Research On Intelligent Diagnosis Methods Of Mechanical Faults Driven By Sample Reliability

Mechanical equipment is an important tool of modern industrial production,whose reliability is related to the safety of people and property at the work site.With the development of intelligent manufacturing concepts,it is especially important to use automated and intelligent methods to monitor and diagnose the health status of mechanical equipment.However,in real industrial scenarios,such phenomena as poor signal quality,wrong labels,and unbalanced number of categories are common in mechanical fault diagnosis tasks.These problems reduce the reliability of the intelligent diagnosis model to a great extent and become the key pain point restricting the effect of intelligent diagnosis.However,the influence of sample quality on diagnostic reliability is rarely considered in the existing research on mechanical intelligent diagnosis.On this basis,this study uses sample reliability to describe the quality levels of samples,combining convolutional neural network,generative adversarial network,graph neural network and other artificial intelligence algorithms to carry out in-depth research on mechanical intelligent diagnosis technology.The main contents are as follows:(1)Aiming at the problems of uneven sample quality and difficult evaluation of sample reliability for mechanical equipment in real industrial scenarios,a data preprocessing method based on sample reliability assessment was proposed.The proposed method considers several requirements for sample reliability assessment,uses the validation loss to quantify the impact of sample quality on model recognition performance,introduces the Influence Function to accelerate the evaluation process of samples,and the efficient calculation of sample reliability value is realized.(2)Aiming at the problems that the sample enhancement task is difficult to conduct and its enhancement effect is poor under the unreliable sample labels,a sample enhancement method based on a reliability screening strategy was proposed.The proposed method can filter out wrong labels with the reliability assessment values(RAV)of training samples,and optimize the training process of the sample enhancement model,so as to improve the quality of the generated sample images and the reliability of the subsequent fault diagnosis.(3)Aiming at the problems that the training process of the fault diagnosis model is not stable and the key features are difficult to extract under the unreliable signal quality,a fault diagnosis method based on a reliability weighting strategy was proposed.The proposed method uses RAVs to construct the sample weight.Class weight and early stopping were also introduced to optimize the procedure of feature extraction.It can effectively reduce the negative influence of the low-quality sample signals on the fault diagnosis model.(4)Aiming at the problems of low accuracy and poor robustness of the fault diagnosis model under unreliable signal quality,a fault diagnosis method based on the dynamicweighted graph convolutional network was proposed,combining sample reliability and sample association information.This method optimizes input graph construction based on sample reliability and proposes a dynamic-weighted graph updating strategy to better mine the correlation between sample nodes,which improves the robustness of the fault diagnosis model.(5)Aiming at the problems of insufficient multi-sensor feature fusion and low model accuracy for fault diagnosis under unreliable signal quality,a multi-sensor multi-head graph attention network(MMHGAT)model was presented,combining sample reliability and sample association information.The feature fusion of the proposed method does not only exist in graph construction,but uses the proposed dynamic fusion process that can dynamically adjust the feature fusion and feature learning process,according to the actual diagnosis effect.Finally,the research contents of this study were summarized.The main innovation points of this study were presented.And the future research was prospected.