Classification of rock fragments produced by tunnel boring machine using convolutional neural networks

Highlights

• A convolutional neural network based method for classifying rock fragments of a boring tunnel is proposed.

• Rock fragment images were all captured at the construction site.

• AlexNet is modified for training and testing on a set of 11,640 images.

• The classifier can achieve 93.8% accuracy and 0.989 as the area under the receiver operator curve.

• The proposed method can effectively identify three classes of rock fragment images.

Abstract

Rock fragmentation in tunnel boring machine (TBM) construction is an important indicator of tunnelling state. Large fragments often correspond to an unhealthy tunnelling state and may cause physical damage to belt conveyors. In this work, a convolutional neural network (CNN), AlexNet, was modified to perform image classification based on the volume of the largest fragment. First, all images were captured at the tunnel construction site. Second, a classification scheme was proposed according to the geological state of the site. After pre-processing raw images, different categories of images were labelled manually to train and evaluate CNN models. Finally, an optimised parameter setting for rock fragment recognition was carried out. Experimental results showed that the extracted features are reliable and can distinguish different types of images well. This competitive performance demonstrates that the proposed methodology can accurately detect large rock fragments by images at the site.

Introduction

As an important monitoring item for tunnel constructors, fragments cut by tunnel boring machine (TBM) can directly reflect the fragmentation of rock mass during excavation. The degree of rock fragmentation plays an essential role in controlling and reducing the overall production cost. Rock fragmentation is related to many factors, including rock mass properties, TBM specifications (such as TBM diameter, cutter number, cutter spacing, cutter size, and shape), and TBM performance parameters (such as thrust force, torque, revolution per minute, and penetration rate). Therefore, it is invaluable to find the connection between rock fragmentation and these factors. Previous studies have shown how rock mass [1], cutter head principal parameters [2], and penetration rate per revolution [3] influence rock fragmentation, and how to identify rock mass rating based on TBM performance parameters and rock fragmentation [4].

Rock crushing during TBM construction is mainly reflected in large rock fragments. The appearance of large rock fragments can indicate the size or shape of muck is nonuniform, reflecting the unreasonable setting of performance parameters. In addition, problems with the conveyor belt mainly occur due to the size or shape nonuniformity of fragments [5]. Therefore, in the excavation process, the monitoring personnel often needed to uninterruptedly observe the fragments falling on the belt conveyor, pick out the large fragments and adjust the excavation parameters in time. For medium-sized fragments that will not damage the belt, it is necessary to adjust tunnel parameters properly. This method is widely used by engineers to arrive at an approximation [6]. However, its long-time operation and failure to collect essential information make the process cumbersome and inefficient, let alone dangerous.

To improve this situation, many studies have been carried out to determine the size distribution of rock fragments. The conventional approach is the sieving analysis in which a group of screens with different mesh sizes are used [7]. Although it is arguably the most accurate technique among others, it is impractical because it is very time-consuming and needs many resources [8]. Moreover, in TBM construction, it is difficult to achieve fast calculation of rock fragmentation on a belt conveyor and timely treatment.

To solve this problem, many researchers have studied the methods based on digital image processing technology. In digital image analysis, the most difficult step to measure the size of rock fragments is image segmentation [9]. Extensive studies have been carried out on this topic. The existing traditional segmentation approaches for rock fragment images could be summarised as (1) thresholding [10,11], (2) edge detection [[12], [13], [14], [15], [16], [17]], (3) region growing or split & merge [[18], [19], [20]], and (4) morphology [8,9,21].

However, due to the characteristics of rock fragments, some drawbacks can be expected when using traditional image segmentation algorithms:

• Nonuniform lighting, shadows, noise (mud, water, or other fine materials), overlapping, and blocking of fragments and the great range in fragment sizes make delineation extremely difficult using simple edge-detection;

• Background and foreground are uneven when the amount of rock and water on the belt conveyor changes for which a simple thresholding algorithm cannot be applied to segment;

• Each rock fragment may possess a textured surface and multiple faces, which often causes over-segmentation while smaller adjacent rock fragments can also be grouped into larger fragments;

• There is a problem with correctly extracting 3-D information from 2-D images. The fragment sizes in the third dimension need to be assumed [12].

Deep learning methods, especially convolutional neural networks (CNN), have dominated the field of computer vision in recent years. The convolutional kernels of CNN can automatically extract deep features instead of relying on manual feature extraction [68]. After LeNet-5 [22,23], many CNN structures have been proposed in which AlexNet [24] has eight layers, VGG [25] has 16-layer and 19-layer models, and in Google Inception [[26], [27], [28], [29]], the inception module was proposed to solve the problems of excessive training parameters, over-fitting, and gradient dispersion caused by very deep networks. Inspired by the success of CNN for image classification tasks, researchers have adapted CNN to semantic segmentation for feature extraction [30,31]. This method realises the ‘pixel-level’ classification of images and provides a new idea for segmentation. However, this method makes ‘pixel-level’ labelling and manual editing necessary to calculate fragment size distribution. These steps greatly increase the complexity of work. Because of the excellent ability of deep CNN to extract high-level features, many studies directly use it as a feature extractor. Then, a fully connected layer or other classifiers, such as a support vector machine, are used to complete the final classification at ‘image-level’ [[32], [33], [34], [35], [36]].

In this work, the deep CNN was used to automatically classify rock fragment images captured by our timing capture camera mounted vertically above the belt conveyor. Recognising the grading range of the maximum fragment volume in the image is required to determine whether manual picking or adjustment of the tunnelling parameters. Fig. 1 summarises our method flow for rock fragments classification. Section 2 summarises the related research in the image-based rock fragment inspection domain. Section 3 presents the acquisition of rock fragment images and the establishment of the dataset; Section 4 demonstrates the architecture of the CNN model and the training details of the proposed algorithm; Section 5 shows classification performance of the CNN model and the details of high-level features extracted from the target dataset; The results are discussed in Section 6; Section 7 concludes the article.