Image-based thickener mud layer height prediction with attention mechanism-based CNN.

Area of Science:

• Industrial Process Monitoring
• Machine Learning Applications
• Image Analysis

Background:

• Real-time detection of mud layer height in thickeners is critical for process control but challenging with existing methods.
• Existing techniques lack the precision and speed required for effective industrial monitoring.

Purpose of the Study:

• To develop a soft sensor model for real-time mud layer height prediction.
• To enhance the accuracy and efficiency of mud layer height detection in industrial thickeners.

Main Methods:

• Proposed an end-to-end mud layer height prediction method utilizing a convolutional neural network (CNN).
• Integrated sequential attention mechanisms, including a novel regional spatial attention, to extract dynamic features from image samples.
• Employed a regional spatial attention mechanism that assigns higher weights to important feature regions within the spatial feature map.

Main Results:

• The proposed CNN with channel and regional spatial attention mechanisms improved both prediction precision and calculation speed.
• The method effectively avoids the loss of channel or spatial attention information.
• Experimental validation on a thickener mud layer dataset demonstrated the feasibility and effectiveness of the prediction method.

Conclusions:

• The attention mechanism-based CNN offers a viable solution for real-time mud layer height monitoring in industrial thickeners.
• The developed method enhances process control by providing accurate and rapid detection of a key quality index.