TL;DR
This paper introduces an improved entropy-weighted TOPSIS method for a multi-source data fusion evaluation system, designed to tackle data scale inconsistencies in industrial settings. By incorporating a dynamic fusion strategy and an evaluation system, the approach effectively filters out underperforming data sources, leading to more robust and accurate decision-level fusion for intelligent manufacturing.
What Problem Does This Paper Address?
The rapid advancement of industrial internet technology necessitates digital transformation in the manufacturing sector. A key enabler for this transformation is multi-source data fusion, which aims to combine information from various sources to gain a more comprehensive understanding. However, a significant challenge arises from data scale inconsistencies across different sources, making it difficult to fuse models and leading to low fusion accuracy. This paper addresses this by proposing a novel evaluation system and dynamic fusion strategy to improve decision-level fusion for industrial applications.
Methods
The core of this paper is an improved entropy-weighted TOPSIS method integrated into a decision-level fusion evaluation system. This system is designed to:
- Address Data Scale Inconsistency: The TOPSIS method, a multi-criteria decision-making technique, is adapted to evaluate and rank different data sources and models.
- Dynamic Fusion Strategy: The system dynamically assesses the contribution of each data source and model. If a source consistently underperforms (contribution below a threshold for two consecutive iterations), it is removed from the fusion process. This iterative filtering helps to eliminate noisy or irrelevant data.
- Feature Extraction and Fusion: For single-source data, the paper proposes novel feature extraction layers: a Max-AVG pooling layer and a residual dilated convolution layer. These are combined with a transfer learning VGG19 multilayer feature fusion network to handle small sample sizes and improve robustness.
- Decision-Level Fusion: For multi-source data, the paper utilizes an AdaBoost algorithm for decision-level fusion, enhanced by the proposed evaluation system.
Datasets
The experiments were conducted using a custom dataset collected from a tool wear diagnosis scenario. This dataset comprises seven data sources:
- Milling force signals (X, Y, Z axes)
- Vibration signals (X, Y, Z axes)
- Surface knife mark images
The data was prepared by categorizing tool wear into four states: initial wear, normal wear, rapid wear, and abnormal vibration. The knife mark images were collected with a specific resolution (200x200 pixels).
Key Insights and Ideas
- Dynamic Filtering is Crucial: The most significant insight is the effectiveness of dynamically removing underperforming data sources. This "pruning" approach is more efficient and robust than attempting to fuse all available data, especially when dealing with noisy or inconsistent industrial data.
- TOPSIS for Data Source Evaluation: Applying TOPSIS not just for model evaluation but also for selecting and weighting data sources is a novel application that proves effective in improving fusion performance.
- VGG19 as a Robust Backbone: Leveraging a pre-trained VGG19 network, with modifications to its later layers and the addition of custom feature extraction modules, provides a strong foundation for handling complex industrial data, particularly in scenarios with limited samples.
- Feature Extraction Matters for Image Data: The proposed Max-AVG pooling and residual dilated convolution layers demonstrate improved performance in extracting relevant features from surface images, outperforming a baseline VGG19 network.
Key Results
- Improved Accuracy: The proposed multi-source data fusion method achieved a test accuracy of 99.2% (±0.5% variation) in the tool wear status diagnosis task, outperforming single-source models.
- Enhanced Stability: The dynamic fusion strategy significantly improved the stability of the fusion model, especially in the later stages of training, compared to simpler fusion approaches.
- Reduced Model Complexity: By simplifying the network structure and avoiding overly complex feature extraction layers for certain data types (like raw signals), the proposed method reduces training difficulty and deployment overhead.
- Comparative Performance: The proposed method outperformed other mainstream multi-source data fusion algorithms (Uncertainty Weighting, Dynamic Weight Averaging, Dynamic Evidential Fusion) in terms of accuracy and average inference time, especially when considering the collaborative architecture.
Limitations and Future Work
- Focus on Algorithmic Performance: The paper primarily focuses on the algorithmic aspects of data fusion and acknowledges that further research is needed on raw data distribution and quality improvement during data collection.
- Generalizability: While validated on tool wear diagnosis, the authors suggest that future work should explore the application of this method to other industrial scenarios.
- Hyperparameter Sensitivity: Although the dynamic strategy aims to reduce sensitivity, the performance of the fusion system might still be influenced by hyperparameter choices within the TOPSIS and AdaBoost components.
Summary updated