Research On Intelligent Fault Diagnosis Of Harmonic Drive Based On Deep Learning

As a core component with excellent transmission characteristics,harmonic drive is widely used in the joint transmission system of industrial robots.Its running performance directly affects the processing performance of industrial robots.However,the domestic harmonic drive generally has problems such as high failure rate and poor consistency,which seriously restricts the development of our industrial robot industry.The traditional harmonic drive fault detection method depends on the mature experience of technicians,the accuracy is difficult to guarantee and the efficiency is low.The existing intelligent fault diagnosis methods for common rotary equipment cannot be applied well to the precision integrated harmonic drive with elastic deformation characteristics and operating under variable working conditions of flexible series industrial robots.In the recent ten years,scholars at home and abroad have carried out a series of studies on motion characteristics,mechanical properties,and fatigue analysis of harmonic drive.However,research on fault diagnosis of the harmonic drive is still very scarce.Although a few researchers have carried out fault diagnosis of the inner and outer ring of flexible thinwall elliptical bearings in harmonic drives,these studies only focused on the components that can not well reflect the real fault characteristics of complete harmonic drives operating in complex working conditions.Therefore,this paper takes the harmonic drive as the research object,through the signal collected by the multiple vibration acceleration sensors installed at the end of the industrial robot and the multimodal signal of the servo drive system itself,to carry out the intelligent fault diagnosis of the harmonic drive based on deep learning,which has important guiding significance for improving the performance of the industrial robot.The main contents include the following four aspects:1)The original vibration acceleration signal of the harmonic drive is coupled with external strong noise interference,and the fault characteristics are weak,which will adversely affect the classification accuracy of the model,a fault detection method based on a Multi-scale Convolutional Neural Network(MSCNN)is proposed.First,the vibration acceleration signal of the harmonic drive is collected by multiple sensors to obtain the spectrum characteristics of each component,and data fusion is carried out by cascading each group of signals in turn.Second,the MSCNN composed of a multi-scale decomposition module and a classical 2D Convolutional Neural Network(2D-CNN)is constructed to automatically extract the comprehensive features of faults.Finally,the proposed method is tested on the vibration test bench of a real industrial robot,and it is verified that the difference between normal and fault characteristics can be expanded effectively,and the fault detection accuracy of the harmonic drive can be improved under varying working conditions.2)Most of the vibration data of harmonic drive are normal,and the fault data are scarce and unbalanced,which is difficult to apply to various machine learning methods based on a large amount of training data and data balance assumption,a fault diagnosis method based on Generative Adversarial Network(GAN)is proposed.First,the data collected from multiple vibration acceleration sensors are preprocessed by Fast Fourier Transform(FFT)to obtain the frequency spectrum of the vibration signal.Second,the multi-group generative adversarial networks are used to generate various kinds of fault data successively.Third,a data selection module composed of Euclidean distance and Mahalanobis distance is designed to filter and purify the generated data successively,so as to effectively solve the problem of GAN model instability.Finally,high-quality generated data and real data were selected to form a new balanced dataset,and then a multi-scale convolutional neural network was used for multiclassification diagnosis of various types of health status samples.Compared with the experimental results obtained by various traditional deep neural network methods,the proposed method is proved to have excellent multi-classification capability and robustness.3)The combination of compound faults driven by harmonic drives is diverse.However,only a variety of single-label fault data can be collected,and there is a lack of compound fault data of various types,and it is difficult to fully and accurately identify various components of compound faults.A novel compound fault diagnosis method based on Deep Capsule Graph Convolutional Network(DCGCN)is proposed.First,the frequency spectrum of fault signals is obtained by multiple vibration sensors and the label diagram of adjacency matrix is constructed.Second,not only the vector representation of capsule units in the deep capsule network is used to further obtain the fault features containing the pose information,but also the graph convolutional network is used to learn the relationship features between various single-label faults,so as to enhance the compound diagnosis ability.Third,the two networks are combined to obtain diagnosing results.Finally,a dynamic routing algorithm and marginal loss function are used to optimize DCGCN.By using a variety of single fault datasets and compound fault datasets,it is proved that the method can directly realize accurate fault identification of compound faults and obtain excellent compound fault diagnosis performance just by learning various types of single-label fault samples.4)Harmonic drive often works at too high speed and load,the occurrence of failure is unpredictable.Therefore,it is important to implement online fault diagnosis.This paper proposes an online intelligent fault diagnosis method of harmonic drive based on the Semisupervised Contrastive Graph Generative Network(SCGGN)via multimodal data.First,the multimodal signals(including current and motor encoder signals)of the robot servo system were collected in real time by data acquisition software,and the Euclidean distance was used to analyze the similarity of the normalized time/frequency domain data and construct the fault graph.Second,Multilayer Graph Convolution(MGCN)and Hierarchical Graph Convolution(HGCN)subnetworks are used to obtain complementary fault features in local and global views,respectively.Third,a contrastive learning network is constructed to obtain high-dimensional features of fault data through unsupervised learning,and data clustering is carried out to obtain multi-classification output results.Finally,SCGGN is optimized by combining the learnable loss function to improve the multi-classification performance of the model.Through the online fault diagnosis test of harmonic drive,it is verified that the proposed method can realize the accurate diagnosis of unknown fault samples by learning a few labeled fault samples and a large number of unlabeled samples,and the computational efficiency can meet the actual demand.The proposed method effectively improves the classification accuracy and generalization ability of multi-sensor data fusion,multi-scale feature extraction,scarce fault sample generation,compound fault diagnosis,multimodal signal online fault diagnosis tasks of harmonic drive under the interference of variable and strong noise,and provides a new means and ideas for intelligent fault diagnosis of integrated harmonic drive under complex environment.It is the frontier research on intelligent fault diagnosis of precision integrated harmonic drive used in industrial robots at home and abroad,which has important engineering significance for improving product quality and machining performance.