Belt Tearing Detection Method Based On Deep Convolution Generation Adversarial Network

Belt conveyor is a very important tool required for underground mining in coal mines,and it shoulders a very important responsibility in safe production.In the production process,due to the harsh working environment,the frequent occurrence of sharp objects such as gangue mixed with angles and fine rods,the belts are often accompanied by scratches,cracks and even serious breakages such as tears,which can lead to a huge threat to the safety of workers and property in the mines.For this reason,the study of how to quickly and accurately detect breakage of belts on belt conveyors is a topical research issue of current relevance.However,existing detection methods,for belt breakage images,can only accurately detect one type of breakage,which is not conducive to later belt protection and detection.Most image-based detection methods require pre-processing operations such as binarization,edge extraction and image denoising,which can result in lengthy computation time.In recent years,with the advent of deep learning,its advantages in training networks to extract target features using large-scale data.The Deep Convolutional Generative Adversarial Network(DCGAN)is one of the deep learning models.When DCGAN and its improved algorithms are applied to belt damage detection,the use of batch normalization for up-sampling feature extraction makes the pixel space in the generator unevenly covered and prone to artefacts,leading to bias in the features learned by the generator,which in turn affects the accurate detection of belt damage types.Secondly,the discriminator mainly uses a binary sigmoid function,which can only output two categories and cannot identify three types of damage.In addition,the discriminator and generator typically use the same learning rate,and during training the generator is updated several times while the discriminator is updated only once,which allows the discriminator to reach a local optimum prematurely,leading to pattern collapse.Furthermore,due to the large number of network layers in the generator and discriminator models,the generator and discriminator go through multiple convolution layers,which may make the learned features partially missing,resulting in incomplete features and affecting the accurate detection of belt tears.To address the above problems,this paper proposes two methods of belt damage detection based on deep convolution generative adversarial networks and conditional deep convolution generative adversarial networks,with the following main contributions.(1)For the belt damage detection method of the deep convolution generative adversarial network,1)In response to the use of batch normalization in the generator,artifacts are easily generated in the generated conveyor belt images,which affects the accurate detection of conveyor belt damage,and batch normalization tends to cause lengthy computation time and memory usage,this paper removes the batch normalization from the generator,which not only improves the accuracy of breakage detection,but also reduces the training time of the network.2)To address the fact that the discriminator of deep convolution generative adversarial networks cannot be used to identify multiple object types,this paper uses a multi-class softmax function for the output of the discriminator to output a vector of category probabilities to accurately classify the scratches,scrapes and tears that occur in conveyor belt damage.3)To address the problem that the discriminator and the generator use the same learning rate,which makes the model prone to collapse,this paper introduces two time-scale update rule,in which the generator and the discriminator use different learning rates and are updated in accordance with the scale,which can not only maintain the counterbalance between the generator and the discriminator,but also improve the training speed of the discriminator and make the conveyor belt damage detection better in real time.(2)For the belt damage detection method of the conditional deep convolution generative adversarial network,in order to avoid incomplete features learned by the generator and discriminator,the generator and discriminator models are connected by skip-layer during training in this paper,which not only improves the convergence speed,but also avoids partial loss of features,thus improving detection accuracy.