Research On Intelligent Fault Diagnosis Of Electric Motor Based On Deep Learning

With the continuous development of global science and technology,electric motors have gradually become one of the most important equipment in various industrial fields,and are widely used in electric vehicles,CNC lathes,compressors,pumps and other fields,which is an indispensable part of the modern industrial system.As an important rotating part of the modern industrial system,electric motors often work in harsh environments and complex working conditions,and are permanently in high operating speeds and high load rates.Long term operation can cause fatigue damage,overheating,or insulation material damage inside electric motors,which leads to the decline in the mechanical properties of the important parts and serious wear and tear.Eventually,it will evolve step by step into more serious faults,leading to equipment damage and shutdown.Electric motor fautls bring challenges and threats to the safe and stable operation of the entire industrial production system and the personal safety of staff.Therefore,in order to ensure the safety and stability of electrical equipment,avoid the occurrence of major safety accidents and promote the sustainable development of the industrial field.It is of great significance to strengthen the research on the fault diagnosis technology of electrical equipment,and diagnose the fault location and type of electric motors timely and accurately.Traditional motor faults diagnosis methods mainly rely on the accuracy of modeling or rely on the experience of diagnostic experts to manually extract the fault features for judgment.These traditional methods are not only inefficient,but also difficult to find potential faults.Therefore,the development of automated and intelligent motor fault diagnosis technology has become an inevitable trend.Aiming at the above problems,this dissertation carries out a series of deep learning-based research on intelligent diagnosis of common mechanical and electrical faults of electric motors due to the impact on the performance of diagnostic models in the intelligent fault diagnosis of electric motors caused by the varying working conditions,limited fault data samples,environmental noise and different data sampling frequencies,etc.The main research work of this dissertation is as follows:1.A capsule network-based electric motor fault diagnosis model under varying operating conditions is proposed to address the impact of changing motor operating conditions on fault diagnosis performance.The proposed diagnosis model combines a wide convolutional kernel convolutional neural network(WDCNN)with a capsule network.The WDCNN network can increase the perception field of neural network for motor fault data and better capture the time-dependent characteristics of motor operation data.Unlike traditional scalar neurons,capsule networks transfer data features internally in the form of vector neurons,which can store more motor fault data features.By combining the advantages of the aforementioned two networks,the diagnostic performance of the model is ensured when the motor operating conditions change.Finally,a finite element simulation model of the motor was built on ANSYS Maxwell simulation software,and the motor fault data were simulated and collected to verify the validity of the proposed fault diagnosis model.The experimental results show that compared with other methods,the fault diagnosis method based on the combination of WDCNN and capsule network proposed in this dissertation can have better fault diagnosis performance and stronger generalization performance under varying operating conditions.2.Aiming at the problems of overfitting and performance degradation of existing fault diagnosis models in the case of changes in motor operating conditions and limited data samples.A few-shot learning based fault diagnosis model based on the combination of residual network and Gaussian process is proposed.Unlike traditional deep learning methods that require a large amount of data to train the diagnosis model,this dissertation proposes a fault diagnosis strategy with limited data based on meta-learning,which views fault diagnosis under different operating conditions as different few-shot classification tasks.Firstly,a deep residual network is applied to extract fault features from motor operation data;then,the encoded feature vectors are input into a Gaussian process network with kernel transfer capability for fault detection and classification of motors under different operating conditions.First,the performance of the proposed model in this dissertation is preliminarily verified on the simulation model-based dataset and an opensource dataset,respectively.Finally,in order to further validate the effectiveness of the proposed method on a real platform,a three-phase induction motor experimental platform was constructed for further verification.The results of the relevant experiments show that compared with traditional methods,the proposed method has better performance under different operating conditions with limited data samples.When conducting fault diagnosis tests under new operating conditions without participating in model training,the method proposed in this dissertation also shows better generalization performance without parameter updates.In addition,by comparing on a real three-phase induction motor experimental platform,the experimental results show that the proposed method also exhibits better fault diagnosis performance.3.Aiming at the effects of data noise,working condition variations,and limited available data on the performance of the fault diagnosis model.In this dissertation,we further explore the few-shot fault diagnosis strategy for electric motors with limited data samples,while considering the impact of different levels of data noise on the performance of diagnosis models.A few-shot fault diagnosis strategy for motor bearings based on model-agnostic meta-learning is proposed.In this study,fault diagnosis under different working conditions is regarded as a fast adaptation problem between different tasks.Firstly,motor bearing fault diagnosis under various operating conditions is regarded as a series of interrelated tasks.In order to learn the characteristics and patterns of specific fault diagnosis tasks under different operating conditions,a basic-learner with a convolutional neural network as its backbone is used to train on each fault diagnosis task and realize fault classification.Then,the meta-learner summarizes the training experience on all the specific fault diagnosis tasks and provides the basic-learner with the initial parameters of the model that can be quickly adapted to the new fault diagnosis tasks.Through the collaborative work of basic-learners and meta-learners,a basic-learner that can quickly adapt to new fault diagnosis tasks with limited data samples can be obtained.The performance of the proposed diagnostic strategy is validated on two public motor datasets and the three-phase induction motor experimental platform mentioned above,respectively.The relevant experimental results show that the proposed method can achieve fast adaptation to fault diagnosis tasks in different scenarios,and can achieve better fault diagnosis performance using fewer parameter update steps.Moreover,the proposed method in this dissertation still has better diagnostic performance on the built induction motor experimental platform.4.Most of the current fault diagnosis methods do not consider the impact of different data sampling frequencies on the model performance.A multi-task learning framework based on a knowledge sharing mechanism is proposed in this dissertation.This method improves the performance of tasks with low sampling frequency data input by sharing useful fault feature information from high sampling frequency data with low sampling frequency data.The performance of the proposed fault diagnosis framework is validated on two publicly available datasets,and the experimental results show that the knowledge sharing mechanism proposed in this dissertation can effectively improve the fault diagnosis accuracy of the model.