Three-Stage Vision Screens Coal Belt Objects in 2 s
Belt conveying is the main transport method in a coal mine, and large foreign objects mixed into the coal flow can tear the belt, block a transfer point and even damage equipment or cause an accident. This cell mounts an industrial camera above the belt and applies three processing stages: adaptive weighted multi-scale Retinex enhancement, improved multi-scale template matching that combines frame differencing with an area rule, and an MLP network optimised by a grey wolf algorithm. It recognises and triggers screening within a 2 second cycle, with field accuracy above 95 percent, subject to site acceptance results.
The image is already damaged before the model sees it
Underground illumination is low and dust is heavy, so captured images have poor contrast and lost shadow detail. Running recognition directly on that input leaves even a strong model guessing at a blurred picture. The first task is therefore not recognition but restoring the image to a readable state.
Why three stages instead of one large network
Each stage has a job: enhancement guarantees input quality; template matching screens suspected objects out of the moving coal flow; the network makes the final decision on candidates. A single network asked to do all three would need far more capacity and would still be limited by the input.
Template matching screens; the network decides; a grey wolf tunes
Frame differencing uses differences between consecutive frames to catch moving targets, and the area rule removes small scattered coal pieces by size. Together they extract suspected foreign objects from a moving scene and cut the downstream network load sharply, so the network only judges a small candidate set, where speed and accuracy are easier to guarantee.
The grey wolf algorithm is a bio-inspired optimiser used to tune the MLP parameters so the network separates large foreign objects more accurately. It does not change network structure; it changes the quality of the trained parameters, which is especially valuable when samples are limited.
Two seconds is set by the coal flow
The belt runs continuously, so the system must process and decide within the short time the coal flow crosses the field. The 2 second cycle keeps recognition and screening action aligned with the flow, so an object does not pass the detection zone unseen. A slower cycle makes even precise screening pointless, because the object is already downstream.
Recognition is only useful if the diverter fires
When a large object is recognised, the system outputs a signal that drives the screening mechanism to remove it while coal continues to flow. When nothing is found, the coal passes normally. That linkage closes the loop; recognition without actuation is just a report.
Accuracy is a function of conditions
The field accuracy above 95 percent was achieved under specific underground conditions and coal quality. Illumination, dust concentration, coal flow speed and the mix of object types all affect performance. The system comprises an industrial camera, supplementary lighting, an image-processing industrial computer and the screening mechanism, and the protection and real-time behaviour of every part must be checked against site conditions.
Questions about the underground setup
Which foreign objects are targeted?
Large objects mixed into the coal flow, such as metal pieces and bulky debris. They can tear the belt or block a transfer point, so they must be removed during transport.
What does Retinex enhancement solve?
Low illumination and heavy dust leave images with poor contrast. Adaptive weighted multi-scale Retinex restores detail and contrast, giving later matching and recognition usable input.
Why combine frame differencing with an area rule?
Frame differencing catches moving targets, and the area rule removes small scattered coal. Combined, they screen suspects in a moving flow and reduce the network burden.
How is a grey wolf optimiser different from ordinary training?
It tunes MLP parameters to improve discrimination of large foreign objects without changing network structure, so it is a training-level improvement rather than a new architecture.
Can it go straight onto an existing belt?
The system includes camera, lighting, industrial computer and screening mechanism, with camera and lighting protected for underground conditions and screening linked to results. Layout must be set against local illumination, dust and coal flow speed.