Research On Fault Pattern Recognition Of Mechanical System Key Components Based On Deep Learning

In order to meet the production capacity requirements,under the pressure of large-scale and integrated industrial production,machinery and equipment often work in heavy-load,high-speed,high-temperature and high-pressure conditions without break.In such working condition,key components such as bearings are prone to fatigue,accumulation of tiny lesions,performance degradation,finally lead to failure of the entire mechanical system.Therefore,in order to ensure the reliable and safe operation of the mechanical system,it is important to identify the running health status of the bearing.The traditional fault pattern recognition method artificially analyzes the time domain,frequency domain or timefrequency domain through signal processing,but due to factors such as mechanical system structure and interference components,the speed and effectiveness of the analysis have certain limitations.Intelligent failure pattern recognition methods can overcome this problem to a certain extent.However,most of the existing research on intelligent fault pattern recognition is based on the assumption that the fault has already occurred,and the detection task often focuses on the accurate identification of a single fault.Possible situations such as multiple failures on the same component or sudden failures during operation are ignored to a certain extent.In view of the limitations of traditional signal processing methods and common intelligent fault pattern recognition methods for special scenarios,under the support of the National Natural Science Foundation etc.The key components of rotating machinery are the main research objects.Based on the deep learning theory,a learning model of component fault characteristics is constructed,and based on this,the problems of compound fault pattern recognition and real-time fault detection are studied.First,in order to study the composite failure mode recognition problem of rolling bearings,a recognition model based on an improved Convolutional Deep Belief Networks(CDBN)is proposed.The proposed model utilizes the hierarchical network to extract the deep features of the data,and effectively utilizes the features of each layer to realize the composite failure mode recognition of the bearing.First,the band-pass filtering method is added in the preprocessing stage to enhance the signal;secondly,the Adam optimizer is used in the training stage to speed up the training process;finally,the multi-layer features are combined with the double-layer softmax,so that the features of each layer of the model can be fully obtained.use.In the experimental analysis of the identification of compound failure modes,the proposed model has achieved ideal results for the identification of single and compound failure modes in the dataset.Comparative experiments show that the improved CDBN model has higher classification accuracy than the standard CDBN network.Further,compared with sparse autoencoders,artificial neural networks and deep belief networks,the training error of the method proposed in this paper has a smoother descent process and a higher diagnostic accuracy,which fully proves the superiority of the proposed model.Subsequently,for the real-time detection and identification of mechanical fault modes,this paper proposes a sudden failure identification model based on the Gated Recurrent Unit(GRU)model.The model comprehensively considers a single failure mode classification and remaining life prediction problem,focusing on the diagnostic goal to track the equipment operating state along the time axis and achieve fast failure mode identification.The proposed scheme attempts to address the interpretability of the network and achieve a balance between trend prediction and classification.This process is similar to the voice activation mode of standby equipment.The proposed model is designed to monitor the equipment status.When a sudden failure occurs,the unidirectional GRU network in the model structure is activated accordingly to achieve rapid identification of the failure mode.After activating the network layer,a bidirectional long short term network(LSTM)structure is constructed to detect faults.The segment is processed to ensure that the distinct features of that segment are effectively learned for recognition.The experimental results and comparative results further confirm that the proposed network can solve the tasks of health status monitoring and fault identification.In summary,this paper proposes a CDBN-based fault pattern recognition model to introduce signal enhancement methods in the signal processing stage,use a deep convolutional confidence network for feature learning,and introduce feature fusion in the classification stage to achieve effective diagnosis.In this paper,a data processing method for sudden faults is proposed,and the state data is detected hierarchically by using a recurrent neural network-based sudden fault detection model,so as to achieve the purpose of real-time fault monitoring and classification.