limited labeled samples, expanding the effective training dataset while maintaining accuracy.
Solution Approach 2:
The invention introduces an intermediary selection mechanism that bridges labeled and unlabeled data. By measuring dissimilarity and selecting representative unlabeled samples that are diverse from labeled ones, the intermediary selection process ensures that unlabeled data complements rather than redundant the labeled data, maintaining prediction accuracy.
3Adaptability or versatility
If all unlabeled data samples are directly used in ensemble learning, then data utilization is maximized, but diversity performance and model accuracy deteriorate due to inclusion of useless samples
Solution Approach 1:
Instead of treating all unlabeled data uniformly, the invention applies local quality assessment by measuring the dissimilarity of each unlabeled sample against the labeled dataset. Samples with high dissimilarity (indicating they provide unique information) are selected, while redundant samples are discarded, ensuring high-quality data utilization.
Solution Approach 2:
The invention changes the selection parameter from simple random sampling to dissimilarity-based sampling. By using dissimilarity measurement as the selection criterion, the model dynamically identifies and selects unlabeled samples that maximize diversity and information content, improving both data utilization and model accuracy.
Data Source
AI summary
The present disclosure provides a novel advanced ensemble learning strategy for soft sensor development with semi-supervised model. The main target of the soft sensor is to improve the prediction performance with a limited number of labeled data samples, under the ensemble learning framework. Firstly, in order to improve the prediction accuracy of sub-models for ensemble modeling, a novel sample selection mechanism is established to select the most significantly estimated data samples. Secondly, the Bagging method is employed to both of the labeled and selected data-set, and the two different kinds of datasets are matched based on the Dissimilarity (DISSIM) algorithm. As a result, the proposed method guarantees the diversity and accuracy of the sub-models which are two significant issues of the ensemble learning. In this work, the soft sensor is constructed upon the Gaussian Process Regression (GPR) model.