Research On Drainage Pipe Defect Recognition Method Based On Acoustic Features And Deep Neural Networ

Urban drainage pipe,as an important underground infrastructure to ensure the sustainable development of modern cities,bears the heavy burden of sewage and rainwater treatment.The advantages of environmental protection,energy saving and high efficiency have become the lifeline of the city.Therefore,intelligent detection and defect identification of drainage pipe are of great significance for accurately grasping safe operating conditions and avoiding catastrophic accidents and economic losses.Deep learning provides new approaches in low-frequency active acoustic signal representation,feature extraction,and defect identification,to avoid the application difficulties of acoustic wave modal mechanism and defect identification method for specific research objects.In various pipe operating conditions,there are complex working conditions and continuous multiple blockages.The difficulty lies in how to solve the defect recognition model with high recognition accuracy and good generalization performance when the defect training samples are insufficient or even missing through effective feature learning.Based on the great feature learning ability of deep neural network,in-depth research is conducted on multi-scale blockage feature extraction,individual acoustic signature identification,siphon multi-class defect identification and zero-sample defect identification method.There are four main aspects:(1)For the problem that convolutional neural defect identification methods with single scale and single label output are difficult to effectively separate and identify multiple blockages,an improved Multi-Scale Convolution Neural Network(MSCNN)method is proposed,which extracted transient features of acoustic signatures at different time scales in parallel by setting convolution kernels of different sizes.Using Sigmoid function as activation function,the improved classification layer can ensure that the probability of each type of defect output was independent from each other.The test sample was used to predict multiple label outputs,thereby realizing the separation and identification of multiple blockages.The method has an average recognition accuracy of 94.27% for multiple blockages,which is 6.73% higher than the traditional machine learning method of extreme learning machine,and 3.36%II higher than the single-scale and single-label CNN.Through the multi-label output mechanism,the separation and identification of multiple blockages in straight pipelines is realized,thus significantly improving the practicability and effectiveness of the multiple blockage identification method.(2)As for the problem that the generalization performance of the defect recognition model is degraded because the training sample set collected periodically cannot cover all the developing defect types,a multi-target individual recognition based on the acoustic signature model(AS model)and the Squeeze Net neural network method within pipeline was proposed.Through AS model,the typical common features of the same type of individuals(blockage,3-way piece,and pipeline tail)under multiple pipeline operating conditions were obtained for forming the individual acoustic signature feature maps.Squeeze Net neural network was used to automatically learn the representative features of individual acoustic signatures,thereby eliminating the dependence on manual feature extraction,reducing the weight parameters,simplifying the network structure,and improving the blockage recognition speed.The recognition accuracy rate of this method is 99.03% for individuals with known pipeline operating states,and 98.52% for individuals with unknown pipeline operating states.The identification of unknown pipeline operating states is no longer limited by the defect categories contained in the known training set,thus improving the generalization ability of the defect identification model.(3)Due to the special structure of siphon that undertakes the cushioning function,frequent and time-delayed multi-source defects lead to the failure of MSCNN defect recognition method to effectively extract the time-delay features of siphon multi-type defects.In response to this,a siphon multi-class defect recognition method based on MSCNN and LSTM neural network was proposed.Automatic end-to-end feature extraction from raw acoustic signatures by MSCNN and Long S hort-Term Memory(LSTM)neural network model,local fine-grained features and time-dependent coarse-grained contextual features were captured,thereby realizing the fusion of coarse-grained features and acoustic properties that comprehensively characterize siphon defects.The attention mechanism was introduced into adaptive r eorganization,and the features were screened to effectively describe the differences in the features of various defects and improve the recognition accuracy of multiple defects such as blockage,leakage and weld breakage.The average recognition accuracy of the method for siphon multi-class defects is 98.44%,which is 10.89% higher than the siphon multi-class defect recognition accuracy using only MSCNN model.Meanwhile,Gradient-Weighted Class Activation Map(Grad-CAM)was used to explore the interpretability of the model in multi-class defect recognition based on acoustic signals,which fully verified the effectiveness.(4)A pipeline zero-shot defect identification method based on Transformer neural network and attribute description was proposed for the problems of insufficient defect feature learning and high misdiagnosis rate for pipeline defect samples without historical records in practice.Transformer neural network extracted pipeline defect features of different shapes from the global view.The corr elation between known class defect and unknown class defect information was achieved by building an attribute description matrix to replace the labels of unknown class defects and sharing auxiliary information of attribute descriptions.Based on the CNN-based attribute learner,the unknown class defect features are mapped from features to attributes for attribute prediction.Through the similarity measure of attribute prediction and attribute description matrix,the task of zero-shot pipeline defect identification was completed.According to the experimental results,a pipeline zeroshot defect identification method based on Transformer neural network and attribute description has an average recognition accuracy of 84.71% for zero-sample defects,which provides a new idea for solving the problem of pipeline zero-sample defect recognition.