Research On Fault Diagnosis Of Bolt Loosening Based On Convolutional Neural Network

Bolted connection has the advantages of high reliability,easy installation and maintenance,and strong bearing capacity.It is widely used in engineering fields such as aerospace,civil engineering and bridges.However,because the bolted connection structure usually works under time-varying loads,the bolts in service are often loosened,resulting in connection failure.If the bolt loosening cannot be diagnosed in a timely and effective manner,it will often cause loss of equipment accuracy,economic losses,and even casualties.Therefore,effective detection and quantification of bolt looseness has important engineering significance.Traditional mechanical fault diagnosis methods focus on manual design and extraction of features.Researchers need to have certain prior knowledge,and the entire diagnosis process is inefficient.The convolutional neural network can learn from the fault data autonomously and mine useful features to improve the diagnosis accuracy,and realize the end-to-end bolt loosening fault diagnosis.Aiming at the problem of large sample size and low accuracy required for the application of convolutional neural network to bolt loosening fault diagnosis,a bolt loosening fault diagnosis model based on time-frequency analysis and deep migration learning is established.First,analyze ten typical time-frequency analysis methods;secondly,use the fine-tuned pre-training network AlexNet for training,analyze and compare the feature expression capabilities of each time-frequency method;finally,in order to better understand the network learning process,The activation state of neurons and the extracted features are visualized.In order to observe the model’s ability to extract features layer by layer,t-SNE is used to visualize the test samples and the features extracted after the first layer of convolution,the second layer of convolution,and the Softmax layer.Since noise will seriously affect the final diagnosis accuracy of the model,a bolt looseness fault diagnosis model(TSCNN)based on one-dimensional(1D)and two-dimensional(2D)two-stream convolutional neural networks is established.First of all,the traditional LeNet-5 network is improved,by changing the size and number of convolution kernels,increasing the convolutional layer and pooling layer,using batch normalization and dropout operations,respectively,to establish a one-dimensional convolutional neural network(1DCNN)And two-dimensional convolutional neural network(2DCNN);secondly,by analyzing the influence of the first-layer wide convolution kernel,time-frequency image size and network depth on the diagnosis results,the structural parameters of 1DCNN and 2DCNN are determined;finally,the two The depth-invariant features extracted by the network are combined with two streams to improve the accuracy of the bolt looseness fault diagnosis model.Finally,the effectiveness of the two methods was verified on three experimental platforms:single-bolt steel plate overlap,two-bolt steel plate overlap,and multi-bolt machine tool rails.Compared with traditional methods,the methods based on time-frequency analysis and deep transfer learning have higher recognition accuracy under a small sample of 2000.The accuracy of TSCNN on the three experimental platforms is 100%,99.58%and 99.58%,respectively.In addition,the diagnostic performance of the network under different noise environments is tested without any noise reduction preprocessing.The TSCNN model shows good robustness and high recognition accuracy under different noise environments.