Research On Mechanical Fault Diagnosis Method Based On Attention-guided Multifunctional CNN

With the development of information technology and the progress of the times,China’s modern industrial system is gradually moving towards complexity,and the structure of various mechanical equipment is becoming more and more integrated.Rotating machinery,as an important part of mechanical equipment,is widely used in gas turbine engines,wind turbines,aviation engines and other fields.Due to the long-term operation under complex and unfavorable conditions,it is highly susceptible to failure.In addition,the integrated system presents the characteristics of multi-source coupling,which makes the large volume of data processing and condition monitoring technology urgent to be solved.The demand for fault diagnosis under the background of "big data" exceeds the capabilities of traditional fault diagnosis technologies.As such,intelligent fault diagnosis technology based on deep learning has gained significant attention due to its ability to tackle the challenges posed by the transformation and upgrading of rotating machinery and equipment.However,most of the intelligent fault diagnosis methods only learn some features of the input signal to achieve fault classification,ignoring the impact of working condition changes on the whole system.Therefore,this paper will conduct a deep learning-based fault diagnosis study of bearings and gears with key rotating components in rotating machinery.This paper is dedicated to mining more complementary feature information through multifunctional convolutional neural networks in a parallel learning framework to improve the learning ability and diagnostic accuracy of the network.The main studies of this paper are as follows:(1)To address the problem of poor fault diagnosis accuracy due to background noise in one-dimensional vibration signals,a network model is constructed based on attention mechanism and multi-task learning theory.It consists of a global feature sharing network and special task networks with feature attention modules,allowing for mutual learning and parallel work between different network branches.One of the special task sub-networks is a signal denoising network,which uses coding-decoding modules and wavelet noise reduction to improve the signal-to-noise ratio of the input signal.The other sub-network is a fault diagnosis network used to identify fault classes.The feasibility of this approach is verified by conducting experiments on an open source dataset and noisy environment.(2)Aiming at the problem that the slope parameters of the Leaky Re Lu function and the weight coefficients of the integrated loss function in the PLA-CNN network are difficult to select,a improved fault diagnosis method of PLA-CNN based on the Sparrow Search Algorithm(SSA)is proposed.The initial integrated loss function in the training process of PLA-CNN model is selected as the fitness function of the optimization algorithm,and the built-in parameters of the network are automatically selected by using SSA,which avoids the blindness and uncertainty caused by the empirical presetting of parameter combinations in the original model,and the superiority of the optimized model is demonstrated by experiments.(3)To address the challenge of complex coupling of vibration signals and the difficulty in diagnosing bearing faults under variable working conditions,a novel attention-guided multifunctional parallel learning convolutional neural network model is presented.The model expands the initial approach to enable simultaneous signal noise reduction,fault classification,and working condition recognition,thus providing a comprehensive solution to this problem.The networks at different functional levels learn from each other to extract more complementary feature information,resulting in improved integrated learning capability and diagnostic accuracy of the model.In order to enhance the robustness of the model,a multi attention module is designed and introduced,and the effectiveness of the method is verified in the experiments of multi-classification of variable working conditions.