Industrial robot harmonic reducer fault diagnosis method and system under different working conditions based on digital twinning

Abstract

The invention discloses an industrial robot harmonic reducer fault diagnosis method and system under different working conditions based on digital twinning, and relates to the technical field of harmonic reducer fault diagnosis. According to the technical key points, the method comprises the following steps: firstly, constructing a harmonic reducer fault digital twinborn model based on a dynamic modeling method, and obtaining twinborn data; secondly, providing a digital twinning virtual-real mapping method based on a cyclic generative adversarial network, and realizing virtual-real mapping of twinning data and actually measured data; and then, an improved semi-soft threshold function is introduced to construct a deep residual shrinkage network, noise interference is suppressed, features are extracted, domain adaptation processing is performed on the extracted features in an unsupervised scene, distribution differences among fields are reduced by using MMD, and fault diagnosis under different working conditions is realized. And finally, building a harmonic reducer fault simulation test bed and carrying out experimental verification. According to the method, the fault diagnosis problem of the harmonic reducer of the industrial robot under different working conditions in an unsupervised scene is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of harmonic reducer fault diagnosis, and in particular to a method and system for diagnosing industrial robot harmonic reducer faults under different working conditions based on digital twins. Background Art

[0002] With the continuous improvement of the manufacturing industry, industrial robots have been widely used in traditional manufacturing due to their advantages such as stable and reliable operation and high efficiency. [1] The harmonic reducer is a high-performance reducer with compact structure, light weight, large reduction ratio and precise transmission. As the "joint" of industrial robots, its normal operation plays a vital role in the safe, accurate and reliable operation of industrial robots. [2] The typical structure of a harmonic reducer includes a rigid wheel, a flexible wheel and a wave generator. Since the core component of a harmonic reducer is a flexible structure, the risk of failure of a harmonic reducer is higher than that of a conventional reducer. Abnormalities in a harmonic reducer may lead to reduced production efficiency and unexpected downtime. [3] Therefore, accurate and efficient harmonic reducer fault diagnosis is of great significance for evaluating the overall health of industrial robots and improving their reliability and robustness. [4] .

[0003] At present, the academic community has conducted a lot of research on the transmission characteristics and mechanical analysis of harmonic reducers, but less on fault diagnosis. Reference [5] combines sparse autoencoders with support vector machine methods to build a fault diagnosis model to achieve reliable identification and diagnosis of various faults of industrial robot reducers. Reference [6] uses the wavelet region correlation threshold method to denoise the original sound signal of the harmonic reducer, combines operational modal analysis with variational mode decomposition to achieve fault detection, and achieves good results. With the continuous development of artificial intelligence technology, deep learning fault diagnosis methods have gradually become mainstream. Reference [7] suppresses the noise in the sensor data of the harmonic reducer by combining the wavelet region correlation threshold denoising algorithm, and combines the convolutional long short-term memory network to mine the hidden features of the data to improve the fault detection rate. Reference [8] combines a one-dimensional convolutional neural network with a matrix kernel adaptive model for continuous time domain vibration signals to achieve on-site fault diagnosis of harmonic reducers. Reference [9] uses a multi-scale hybrid convolutional neural network model to improve the accuracy of in-situ diagnosis of harmonic reducers.

[0004] Although the above-mentioned deep learning-based methods can achieve effective diagnosis of equipment, they are all based on the situation that the training data and the test data follow the same distribution. However, in actual industrial production, the working conditions of harmonic reducers are complex and changeable, and the data distribution under different working conditions is different. In addition, labeled data under some working conditions is scarce or difficult to obtain, which makes it difficult to meet the data volume required for training deep learning models, resulting in poor accuracy and generalization of the established diagnostic model.

[10] Therefore, domain adaptation transfer learning is introduced into rotating machinery fault diagnosis to utilize the knowledge and features learned under a certain working condition to improve the diagnostic performance under other working conditions.

[11] Reference

[12] proposed a deep domain adaptation architecture and used the maximum mean discrepancy (MMD) to measure the distribution difference to achieve the rolling bearing fault diagnosis task. Reference

[13] used the minimization of the local maximum mean difference and the classifier loss function to align the distribution of related subdomains and improve the recognition accuracy when the target domain is unlabeled. Reference

[14] used multi-core maximum mean difference in the domain adaptation module to reduce the distribution difference and performed well in cross-domain bearing fault diagnosis.

[0005] However, the actual operation of large-scale mechanical equipment faces challenges such as dynamic operating condition changes and lack of extreme operating condition data, which makes it difficult to fully cover all fault scenarios. This makes fault diagnosis based on transfer learning a major challenge, and there is an urgent need for a method to obtain more fault data.

[15] . With the development of science and technology, the emergence of digital twin technology has opened up new ideas for solving the problem of data acquisition difficulties. In 2003, Professor Grieves first proposed the "mirror space model" and defined the concept of "digital twin". Digital twin is the digital equivalent of virtual and physical products. With the development of new generation information technologies such as industrial Internet, big data, and cloud computing, digital twin is an effective means to achieve the interactive integration of the manufacturing information world and the physical world. At present, many scholars have applied digital twin to fault diagnosis. Reference

[16] combines SolidWorks and Simulink to establish an accurate digital twin model of bearings, obtains twin data and combines transfer learning to achieve fault diagnosis. Reference

[17] constructs a virtual model of mechanical equipment, performs feature enhancement on fault twin data and retrains the network, and finally uses it for fault diagnosis to solve the problem of small samples. Reference

[18] uses finite element analysis to establish a virtual simulation model of bearings to inject faults and generate rich fault data to achieve fault diagnosis in response to problems such as insufficient data collection and high cost. Reference

[19] used Simscape to establish a three-cylinder pump DT model and generated simulation data as the source domain, and combined it with a transfer learning model to solve the fault diagnosis problem when the fault signal was insufficient.

[0006] At the same time, considering the complex working environment and high background noise of harmonic reducers, how to accurately extract fault features from signals is a challenge. Reference

[20] proposed a deep residual contraction network. The network introduced the soft threshold as a contraction layer into the residual unit of ResNet, which confirmed the superiority of the soft threshold in filtering out noise-related features in the signal. Reference

[21] constructed a deep residual contraction network for bearing fault diagnosis. The network can extract more accurate fault features and reduce the impact of noise on model training.

[0007] The above method obtains fault data by building a digital twin model, extracts features using a deep residual shrinkage network, and combines the domain adaptation method to achieve good results in fault diagnosis of equipment under different working conditions. However, the main idea of ​​the above digital twin method is to build a virtual entity to help realize fault diagnosis by providing fault data, stress data or other physical data. This process often ignores the interactive connection between the physical world and the virtual world, and cannot fully realize the exchange of information. When the digital twin method is implemented with the help of finite element simulation software such as Unity 3D, although interaction can be achieved, the amount of calculation is large, and the computer hardware requirements are often high. In particular, the simulation of highly complex models will also limit the interactive function, increasing the development cost and time. At the same time, the soft threshold function in the deep residual shrinkage network structure will have a constant deviation in the denoising process, causing the fault-related information in the signal to be filtered out during denoising, causing signal distortion and reducing the accuracy of fault diagnosis. Summary of the invention

[0008] In view of the above problems, the present invention proposes a fault diagnosis method and system for an industrial robot harmonic reducer under different working conditions based on digital twins.

[0009] According to one aspect of the present invention, a fault diagnosis method for an industrial robot harmonic reducer under different working conditions based on digital twin is proposed, and the method comprises the following steps:

[0010] Step 1: Data collection and preprocessing

[0011] A harmonic reducer vibration signal experimental platform was built as a physical entity. The one-dimensional vibration signal data of all states of different working conditions were collected using a vibration acceleration sensor, and the data were converted into a two-dimensional time-frequency image by wavelet transform to obtain a measured data set in preparation for subsequent virtual-real mapping.

[0012] Step 2: Construction of harmonic reduction virtual entity

[0013] According to the relevant parameter information and motion mechanism of the physical entity of the harmonic reducer, the dynamic equations of the flexible wheel, rigid wheel and flexible bearing are established respectively. Then, by injecting faults and related speed information, the digital twin model of the harmonic reducer fault is obtained, and the simulation data of different fault types are obtained by solving.

[0014] Step 3: Digital twin virtual-real mapping

[0015] Construct and train the CycleGAN virtual-to-real mapping network, input the simulated data as X-domain data and the measured data as Y-domain data into the network respectively, and obtain the mapped image data through continuous iteration of cycle consistency loss and use it as the source domain sample set;

[0016] Step 4: Unsupervised domain adaptation network model construction

[0017] An improved semi-soft threshold function is introduced to construct a deep residual shrinkage network to extract data features of the source domain with known labels for a certain working condition and the target domain with unknown labels for other working conditions, while suppressing noise interference. Combined with the MMD domain adaptation method, the maximum mean difference between the source domain and the target domain is calculated to minimize the classification loss and domain loss to reduce the distribution differences and label differences between different working conditions, and continuously train the network.

[0018] Step 5: Multi-state identification of harmonic reducer

[0019] The target domain test samples are input into the trained network model, and finally the health status diagnosis results of the industrial robot harmonic reducer under different working conditions are obtained.

[0020] Furthermore, in step 2, based on the relevant parameter information and motion mechanism of the physical entity of the harmonic reducer, the dynamic equations of the flexible wheel, the rigid wheel and the flexible bearing are established respectively, including:

[0021] The dynamic equation of the flexible wheel is expressed as:

[0022] F m =K m (y out cosθ i +x out sinθ i )

[0023] Among them, K m is the meshing stiffness of the flexspline teeth, y out is the vertical displacement of the outer ring, x out is the horizontal displacement of the outer ring, θ i is the input end angle;

[0024] The rigid wheel dynamic equation is expressed as:

[0025]

[0026]

[0027] Among them, m is the mass of the rigid wheel, x c is the horizontal displacement of the rigid wheel, y c is the vertical displacement of the rigid wheel, F m F is the meshing force between the rigid wheel and the flexible wheel, t F is the tangential component of the meshing force on the outer circle of the rigid wheel, r is the axial component of the meshing force; α t Indicates the pressure angle of the bearing, k y Indicates the vertical stiffness coefficient, k x represents the horizontal stiffness coefficient;

[0028] The dynamic equation of the flexible bearing is expressed as:

[0029]

[0030] Among them, m in is the mass of the inner ring, m out is the mass of the outer ring, c in is the connection damping between the inner ring shaft and the shaft fixing point, c out is the connection damping between the outer ring base and the ground foundation, k in k is the connection stiffness between the inner ring shaft and the shaft fixing point, out The connection stiffness between the outer ring base and the ground foundation, F HX With F HY are the components of the elastic restoring force generated by the contact between the inner and outer rings and the rolling elements in the x and y directions, e is the eccentricity, ω is the angular frequency of the inner ring rotation, g is the gravitational acceleration, and t is the time; in Indicates the horizontal displacement of the inner ring; y in Indicates the vertical displacement of the inner ring; x out Represents the horizontal displacement of the outer ring; y out represents the vertical displacement of the outer ring; θ represents the rotation angle of the input end.

[0031] Furthermore, the CycleGAN virtual-to-real mapping network in step 3 is constructed as follows:

[0032] CycleGAN is a ring network consisting of two mirror-symmetric GANs, including two generators G X→Y and G Y→X , two discriminators D X and D Y ; Throughout the process, the generator is used to generate images that are similar enough to the target domain to deceive the discriminator; the discriminator determines whether the generated image is realistic enough; the total loss function of the CycleGAN network is:

[0033] L CycleGAN =L GAN (D X ,G Y→X )+L GAN (D Y ,G X→Y )+λL cyc (G X→Y ,G Y→X )

[0034] Where, L GAN (D X ,G Y→X ) represents the discriminator D X The adversarial loss function L GAN (DY ,G X→Y ) represents the discriminator D Y The adversarial loss function L cyc (G X→Y ,G Y→X ) represents the cycle-consistent loss function; λ is the control ratio.

[0035] Furthermore, the adversarial loss function of GAN in step 3 is as follows:

[0036] L GAN (D X ,G Y→X )=E x～data(x) [logD X (x)]+E y～data(y) [log(1-D X (G Y→X (y)))]

[0037] L GAN (D Y ,G X→Y )=E y～data(y) [logD Y (y)]+E x～data(x) [log(1-D Y (D X→Y (x)))]

[0038] In the formula, E x～data(x) With E y～data(y) are the losses of two pairs of domains X and Y;

[0039] The formula of the cycle-consistent loss function is expressed as:

[0040] L cyc (G X→Y ,G Y→X )=E x～data(x) [||G Y→X (G X→Y (x))-x||1]+E y～data(y) [||G X→Y (G Y→X (y))-y||1]

[0041] In the formula, G Y→X (G X→Y (x)) and G X→Y (G Y→X (y)) represent the domain transformation of the image by the two generators; x is the input feature of the X domain, and y is the input feature of the Y domain.

[0042] Furthermore, the deep residual shrinkage network described in step 4 is based on the deep residual network and introduces an attention mechanism and a semi-soft threshold; wherein the attention mechanism automatically learns a set of weights through a small sub-network to weight each channel of the feature; and the semi-soft threshold function is used for signal denoising.

[0043] Furthermore, the semi-soft threshold function formula is expressed as:

[0044]

[0045] Among them, x is the input feature, y is the output feature, and τ is the threshold.

[0046] Furthermore, the MMD domain adaptation method described in step 4 is combined to calculate the maximum mean difference between the source domain and the target domain, minimize the classification loss and domain loss to reduce the distribution difference and label difference between different working conditions, and continuously train the network including: achieving cross-domain adaptation of features by minimizing the MMD distance between the source domain and the target domain. The MMD calculation formula is expressed as

[0047]

[0048] Where H is the regenerated Hilbert space, φ(·) is the mapping function, and n s and n t are the number of samples in the source domain and the target domain respectively; X s represents the probability distribution of the source domain, X t Represents the probability distribution of the target domain;

[0049] The MMD domain adaptation method includes a feature extractor F and a classifier G; wherein source domain labeled data and target domain unlabeled data are respectively input, and the fault features of the source domain labeled data and the target domain unlabeled data are extracted through the feature extractor F; in the classifier G, the fault features are input, and after passing through a fully connected layer, the difference between the source domain and the target domain is reduced by calculating the MMD loss of the source domain labeled data and the target domain unlabeled data, and then the classification information is output and the classification loss is obtained; and the iterative training model is continuously updated by minimizing the classification loss of the source domain labeled samples and minimizing the MMD loss between the source domain and the target domain samples.

[0050] Furthermore, the calculation formula of the classification loss in step 4 is expressed as:

[0051]

[0052] Among them, K is the number of classification categories, n s is the number of source domain samples, represents the true label of source domain sample i in category k, It represents the probability that the source domain sample i belongs to category k predicted by the model.

[0053] Furthermore, all states of different working conditions described in step 1 include: mild, moderate and severe faults of the inner ring of the flexible bearing; mild, moderate and severe faults of the outer ring of the flexible bearing; severe fault of the rolling body weight; rigid wheel missing tooth fault; flexible wheel missing tooth fault; normal state.

[0054] According to another aspect of the present invention, a fault diagnosis system for industrial robot harmonic reducer under different working conditions based on digital twin is proposed, and the system includes:

[0055] The data acquisition and preprocessing module is configured to build a harmonic reducer vibration signal experimental platform as a physical entity, use a vibration acceleration sensor to collect one-dimensional vibration signal data of all states of different working conditions, and perform wavelet transform on it to convert it into a two-dimensional time-frequency image to obtain a measured data set, in preparation for subsequent virtual-real mapping;

[0056] The harmonic reducer virtual entity construction module is configured to establish the dynamic equations of the flexible wheel, rigid wheel and flexible bearing according to the relevant parameter information and motion mechanism of the harmonic reducer physical entity, and then obtain the digital twin model of the harmonic reducer fault by injecting fault and related speed information, and solve to obtain simulation data of different fault types;

[0057] The digital twin virtual-to-real mapping module is configured to build and train the CycleGAN virtual-to-real mapping network, input the simulated data as the X-domain data and the measured data as the Y-domain data into the network respectively, and obtain the mapped image data through continuous iteration of cycle consistency loss as the source domain sample set;

[0058] The unsupervised domain adaptation network model training module is configured to introduce an improved semi-soft threshold function to construct a deep residual shrinkage network, extract the data features of the source domain with known labels for a certain working condition and the target domain with unknown labels for other working conditions, and suppress noise interference; combined with the MMD domain adaptation method, by calculating the maximum mean difference between the source domain and the target domain, the classification loss and domain loss are minimized to reduce the distribution difference and label difference between different working conditions, and the network is continuously trained;

[0059] The harmonic reducer multi-state recognition module is configured to input the target domain test samples into the trained network model, and finally obtain the health status diagnosis results of the industrial robot harmonic reducer under different working conditions.

[0060] The beneficial technical effects of the present invention are:

[0061] The present invention proposes a method and system for fault diagnosis of harmonic reducers of industrial robots under different working conditions based on digital twins, and specifically proposes a digital twin domain adaptation method based on dynamic modeling and CycleGAN virtual-to-real mapping to realize fault diagnosis of harmonic reducers of industrial robots under different working conditions. The present invention uses the harmonic reducer dynamics model to generate fault simulation data, and combines the CycleGAN network virtual-to-real mapping method to realize the whole process of digital twins, and solve the problems of few fault samples and missing data labels; introduces an improved semi-soft threshold function instead of the soft threshold function to construct a deep residual shrinkage network for feature extraction network, and combines the MMD domain adaptation method to suppress noise interference while effectively extracting fault features, and reduce the distribution difference between the source domain and the target domain, and finally realizes the fault diagnosis of harmonic reducers of industrial robots under different working conditions in unsupervised scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification, and are used to further illustrate the preferred embodiments of the present invention and explain the principles and advantages of the present invention.

[0063] Figure 1 It is a schematic diagram of the dynamic model of flexible bearing.

[0064] Figure 2 This is a schematic diagram of a local failure of the inner ring of a flexible bearing.

[0065] Figure 3 It is a schematic diagram of local failure of the outer ring of the flexible bearing.

[0066] Figure 4 It is a schematic diagram of local failure of the rolling element of a flexible bearing.

[0067] Figure 5 This is the CycleGAN network structure diagram.

[0068] Figure 6(a) is a schematic diagram of the CycleGAN network generator structure; Figure 6(b) is a schematic diagram of the CycleGAN network discriminator structure.

[0069] Figure 7 This is a comparison chart of different threshold functions.

[0070] Figure 8 This is a schematic diagram of the domain adaptation training process based on the MMD metric.

[0071] Fig. 9 This is a flowchart of fault diagnosis of industrial robot harmonic reducer under different working conditions based on digital twins described in the present invention.

[0072] Fig.10This is an example diagram of the harmonic reducer vibration signal acquisition experimental platform.

[0073] Fig.11 This is an example diagram of the measurement and control system.

[0074] Fig.12 These are example diagrams of the time domain waveform and envelope spectrum of the IR1 simulation signal under working condition A; (a) is the time domain waveform of the IR1 simulation signal under working condition A; (b) is the envelope spectrum of the IR1 simulation signal under working condition A; (c) is the envelope spectrum of the real IR1 signal under working condition A.

[0075] Fig.13 These are example diagrams of the time domain waveform and envelope spectrum of the OR10 simulation signal under working condition A; (a) is the time domain waveform of the OR10 simulation signal under working condition A; (b) is the envelope spectrum of the OR10 simulation signal under working condition A; (c) is the envelope spectrum of the OR10 real signal under working condition A.

[0076] Fig.14 The following are examples of the time domain waveform and envelope spectrum of the B10 simulation signal under working condition A; (a) is the time domain waveform of the B10 simulation signal under working condition A; (b) is the envelope spectrum of the B10 simulation signal under working condition A; (c) is the envelope spectrum of the real signal of B10 under working condition A.

[0077] Fig.15 It is a comparison chart of the diagnostic results before and after the virtual-real mapping of different fault data.

[0078] Fig.16 It is a comparison chart of the diagnosis results before and after virtual-real mapping of IR1 for different tasks.

[0079] Fig.17 This is a comparison chart of the results before and after virtual-reality mapping of 12 tasks.

[0080] Fig.18 This is a feature visualization of the original residual network ResNet50.

[0081] Fig.19 It is a visualization diagram of the improved DRSN network features.

[0082] Fig. 20 This is the confusion matrix diagram of the original residual network ResNet50.

[0083] Fig.21 It is the improved DRSN network confusion matrix diagram.

[0084] Fig. 22 This is a comparison chart of fault diagnosis accuracy under different noise intensities.

[0085] Fig.23 This is a graph showing experimental comparison results between the present invention and other deep learning methods. DETAILED DESCRIPTION

[0086] In order to enable those skilled in the art to better understand the scheme of the present invention, exemplary implementations or embodiments of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described implementations or embodiments are only implementations or embodiments of a part of the present invention, not all of them. Based on the implementations or embodiments of the present invention, all other implementations or embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0087] As an important component of industrial robots, harmonic reducers operate in a complex and changeable environment, and once a fault occurs, it will cause huge losses. Considering the difficulty in collecting actual harmonic reducer vibration data, the small number of fault samples and the lack of data labels, and the difference in data distribution under different working conditions, the embodiment of the present invention proposes a fault diagnosis method for industrial robot harmonic reducers under different working conditions based on digital twins.

[0088] 1. Construction of digital twin model of harmonic reducer

[0089] The digital twin model consists of two parts: virtual entity modeling and virtual-real mapping. The virtual entity of the harmonic reducer is constructed by dynamic modeling, and the virtual-real mapping is realized by using the CycleGAN network.

[0090] 1.1 Harmonic reducer fault dynamics modeling

[0091] The harmonic reducer consists of a rigid wheel, a flexible wheel and a wave generator. It has a complex structure and deformable parts. Since the deformable parts cannot be treated as rigid bodies, the harmonic reducer is split and simplified into separate components for modeling to reduce the difficulty of modeling. According to the dynamic characteristics and fault displacement excitation of the harmonic reducer, the dynamic models of the harmonic reducer system with different fault degrees of the inner ring, outer ring and rolling element of the flexible bearing are constructed respectively.

[0092] Under normal conditions, the flexspline is deformed into an ellipse due to the continuous meshing of the gear teeth during the rotation of the harmonic reducer, and cannot be regarded as a rigid body. Therefore, the action of the flexspline is directly abstracted as the meshing force. The dynamic equation is expressed as

[22]

[0093] F m =K m (y out cosθ i +x out sinθ i ) (1)

[0094] Among them, K m is the meshing stiffness of the flexspline teeth, y out is the vertical displacement of the outer ring, x out is the horizontal displacement of the outer ring, θi is the input corner.

[0095] Under normal conditions, the rigid wheel is stationary during operation and is only subject to the support force of the support and the meshing force of the flexible wheel. The dynamic equation of the rigid wheel is expressed as

[0096]

[0097]

[0098] Where m is the mass of the rigid wheel, x c is the horizontal displacement of the rigid wheel, y c is the vertical displacement of the rigid wheel, F m F is the meshing force between the rigid wheel and the flexible wheel, t F is the tangential component of the meshing force on the outer circle of the rigid wheel, r is the axial component of the meshing force. t Indicates the pressure angle of the bearing, k y Indicates the vertical stiffness coefficient, k x Represents the horizontal stiffness coefficient.

[0099] The modeling of the flexible bearing can be compared to that of a rolling bearing. When in operation, it deforms into an ellipse and contacts the rigid wheel at the long axis end, which generates a meshing force. The four-degree-of-freedom dynamic equation of the flexible bearing is established as shown in equation (4). The dynamic model of the flexible bearing is as follows: Figure 1 shown.

[0100]

[0101] Among them, m in is the mass of the inner ring, m out is the mass of the outer ring, c in is the connection damping between the inner ring shaft and the shaft fixing point, c out is the connection damping between the outer ring base and the ground foundation, k in k is the connection stiffness between the inner ring shaft and the shaft fixing point, out The connection stiffness between the outer ring base and the ground foundation, F HX With F HY are the components of the elastic restoring force generated by the contact between the inner and outer rings and the rolling elements in the x and y directions, e is the eccentricity, ω is the angular frequency of the inner ring rotation, g is the gravitational acceleration, and t is the time; in Indicates the horizontal displacement of the inner ring; y in Indicates the vertical displacement of the inner ring; x out Represents the horizontal displacement of the outer ring; y out represents the vertical displacement of the outer ring; θ represents the rotation angle of the input end.

[0102] After the local fault is introduced, the fault location can be simplified to a rectangular pit. When the rolling element enters and leaves the fault area, the elastic deformation δ i will change, and the formula is expressed as

[0103] δ i =(x in -x out )sinθ i +(y in -y out )cosθ i -c r -H (5)

[0104] Where H is the radial deformation caused by local fault.

[0105] When the inner ring fails, the local fault diagram is as follows Figure 2 As shown in the figure, the radial deformation H generated when the rolling element passes through the fault area in It is expressed as shown in formula (6).

[0106]

[0107] When the outer ring fails, the local fault diagram is as follows Figure 3 As shown in the figure, the radial deformation H generated when the rolling element passes through the fault area out The formula is shown in formula (7).

[0108]

[0109] When a rolling element fails, the local failure diagram is as follows Figure 4 As shown in the figure, the defective ball will contact the inner and outer raceways respectively as the flexible bearing rotates, and the radial deformation H ball The formula is shown in formula (8).

[0110]

[0111] According to the written dynamic equations of each part, the fault dynamics model of the harmonic reducer system is constructed, and the simulation data of various states are obtained by solving it through the Runge-Kutta numerical integration method.

[0112] 1.2 Digital Twin Virtual-Real Mapping Based on CycleGAN

[0113] CycleGAN was proposed by the literature

[23] in 2017. Its main advantage is that it can achieve the mapping between input images and output images without paired training data. CycleGAN is a ring network composed of two mirror-symmetric GANs, including two generators G X→Y and G Y→X, two discriminators D X and D Y , the network structure is as follows Figure 5 shown.

[0114] Throughout the process, the generator is used to generate images that are similar enough to the target domain to deceive the discriminator; the discriminator tries to determine whether the generated image is realistic enough. In order to ensure that images can be converted between different domains, a cycle-consistent loss function is introduced, which is expressed as

[0115] L cyc (G X→Y ,G Y→X )=E x～data(x) [||G Y→X (G X→Y (x))-x||1]+E y～data(y) ]||G X→Y (G Y→X (y))-y||1] (9)

[0116] The adversarial loss function of GAN is as follows:

[0117] L GAN (D X ,G Y→X )=E x～data(x) [logD X (x)]+E y～data(y) [log(1-D X (G Y→X (y)))]

[0118] L GAN (D Y ,G X→Y )=E y～data(y) [logD Y (y)]+E x～data(x) [log(1-D Y (D X→Y (x)))] (10)

[0119] In the formula, E x～data(x) With E y～data(y) are the losses of two domains X and Y; G Y→X (G X→Y (x)) and G X→Y (G Y→X (y)) represent the domain transformation of the image by the two generators; x is the input feature of the X domain, and y is the input feature of the Y domain.

[0120] The total loss function of the entire CycleGAN is:

[0121] L CycleGAN =LGAN (D X ,G Y→X )+L GAN (D Y ,G X→Y )+λL cyc (G X→Y ,G Y→X ) (11)

[0122] Among them, λ is the control ratio.

[0123] Based on the characteristics of CycleGAN, the present invention uses the virtual fault data obtained by dynamic model simulation as the X domain and the real fault data actually collected as the Y domain. Through the continuous adversarial training of the network, the virtual data (X domain) can learn the characteristics of the real data (Y domain) and retain the original characteristics, realizing the mutual mapping of virtual reality. The structures of the generator and the discriminator are shown in Figure 6.

[0124] By constructing a virtual entity of the harmonic reducer and simulating to obtain fault data, the CycleGAN network is used to achieve virtual-real mapping and realize the entire process of the harmonic reducer digital twin.

[0125] 2 Improved Deep Residual Contraction Network

[0126] The deep residual shrinkage network (DRSN) is based on the deep residual network and introduces the attention mechanism and soft threshold.

[0127] 2.1 Attention Mechanism

[0128] The DRSN network uses the compression and excitation network module proposed in the literature

[24] . This module automatically learns a set of weights through a small sub-network to weight each channel of the feature. This module can enhance useful feature channels, reduce redundant feature channels, and improve the accuracy of model classification.

[0129] 2.2 Improved semi-soft threshold function

[0130] The soft threshold function in DRSN is used for signal noise reduction. It can set the features with absolute values ​​lower than the threshold to 0 and shrink the features with absolute values ​​greater than the threshold to 0. The formula of the soft threshold function is expressed as

[0131]

[0132] Among them, x is the input feature, y is the output feature, and τ is the threshold. From formula (12), it can be seen that the soft threshold function has good continuity overall, but when |x|>τ, a constant deviation will always be generated, resulting in the loss of effective information and affecting the diagnosis effect.

[0133] In order to solve this problem, the present invention adopts a semi-soft threshold

[25] The function is introduced into DRSN, and its expression is expressed as:

[0134]

[0135] The soft threshold and semi-soft threshold function images are as follows Figure 7 shown.

[0136] From formula (13) and Figure 7 It can be seen that the semi-soft threshold function retains the continuity of the soft threshold function. When |x|<τ, the near-zero feature can be set to 0; especially when |x|≥τ, as |x| increases, the deviation between the input and output gradually decreases, eliminating the constant deviation problem caused by the soft threshold, retaining the effective features of the signal to the greatest extent, and suppressing noise interference.

[0137] 2.3LeakyReLU function

[0138] The ReLU activation function has fast calculation speed and low calculation complexity, but the gradient is 0 in the negative area, and the neuron is always in a "necrotic" state. Therefore, the present invention uses the LeakyReLU activation function to replace the ReLU function. The LeakyReLU function gives a small gradient in the negative part, which effectively solves the problem that the ReLU output is 0. The LeakyReLU function formula is shown in formula (14):

[0139]

[0140] Among them, k is the leak coefficient, which has a very small value.

[0141] 3 Domain Adaptation Method Based on MMD

[0142] Domain adaptation technology is an important branch of transfer learning. This technology reduces the data distribution differences between different domains by mining domain-invariant features.

[0143] The maximum mean difference is the most commonly used metric to measure the difference between two domain distributions in unsupervised transfer learning. In recent years, it has been widely used in domain adaptation to achieve cross-domain adaptation of features by minimizing the MMD distance between the source domain and the target domain. It is defined as the distance between two distributions in the reproducing kernel Hilbert space (RKHS). The MMD calculation formula is expressed as

[0144]

[0145] Where H is the regenerated Hilbert space, φ(·) is the mapping function, and n s and nt are the number of samples in the source domain and the target domain respectively; X s represents the probability distribution of the source domain, X t Represents the probability distribution of the target domain.

[0146] Therefore, the present invention also uses this function to measure the difference in cross-domain distribution. The MMD-based domain adaptation method includes a feature extractor F and a classifier G. The source domain labeled data x and the target domain unlabeled data y are respectively input, and the x and y signal fault features are extracted through the feature extractor F; in the classifier G, the fault features are input, and the MMD loss L of x and y is calculated by passing through the fully connected layer. MMD Make the difference between the source domain and the target domain smaller, then output the classification information and get the classification loss L c By minimizing the classification loss L of the source domain labeled samples c and minimize the MMD distance L between source and target domain samples MMD , continuously update and iterate to train a better model so that the model can have good label prediction for the target domain unlabeled data y Finally, the unlabeled data in the target domain can be well classified. The calculation formula of the classification loss is expressed as:

[0147]

[0148] Among them, K is the number of classification categories, n s is the number of source domain samples, represents the true label of source domain sample i in category k, It represents the probability that the source domain sample i belongs to category k predicted by the model. The domain adaptation training process based on MMD metric is as follows: Figure 8 shown.

[0149] 4 Fault diagnosis methods for harmonic reducers of industrial robots under different working conditions

[0150] The fault diagnosis process of industrial robot harmonic reducer under different working conditions based on digital twin is as follows: Fig. 9 shown.

[0151] The specific steps are as follows:

[0152] Step 1: Data collection and preprocessing. Build a harmonic reducer vibration signal experimental platform as a physical entity, use a vibration acceleration sensor to collect one-dimensional vibration signal data of all states under different working conditions, and perform wavelet transform on it to convert it into a two-dimensional time-frequency image to obtain a measured data set, in preparation for subsequent virtual-real mapping.

[0153] Step 2: Construction of harmonic reducer virtual entity. According to the relevant parameter information and motion mechanism of the physical entity of the harmonic reducer, the dynamic equations of the flexible wheel, rigid wheel and flexible bearing are established respectively. The digital twin model of the harmonic reducer fault is obtained by injecting fault and related speed information, and the simulation data of different fault types are obtained by solving.

[0154] Step 3: Digital twin virtual-real mapping. Construct and train the CycleGAN virtual-real mapping network, input the simulation data as X-domain data and the measured data as Y-domain data into the network respectively, and obtain the mapped image data through continuous iteration of cycle consistency loss as the source domain sample set.

[0155] Step 4: Unsupervised domain adaptation network model construction. An improved semi-soft threshold function is introduced to construct a deep residual shrinkage network to extract data features of the source domain with known labels for a certain working condition and the target domain with unknown labels for other working conditions, while suppressing noise interference. Combined with the MMD domain adaptation method, the maximum mean difference between the source domain and the target domain is calculated, and the classification loss and domain loss are minimized to reduce the distribution differences and label differences between different working conditions, and the network is continuously trained.

[0156] Step 5: Multi-state identification of harmonic reducer. Input the target domain test samples into the trained network model, and finally obtain the health status diagnosis results of the industrial robot harmonic reducer under different working conditions.

[0157] 5 Application and Analysis

[0158] 5.1 Experimental bench construction

[0159] The experimental data of the present invention are all from the harmonic reducer vibration signal acquisition experimental platform built in the laboratory. The actual platform is shown in the figure. Fig.10 shown.

[0160] The measurement and control system of the harmonic reducer vibration signal acquisition experimental platform is mainly composed of a motor control box, a drive motor, a torque and speed sensor, a reducer to be tested, a load motor, an acceleration sensor and a data acquisition instrument. The schematic diagram is shown in the figure. Fig.11 During the experiment, acceleration sensors were placed at 3, 6, 9 and 12 o'clock on the input and output ends of the harmonic reducer, and a computer was used to collect vibration signals with a frequency of 100kHz.

[0161] 5.2 Introduction to Experimental Dataset

[0162] The present invention takes the green harmonic LCSG-32-50-I model harmonic reducer as an example, and sets the mild, moderate and severe faults of the inner ring of the flexible bearing, the mild, moderate and severe faults of the outer ring of the flexible bearing, the heavy fault of the rolling body, the tooth missing fault of the rigid wheel, the tooth missing fault of the flexible wheel, and the normal state, a total of 10 states. Among them, the mild, moderate and severe faults of the flexible bearing correspond to laser pitting 1 time, 5 times and 10 times respectively. As shown in Table 1, the 10-state brief description method of the harmonic reducer data is shown.

[0163] Table 1 Data representation methods for different status types

[0164]

[0165] There are four working conditions of the harmonic reducer of the present invention, as shown in Table 2.

[0166] Table 2 Harmonic reducer working conditions

[0167]

[0168] A total of 12 different working condition migration tasks were set up in the experiment, and the composition of the data sets used for each task is shown in Table 3.

[0169] Table 3 Composition of the datasets used for each task

[0170]

[0171]

[0172] Taking Task 1 as an example, Task 1 indicates that the source domain is working condition A and the target domain is working condition B. The source domain has 10 states, with 200 labeled samples in each state. The target domain also has 10 states, with 300 unknown label samples in each state, which are used as target domain training samples and target domain test samples in a ratio of 2:1.

[0173] 5.3 Virtual Entity Modeling

[0174] Based on the dynamic equations established in Section 1.1, the dynamic simulation of the fault states of the inner and outer rings and rolling elements of the flexible bearing under different working conditions shown in Table 2 is performed to obtain the dynamic response of the acceleration of the harmonic reducer system. Taking the mild fault (IR1) of the inner ring of the flexible bearing, the severe fault (OR10) of the outer ring of the flexible bearing and the severe fault (B10) of the rolling element of the flexible bearing under working condition A as examples, the time domain and frequency domain diagrams are obtained by simulation and compared with the real signals, as shown in Figure 2. Fig.12 , Fig.13 and Fig.14 shown.

[0175] pass Fig.12 (a) Fig.13 (a) and Fig.14(a) It can be observed that periodic pulses are generated in the simulated time domain signals. Fig.12 (b) Fig.13 (b) and Fig.14 (b) It can be seen that the envelope spectrum of the simulation signal mainly shows the rotation frequency, fault frequency and its multiples, which can well reflect the fault characteristics of the harmonic reducer during rotation.

[0176] Compared with the real signal envelope spectrum collected by the industrial robot harmonic reducer fault test bench, Fig.12 (b) and Fig.12 (c) It can be seen that under working condition A, the fault frequency in the real signal envelope spectrum at IR1 is 216.7 Hz, while the fault frequency of the simulated signal is 216.8 Hz, which is only 0.1 Hz different and basically consistent. Fig.13 (b) and Fig.13 (c) It can be seen that under working condition A, when OR10, the fault frequency in the envelope spectrum of the real signal is 173.9Hz, and the double frequency is 33.6Hz, while the fault frequency of the simulated signal is 172.5Hz, and the double frequency is 33.3Hz. The error is very small and basically consistent. Fig.14 (b) and Fig.14 (c) It can be seen that under working condition A at B10, the fault frequency in the real signal envelope spectrum is 183.1 Hz, and the double rotation frequency is 36.6 Hz, while the fault frequency of the simulation signal is 182.0 Hz, and the double rotation frequency is 33.3 Hz, which are basically consistent, further verifying the effectiveness of the established fault dynamics model.

[0177] 5.4 Comparison experiment before and after virtual-real mapping

[0178] The virtual-to-real mapping method based on CycleGAN can map and fuse the features of simulation data with real data. In order to verify the mapping ability of the network model, firstly, taking Task 2 under working condition A as an example, the data of the five generated fault types (IR5, IR5, OR1, OR5, B10) are combined with the improved DRSN and MMD domain adaptation methods proposed in this invention for comparative experiments. Then, taking Tasks 1 to 4 as examples, one of the fault types (OR1) is selected and combined with the improved DRSN and MMD domain adaptation methods proposed in this invention for comparative experiments. The specific experimental results are shown in Figure 2. Fig.15 and Fig.16 shown.

[0179] Depend on Fig.15It can be seen that the average diagnostic accuracy of the five fault types after virtual-to-real mapping is as high as 98.5%, and the average accuracy of fault diagnosis using only simulated data and real data is 95.7% and 95.6%, respectively, which is an increase of about 2.8% and 2.9%. It can be proved that the data after virtual-to-real mapping in different fault types not only retains the characteristics of the simulated signal, but also learns the characteristics of the real signal, with richer feature information, so that the model can be fully trained and the diagnostic accuracy of the harmonic reducer model is improved. Fig.16 It can be seen that the data after virtual-to-real mapping showed excellent diagnostic effects in tasks 1-4, especially the diagnostic accuracy of task 4 reached 99.8%. Compared with the fault diagnosis results using only simulated signals or real signals alone, the average accuracy was improved by about 3.2% and 3.0%, respectively. This further verifies that the data after virtual-to-real mapping has more fault feature information than the original simulated signal and real signal data, and can show good diagnostic effects in different tasks.

[0180] In order to further verify the effectiveness of the data mapped from virtual to real using the CycleGAN method in various tasks, the simulated signal, the real signal and the signal mapped from virtual to real were input into the CycleGAN network respectively, and comprehensive experiments were conducted in 12 migration tasks. The experimental results were plotted as radar charts, as shown in Figure 2. Fig.17 shown.

[0181] Depend on Fig.17 It can be seen that in various migration tasks, the data after virtual-to-real mapping has an average accuracy of about 3.4% and 3.1% higher than the simulated signal and the real signal, respectively. In summary, the introduction of the CycleGAN virtual-to-real mapping method can better realize the fault diagnosis of industrial robot harmonic reducers under different working conditions.

[0182] 5.5 Comparative Experiments before and after the Introduction of Improved DRSN Network

[0183] In order to verify the effectiveness of the improved DRSN network in the unsupervised fault diagnosis of harmonic reducers under different working conditions, the original residual network ResNet50 and the improved DRSN network with the same ResNet50 structure are used to extract the deep features of the harmonic reducer time-frequency image after virtual-real mapping in the noise-free case, and the t-distributed random neighborhood embedding algorithm is used for feature visualization. Taking Task 2 as an example, the feature visualization results are shown in Figure 2. Fig.18 and Fig.19 shown.

[0184] pass Fig.18 and Fig.19It can be observed that there are sample misclassification phenomena before and after the feature extraction network is improved, but there are still large differences in feature distribution. In the original residual network ResNet50, there are multiple sample misclassification areas such as L1, L2, L3, L5 and L6, and feature aliasing occurs in samples in the L4 area. In the improved DRSN network, only a small number of samples in the L1 and L2 areas are misclassified, and the boundaries of each type of sample are clear. The misclassification phenomenon and sample aliasing phenomenon have been significantly improved. It can be verified that the introduction of the improved DRSN network can better extract the deep features of the harmonic reducer, which helps to improve the accuracy of model fault diagnosis.

[0185] In order to more intuitively observe the effectiveness of the improved DRSN network in the harmonic reducer fault diagnosis problem, a multi-classification confusion matrix is ​​introduced to analyze the diagnosis results. Taking Task 2 as an example, the original residual network ResNet50 and the improved DRSN network are used in combination with the domain adaptation method proposed in this invention to diagnose the harmonic reducer fault. The confusion matrix is ​​drawn as follows: Fig. 20 and Fig.21 shown.

[0186] pass Fig. 20 and Fig.21 It can be seen that when the original residual network is used to extract the fault features of the harmonic reducer samples after virtual-real mapping and perform feature alignment, the CMT, FMT, IR1, OR1 and N health states cannot be correctly identified, resulting in misclassification, especially in the FMT class samples, 9 samples are misclassified as N class states. After the improved DRSN network is used to extract fault features and align features, although the sample misclassification phenomenon still occurs in the FMT class samples, only 3 samples are misclassified as N class states. Compared with the network before the improvement, the number of misclassified samples is significantly reduced, and the classification accuracy of the model for CMT, IR1, OR10 and N class samples is significantly improved. It can be seen that the improved DRSN network can better mine the deep features of the samples and extract richer information from the samples, thereby improving the accuracy of the industrial robot harmonic reducer diagnosis model under different working conditions.

[0187] 5.6 Comparative experiment on fault diagnosis under different noise intensities

[0188] In order to verify the effectiveness of the proposed method in suppressing noise interference, -4dB to 6dB Gaussian white noise is added to the test data to simulate the background noise of the actual factory environment of the harmonic reducer, and the original residual network ResNet50, DRSN and improved DRSN combined with the proposed MMD domain adaptation method are used to perform fault diagnosis on the virtual-real mapping time-frequency diagram of the harmonic reducer with and without noise (None means no noise is added), and a line graph of the diagnosis accuracy under different noise intensities is obtained. Taking Task 5 as an example, the experimental results are shown in Figure 2. Fig. 22 shown.

[0189] Depend on Fig. 22 It can be seen that the addition of noise will affect the results of fault diagnosis. When no noise is added, all three methods have high diagnostic accuracy, and the average accuracy can reach more than about 95.0%. However, with the increase of noise intensity, the performance of ResNet50 network gradually deteriorates, especially when the signal-to-noise ratio is -4dB, the average accuracy is only 83.7%, indicating that ResNet50 does not have the ability to cope with noise interference. The DRSN network still performs well in the diagnosis effect when the signal-to-noise ratio is 2dB~6dB, which can prove that the soft thresholding function in the network has a certain filtering effect on noise; but when the signal-to-noise ratio is -4dB~0dB, due to too much noise signal, the constant deviation problem in the soft threshold function leads to the loss of original signal characteristics while suppressing noise, resulting in a decrease in accuracy. The improved DRSN network uses a semi-soft threshold function to solve the constant deviation problem, effectively suppress noise and retain useful information, so that the average diagnostic accuracy is at least about 0.8% higher than before the improvement, and at most about 2.1%, proving that the proposed method has strong noise resistance.

[0190] 5.7 Comparison of the proposed method with other deep learning methods

[0191] In order to verify that the proposed method has better effect in fault diagnosis of industrial robot harmonic reducer in unsupervised scenarios under different working conditions, the present invention selects four deep transfer learning methods, namely dynamic adversarial domain adaptation network (DAAN), deep adaptation network (DAN), domain adversarial neural networks (DANN) and deep subdomain adaptation network (DSAN), for comparative experiments. In the experiment, all methods use the images after virtual-to-real mapping of digital twins as the source domain data set and the real data as the target domain, and are tested in 12 migration tasks. The results are as follows: Fig.23 shown.

[0192] Depend on Fig.23 The experimental results show that the proposed method has a significantly higher accuracy than other deep learning methods in 12 migration task tests, with an average accuracy of 99.2%. Compared with the deep learning methods of DAAN, DAN, DANN and DSAN, the average accuracy is increased by about 5.2%, 7.9%, 5.8% and 3.4% respectively. The experiment shows that the proposed method can effectively realize the fault diagnosis of industrial robot harmonic reducer under different working conditions.

[0193] 6 Conclusion

[0194] In view of the difficulty in collecting vibration data of harmonic reducers of actual industrial robots, the small number of fault samples and missing data labels, and the large differences in data distribution under different working conditions, this paper proposes a digital twin MMD domain adaptive fault diagnosis method for unsupervised scenarios. The data collected by the industrial robot harmonic reducer fault test bench was verified, and the following conclusions were obtained:

[0195] (1) A method for constructing a digital twin model of a harmonic reducer fault based on dynamic modeling is proposed. According to the dynamic characteristics of the harmonic reducer and the fault displacement excitation, a dynamic model of a harmonic reducer system with different fault degrees of the inner ring, outer ring and rolling elements of the flexible bearing is constructed. The effectiveness of the constructed fault dynamic model is proved by comparing it with the time-frequency characteristics of the real signal.

[0196] (2) A digital twin virtual-to-real mapping method based on CycleGAN is proposed to solve the problem of deviation between simulated signal data and measured data. In 12 migration task experiments, the image data after virtual-to-real mapping between simulated signals and real signals was input into the network for diagnosis. Compared with the single signal before virtual-to-real mapping, the average accuracy rate was improved by at least 2.8% and at most 3.4%, which proved the effectiveness of the CycleGAN digital twin virtual-to-real mapping method.

[0197] (3) An improved deep residual shrinkage network combined with MMD domain adaptation method is proposed, and the semi-soft threshold function is used to suppress noise interference and effectively extract fault features. In 12 migration task experiments, the improved deep residual shrinkage network combined with MMD can better extract domain-invariant features of the source domain and the target domain. Compared with the original residual network ResNet50 combined with MMD, the average accuracy is improved by 10.9%.

[0198] (4) A method for fault diagnosis of harmonic reducers of industrial robots under different working conditions based on digital twins is proposed. This method effectively alleviates the shortage of data and enriches the characteristics of simulation data after mapping a large amount of fault data obtained by simulation. At the same time, it realizes the fault diagnosis of harmonic reducers under different working conditions in unsupervised scenarios. Experimental verification shows that the average accuracy of the proposed method in 12 migration tasks is as high as 99.2%, which is at least 3.4% higher than the accuracy of other deep learning methods.

[0199] Another embodiment of the present invention provides a fault diagnosis system for industrial robot harmonic reducer under different working conditions based on digital twins, the system comprising:

[0200] The data acquisition and preprocessing module is configured to build a harmonic reducer vibration signal experimental platform as a physical entity, use a vibration acceleration sensor to collect one-dimensional vibration signal data of all states of different working conditions, and perform wavelet transform on it to convert it into a two-dimensional time-frequency image to obtain a measured data set, in preparation for subsequent virtual-real mapping;

[0201] The harmonic reducer virtual entity construction module is configured to establish the dynamic equations of the flexible wheel, rigid wheel and flexible bearing according to the relevant parameter information and motion mechanism of the harmonic reducer physical entity, and then obtain the digital twin model of the harmonic reducer fault by injecting fault and related speed information, and solve to obtain simulation data of different fault types;

[0202] The digital twin virtual-to-real mapping module is configured to build and train the CycleGAN virtual-to-real mapping network, input the simulated data as the X-domain data and the measured data as the Y-domain data into the network respectively, and obtain the mapped image data through continuous iteration of cycle consistency loss as the source domain sample set;

[0203] The unsupervised domain adaptation network model training module is configured to introduce an improved semi-soft threshold function to construct a deep residual shrinkage network, extract the data features of the source domain with known labels for a certain working condition and the target domain with unknown labels for other working conditions, and suppress noise interference; combined with the MMD domain adaptation method, by calculating the maximum mean difference between the source domain and the target domain, the classification loss and domain loss are minimized to reduce the distribution difference and label difference between different working conditions, and the network is continuously trained;

[0204] The harmonic reducer multi-state recognition module is configured to input the target domain test samples into the trained network model, and finally obtain the health status diagnosis results of the industrial robot harmonic reducer under different working conditions.

[0205] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments are contemplated within the scope of the invention thus described. The disclosure of the present invention is intended to be illustrative rather than restrictive of the scope of the invention, which is defined by the appended claims.

[0206] The documents cited in the present invention are as follows:

[0207] [1] Chen Renxiang, Zhang Yong, Hu Xiaolin, et al. Health status assessment of harmonic reducer based on deep fusion of multi-sensor information[J]. Vibration and Shock, 2022, 41(7): 139-144+152.

[0208] [2] Kang Shouqiang, Zhang Weidong, Wang Yujing, et al. Fault diagnosis method of harmonic reducer under different working conditions based on information fusion subdomain adaptation[J]. Chinese Journal of Scientific Instrument, 2024, 45(3): 60-71.

[0209] [3]Jia Y,Li Y,Xu M,et al.A fault diagnosis scheme for harmonicreducer under practical operating conditions[J].Measurement,2024,227:114234.

[0210] [4]Sun S,Ding H.Fault Feature Extraction Based on Unsupervised GraphEmbedding for Harmonic Reducers Diagnosis[C] / / International Conference onMechanical System Dynamics.Singapore:Springer Nature Singapore,2023:3375-3389.

[0211] [5]Long J,Mou J,Zhang L,et al.Attitude data-based deep hybridlearning architecture for intelligent faultdiagnosis ofmulti-jointindustrialrobots[J].Journal ofmanufacturing systems,2021,61:736-745.

[0212] [6]Liu L,Zhi Z,Yang Y,et al.Harmonic reducer fault detection withacoustic emission[J].IEEE Transactions on Instrumentation and Measurement,2023.

[0213] [7]Zhi Z,Liu L S,Liu D T,et al.Fault detection ofthe harmonic reducerbased on CNN-LSTM with a novel denoising algorithm[J].IEEE Senseors Journal,2022,22(03):2572-2581.

[0214] [8]Zhou

[0215] [9]He YM, Chen JH, Zhou X, et al. In-situ fault diagnosis for the harmonic reducer of industrial robots via multi-scale mixed convolutional neural networks [J]. Journal of Manufacturing Systems, 2023, 66: 233-247.

[0216]

[10] Hou Zhaoguo, Wang Huawei, Xiong Minglan, et al. Gearbox fault diagnosis based on transfer learning and weighted multi-channel fusion [J]. Vibration and Shock, 2024, 43(10): 106-113.

[0217]

[11] Zhou Xiangqi, Fu Zhongguang, Gao Yucai. Application of unsupervised domain adaptation transfer learning in rotating machinery fault diagnosis[J]. Vibration and Shock, 2024, 43(10): 106-113.

[0218]

[12] Lu N, Xiao H, Sun Y, et al. A new method for intelligent faultdiagnosis of machines based on unsupervised domain adaptation[J]. Neurocomputing, 2021, 427: 96-109.

[0219]

[13] Wu Jingran, Liu Jianhua, Cui Ran. Unsupervised bearing fault diagnosis based on subdomain adaptation[J]. Vibration and Shock, 2021.40(15): 34-40.

[0220]

[14] Wan L, Li Y, Chen K, et al. A novel deep convolution multi-adversarial domain adaptation model for bearing rolling fault diagnosis [J]. Measurement, 2022, 191: 10752.

[0221]

[15] Liu S, Qi Y, Gao X, et al. Transfer learning-based multipledigitaltwin-assisted intelligent mechanical fault diagnosis[J]. Measurement Science and Technology, 2023, 35(2): 025133.

[0222]

[16] Xie XY, Yang, ZC, Wu WH, et al. Diagnosis method for bearing based on digital twin[J]. Mathematical Problems in Engineering, 2022(2022): 115-124.

[0223]

[17] Xia Jingyan, Huang Ruyi, Chen Zhuyun, et al. Equipment small sample diagnosis method driven by fusion of twin data and feature enhancement[J]. Science China: Technological Sciences, 2023, 53(7): 1202-1213.

[0224]

[18] Zhao C, Zhang Q, Wang C, et al. Digital twin based bearing fault simulation modeling strategy and display dynamics [C] / / 2023 6th International Symposium on Autonomous Systems (ISAS). IEEE, 2023: 1-5.

[0225]

[19] Wang JR, Zhang ZZ, Liu ZL, et al.Digital twin aided adversarialtransfer learning method for domain adaptation fault diagnosis[J].ReliabilityEngineering and System Safety, 2023(234): 567-678.

[0226]

[20] Zhao MH, Zhong SS, Fu XY, et al. Deep residual shrinkage networks for fault diagnosis [J]. IEEE Transactions on Industrial Informatics, 2019, 16(7): 4681-4690.

[0227]

[21] Liu Xuzhou, Li Xiaozhong. Fault diagnosis of variable-operating bearings based on deep residual shrinkage network and transfer learning[J]. Journal of Tianjin University of Science & Technology, 2023, 38(4): 76-80.

[0228]

[22] Zhong Bowen. Dynamic simulation and fault characteristic analysis of harmonic drive system[D]. Heilongjiang: Harbin Institute of Technology, 2023.

[0229]

[23] Zhu JY, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks [C] / / Proceedings of the IEEEinternational conference on computer vision. 2017: 2223-2232.

[0230]

[24] Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C] / / Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 7132-7141.

[0231]

[25] Lei S,Lu M,Lin J,et al.Remote sensing image denoising based onimproved semi-soft threshold[J].Signal,Image And Video Processing,2021,15:73-81.

Claims

1. A fault diagnosis method for industrial robot harmonic reducer under different working conditions based on digital twin, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing A harmonic reducer vibration signal experimental platform was built as a physical entity. The one-dimensional vibration signal data of all states of different working conditions were collected using a vibration acceleration sensor, and the data were converted into a two-dimensional time-frequency image by wavelet transform to obtain a measured data set in preparation for subsequent virtual-real mapping. Step 2: Construction of harmonic deceleration virtual entity According to the relevant parameter information and motion mechanism of the physical entity of the harmonic reducer, the dynamic equations of the flexible wheel, rigid wheel and flexible bearing are established respectively. Then, by injecting faults and related speed information, the digital twin model of the harmonic reducer fault is obtained, and the simulation data of different fault types are obtained by solving. Step 3: Digital twin virtual-real mapping Construct and train the CycleGAN virtual-to-real mapping network, input the simulated data as X-domain data and the measured data as Y-domain data into the network respectively, and obtain the mapped image data through continuous iteration of cycle consistency loss and use it as the source domain sample set; Step 4: Unsupervised domain adaptation network model construction An improved semi-soft threshold function is introduced to construct a deep residual shrinkage network to extract data features of the source domain with known labels for a certain working condition and the target domain with unknown labels for other working conditions, while suppressing noise interference. Combined with the MMD domain adaptation method, the maximum mean difference between the source domain and the target domain is calculated to minimize the classification loss and domain loss to reduce the distribution differences and label differences between different working conditions, and continuously train the network. Step 5: Multi-state identification of harmonic reducer The target domain test samples are input into the trained network model, and finally the health status diagnosis results of the industrial robot harmonic reducer under different working conditions are obtained.

2. The fault diagnosis method for industrial robot harmonic reducer under different working conditions based on digital twin according to claim 1 is characterized in that: In step 2, according to the relevant parameter information and motion mechanism of the physical entity of the harmonic reducer, the dynamic equations of the flexible wheel, rigid wheel and flexible bearing are established respectively, including: The dynamic equation of the flexible wheel is expressed as: F m =K m (y out cosθ i +x out sinθ i ) Among them, K m is the meshing stiffness of the flexspline teeth, y out is the vertical displacement of the outer ring, x out is the horizontal displacement of the outer ring, θ i is the input end angle; The rigid wheel dynamic equation is expressed as: Among them, m is the mass of the rigid wheel, x c is the horizontal displacement of the rigid wheel, y c is the vertical displacement of the rigid wheel, F m F is the meshing force between the rigid wheel and the flexible wheel, t F is the tangential component of the meshing force on the outer circle of the rigid wheel, r is the axial component of the meshing force; α t Indicates the pressure angle of the bearing, k y Indicates the vertical stiffness coefficient, k x represents the horizontal stiffness coefficient; The dynamic equation of the flexible bearing is expressed as: Among them, m in is the mass of the inner ring, m out is the mass of the outer ring, c in is the connection damping between the inner ring shaft and the shaft fixing point, c out is the connection damping between the outer ring base and the ground foundation, k in k is the connection stiffness between the inner ring shaft and the shaft fixing point, out The connection stiffness between the outer ring base and the ground foundation, F HX With F HY are the components of the elastic restoring force generated by the contact between the inner and outer rings and the rolling elements in the x and y directions, e is the eccentricity, ω is the angular frequency of the inner ring rotation, g is the gravitational acceleration, and t is the time; in Indicates the horizontal displacement of the inner ring; y in Indicates the vertical displacement of the inner ring; x out Represents the horizontal displacement of the outer ring; y out represents the vertical displacement of the outer ring; θ represents the rotation angle of the input end.

3. The fault diagnosis method of industrial robot harmonic reducer under different working conditions based on digital twin according to claim 2 is characterized in that: The CycleGAN virtual-to-real mapping network described in step 3 is constructed as follows: CycleGAN is a ring network consisting of two mirror-symmetric GANs, including two generators G X→Y and G Y→X , two discriminators D X and D Y ; Throughout the process, the generator is used to generate images that are similar enough to the target domain to deceive the discriminator; the discriminator determines whether the generated image is realistic enough; the total loss function of the CycleGAN network is: L CycleGAN =L GAN (D X ,G Y→X )+L GAN (D Y ,G X→Y )+λL cyc (G X→Y ,G Y→X ) Where, L GAN (D X ,G Y→X ) represents the discriminator D X The adversarial loss function L GAN (D Y ,G X→Y ) represents the discriminator D Y The adversarial loss function L cyc (G X→Y ,G Y→X ) represents the cycle-consistent loss function; λ is the control ratio.

4. The fault diagnosis method for industrial robot harmonic reducer under different working conditions based on digital twin according to claim 3 is characterized in that: The adversarial loss function of GAN is as follows: L GAN (D X ,G Y→X )=E x～data(x) [logD X (x)]+E y～data(y) [log(1-D X (G Y→X (y)))] L GAN (D Y ,G X→Y )=E y～data(y) [logD Y (y)]+E x～data(x) [log(1-D Y (D X→Y (x)))] In the formula, E x-data(x) With E y-data(y) are the losses of two pairs of domains X and Y; The formula of the cycle-consistent loss function is expressed as: L cyc (G X→Y ,G Y→X )=E x～data(x) [||G Y→X (G X→Y (x))-x||1]+E y～data(y) [||G X→Y (G Y→X (y))-y||1] In the formula, G Y→X (G X→Y (x)) and G X→Y (G Y→X (y)) represent the domain transformation of the image by the two generators; x is the input feature of the X domain, and y is the input feature of the Y domain.

5. The fault diagnosis method of industrial robot harmonic reducer under different working conditions based on digital twin according to claim 4 is characterized in that: The deep residual shrinkage network described in step 4 is based on the deep residual network and introduces an attention mechanism and a semi-soft threshold. The attention mechanism automatically learns a set of weights through a small sub-network to weight each channel of the feature. The semi-soft threshold function is used for signal denoising.

6. The fault diagnosis method of industrial robot harmonic reducer under different working conditions based on digital twin according to claim 5 is characterized in that: The semi-soft threshold function formula is expressed as: Among them, x is the input feature, y is the output feature, and τ is the threshold.

7. The fault diagnosis method for industrial robot harmonic reducer under different working conditions based on digital twin according to claim 6 is characterized in that: In step 4, the MMD domain adaptation method is combined to calculate the maximum mean difference between the source domain and the target domain, minimize the classification loss and domain loss to reduce the distribution difference and label difference between different working conditions, and continuously train the network including: achieving cross-domain adaptation of features by minimizing the MMD distance between the source domain and the target domain. The MMD calculation formula is expressed as Where H is the regenerated Hilbert space, φ(·) is the mapping function, and n s and n t are the number of samples in the source domain and the target domain respectively; X s represents the probability distribution of the source domain, X t Represents the probability distribution of the target domain; The MMD domain adaptation method includes a feature extractor F and a classifier G; wherein source domain labeled data and target domain unlabeled data are respectively input, and the fault features of the source domain labeled data and the target domain unlabeled data are extracted through the feature extractor F; in the classifier G, the fault features are input, and after passing through a fully connected layer, the difference between the source domain and the target domain is reduced by calculating the MMD loss of the source domain labeled data and the target domain unlabeled data, and then the classification information is output and the classification loss is obtained; and the iterative training model is continuously updated by minimizing the classification loss of the source domain labeled samples and minimizing the MMD loss between the source domain and the target domain samples.

8. The fault diagnosis method for industrial robot harmonic reducer under different working conditions based on digital twin according to claim 7 is characterized in that: The calculation formula of the classification loss described in step 4 is expressed as: Among them, K is the number of classification categories, n s is the number of source domain samples, represents the true label of source domain sample i in category k, It represents the probability that the source domain sample i belongs to category k predicted by the model.

9. The fault diagnosis method of industrial robot harmonic reducer under different working conditions based on digital twin according to claim 1 is characterized in that: All states of different working conditions described in step 1 include: mild, moderate and severe faults of the inner ring of the flexible bearing; mild, moderate and severe faults of the outer ring of the flexible bearing; severe fault of the rolling body weight; rigid wheel missing tooth fault; flexible wheel missing tooth fault; normal state.

10. The industrial robot harmonic reducer fault diagnosis system under different working conditions based on digital twin is characterized by: include: The data acquisition and preprocessing module is configured to build a harmonic reducer vibration signal experimental platform as a physical entity, use a vibration acceleration sensor to collect one-dimensional vibration signal data of all states of different working conditions, and perform wavelet transform on it to convert it into a two-dimensional time-frequency image to obtain a measured data set, in preparation for subsequent virtual-real mapping; The harmonic reducer virtual entity construction module is configured to establish the dynamic equations of the flexible wheel, rigid wheel and flexible bearing according to the relevant parameter information and motion mechanism of the harmonic reducer physical entity, and then obtain the digital twin model of the harmonic reducer fault by injecting fault and related speed information, and solve to obtain simulation data of different fault types; The digital twin virtual-to-real mapping module is configured to build and train the CycleGAN virtual-to-real mapping network, input the simulated data as the X-domain data and the measured data as the Y-domain data into the network respectively, and obtain the mapped image data through continuous iteration of cycle consistency loss as the source domain sample set; The unsupervised domain adaptation network model training module is configured to introduce an improved semi-soft threshold function to construct a deep residual shrinkage network, extract the data features of the source domain with known labels for a certain working condition and the target domain with unknown labels for other working conditions, and suppress noise interference; combined with the MMD domain adaptation method, by calculating the maximum mean difference between the source domain and the target domain, the classification loss and domain loss are minimized to reduce the distribution difference and label difference between different working conditions, and the network is continuously trained; The harmonic reducer multi-state recognition module is configured to input the target domain test samples into the trained network model, and finally obtain the health status diagnosis results of the industrial robot harmonic reducer under different working conditions.

Images