1. The harmonic reducer fault diagnosis method based on the generative countermeasure network is characterized by comprising the following steps:

s1, preprocessing data, collecting vibration acceleration signals of the harmonic reducer through three vibration acceleration sensors, processing the data by using Fast Fourier Transform (FFT) to extract the characteristics of original signals, and constructing an original data set by using normalized data;

s2, generating data, enhancing fault data by using a generating type countermeasure network GAN composed of a convolution layer and a full connection layer, and generating multiple types of fault data by using multiple generating type countermeasure networks GAN;

s3, selecting data, filtering and purifying the generated data by using a data selection module consisting of data filtering and data purification, and screening the generated data;

and S4, fault classification is carried out, a new balanced data set is formed by utilizing the real data and various generated fault data, and the multi-scale convolutional neural network MSCNN formed by multi-scale decomposition and convolutional neural network CNN is used as a classifier to carry out multi-classification of harmonic reducer faults.

2. The harmonic reducer fault diagnosis method based on the generative countermeasure network as claimed in claim 1, wherein the step S1 is implemented as follows:

s11, collecting vibration acceleration signals of the harmonic reducer through three vibration acceleration sensors, and resampling the vibration acceleration signals to N times of the original number by using a movable and fixed sampling window;

s12, extracting the characteristic information of the resampled vibration acceleration signal by using Fast Fourier Transform (FFT) to form an original data set;

s13, normalizing each data sample according to the formulas (1) and (2), and converting the data samples into 0-1 intervals to accelerate the convergence rate of the training network, wherein the specific formula is as follows:

wherein Result is a pixel Value, Value is each data Value on the signal, Range is a maximum Range Value of the signal, and scaled Value is a normalized Value.

3. The harmonic reducer fault diagnosis method based on the generative countermeasure network as claimed in claim 2, wherein in step S11, three vibration acceleration sensors are installed at the end of the harmonic reducer, and the sensors are orthogonal in pairs and installed in X, Y, Z three mutually orthogonal directions respectively.

4. The method as claimed in claim 1, wherein each generative countermeasure network GAN includes a generator and a discriminator in step S2, and the subsequent data selection is performed by generating one-dimensional fault data of one type, and adjusting the fault data generated by various fault types and the number of different fault types using data rolling and averaging operations.

5. The harmonic reducer fault diagnosis method based on the generative countermeasure network as claimed in claim 1, wherein the step S3 is implemented as follows:

s31, filtering original data outside the real data distribution through the Mahalanobis distance, and setting the sample of the gamma small category in the real data as:

S_γ＝(S_{γ_1}，S_{γ_2}，…，S_{γ_m})

γ＝(1,2，…，N)

wherein m is the number of samples;

sample S_{γ_i}(i ═ (1,2, …, m)) and dataset S_γThe mahalanobis distance between can be calculated as follows:

wherein the content of the first and second substances,representing the centroid of the real data space as the average value of each dimension; covariance matrix cov (S)_γ) Reversible;

using Principal Component Analysis (PCA) to reduce the dimension of the variable, and setting a minimum sample space for real data to determine generated data;

s32, further purifying the filtered data by using Euclidean distance, and obtaining two points y₁And y₂The expression of the euclidean distance between is as follows:

wherein u is the total dimension of a certain point; k refers to the kth dimension of the point;

s33, building a balanced data set by using the generated data which is close to the true distribution and is reserved in the steps S1 and S2 for subsequent training.

6. The harmonic reducer fault diagnosis method based on the generative countermeasure network as claimed in claim 1, wherein the multi-scale decomposition in step S4 is implemented as follows:

s41, extracting the complementation and characteristic information of each scale signal by utilizing multi-scale decomposition, and setting x_iIs the original signal x ═ x₁，x₂，…，x_NH value of, where N is the original sample total number; obtaining a continuous signal { y (S) } from the original signal by using a multiscale factor S; with a four-layer scale decomposition, each element of the scale signal is represented as:

when the value of the multi-scale factor is S, the total number of samples in the scale is N/S; j represents the jth sample out of the total number of samples at the scale; y is_j ^(s)When the scale factor is S, j point data in the total sample under the scale is represented;

s42, according to the size of the input unit of the convolutional network and the multi-scale factor coefficient, respectively intercepting 3072, 1536, 1024 and 716 continuous data of four scales to form each unit, and then sequentially connecting the four units to form each two-dimensional convolutional neural network 2D-CNN input unit to perform subsequent convolutional neural network CNN training.

7. The harmonic reducer fault diagnosis system based on the generative countermeasure network is characterized by comprising a generative countermeasure network GAN model, a fault data generation module, an industrial scene module, a real data collection module, a local database module, a remote server, a balanced data set module, a multi-scale convolutional neural network MSCNN model and a fault diagnosis result module;

the method comprises the following steps of off-line diagnosis, wherein a generative confrontation network GAN model is connected with a fault data generation module, and the generative confrontation network GAN model transmits generated fault data into the fault data generation module;

the industrial scene module is connected with the real data collecting module, real data collected through an industrial scene are transmitted to the real data collecting module, the real data collecting module transmits the real data to the local database through wireless connection, a new balance data set is formed by the generated fault data and the real data in the local database and is input into the balance data set module, and a fault diagnosis result is obtained through the multi-scale convolutional neural network MSCNN model;

and in on-line diagnosis, the industrial scene module is connected with the real data collecting module, real data are uploaded to a remote server for storage through wired network connection, and then a fault diagnosis result is obtained through a multi-scale convolutional neural network MSCNN model.

8. The harmonic reducer fault diagnosis system based on the generative countermeasure network as claimed in claim 7, wherein the online diagnosis operation shares the collected real data with the local database through the remote server, updates the local database and optimizes the MSCNN model of the multi-scale convolutional neural network.

Background

Harmonic reducers are widely used in industrial robots to increase the amount of torque transmitted to or from one shaft end due to their advantages of high gear ratio, backlash free, high compactness and lightness, good resolution and excellent repeatability. The complex design of the harmonic reducer is sensitive to manufacturing and assembly errors, and abnormal vibrations are related to the operating conditions, and even small errors can result in excessive vibrations, thereby compromising the performance of the robot. In addition, harmonic reducers are highly nonlinear systems that are typically coupled to other external electromechanical systems, and thus the vibration signal typically exhibits multi-scale characteristics. Most importantly, the fault types of the harmonic reducer are complex and various, and fault signals are difficult to collect, so that different types of health state data are unbalanced, and the diagnostic performance of the harmonic reducer is adversely affected.

The fault diagnosis of existing harmonic drives is mainly done by shop technicians using simple instruments and relies on technical experience, resulting in the decision being subjective and often inaccurate. There is little research and satisfactory results associated with harmonic reducer condition monitoring and fault diagnosis. The existing documents only research the components of the harmonic reducer, such as a reduction gear and a bearing, but not the whole harmonic reducer finished product, so that the real operation condition of the whole harmonic reducer is difficult to reflect, and the existing time-frequency domain analysis method is used, so that the diagnosis effect needs to be improved. Various machine learning methods based on data balance are not suitable for harmonic reducers because they must have a large amount of data of various types in a well-balanced manner, although they can achieve high classification accuracy in fault diagnosis of mechanical devices such as roller bearings.

Disclosure of Invention

In order to solve the technical problems in the prior art, the invention provides a harmonic reducer fault diagnosis method and system based on a generative countermeasure network.

The method is realized by adopting the following technical scheme: the harmonic reducer fault diagnosis method based on the generative countermeasure network mainly comprises the following steps:

s1, preprocessing data, collecting vibration acceleration signals of the harmonic reducer through three vibration acceleration sensors, processing the data by using Fast Fourier Transform (FFT) to extract the characteristics of original signals, and constructing an original data set by using normalized data;

s2, generating data, enhancing fault data by using a generating type countermeasure network GAN composed of a convolution layer and a full connection layer, and generating multiple types of fault data by using multiple generating type countermeasure networks GAN;

s3, selecting data, filtering and purifying the generated data by using a data selection module consisting of data filtering and data purification, and screening the generated data;

and S4, fault classification is carried out, a new balanced data set is formed by utilizing the real data and various generated fault data, and the multi-scale convolutional neural network MSCNN formed by multi-scale decomposition and convolutional neural network CNN is used as a classifier to carry out multi-classification of harmonic reducer faults.

The system of the invention is realized by adopting the following technical scheme: the harmonic reducer fault diagnosis system based on the generative countermeasure network comprises a generative countermeasure network GAN model, a fault data generation module, an industrial scene module, a real data collection module, a local database module, a remote server, a balanced data set module, a multi-scale convolutional neural network MSCNN model and a fault diagnosis result module;

the method comprises the following steps of off-line diagnosis, wherein a generative confrontation network GAN model is connected with a fault data generation module, and the generative confrontation network GAN model transmits generated fault data into the fault data generation module;

the industrial scene module is connected with the real data collecting module, real data collected through an industrial scene are transmitted to the real data collecting module, the real data collecting module transmits the real data to the local database through wireless connection, a new balance data set is formed by the generated fault data and the real data in the local database and is input into the balance data set module, and a fault diagnosis result is obtained through the multi-scale convolutional neural network MSCNN model;

and in on-line diagnosis, the industrial scene module is connected with the real data collecting module, real data are uploaded to a remote server for storage through wired network connection, and then a fault diagnosis result is obtained through a multi-scale convolutional neural network MSCNN model.

Compared with the prior art, the invention has the following advantages and beneficial effects:

1. according to the method, a large amount of high-quality fault data of various types of harmonic reducers are generated through the generation type countermeasure network GAN, a balanced data set is constructed by the generated fault data and real data, and then fault diagnosis is carried out by using the multi-scale convolutional neural network MSCNN, so that the multi-classification precision of the harmonic reducers is improved under the condition of data imbalance.

2. The invention can be used for solving the problem of the insufficiency of various fault samples and improving the accuracy of fault diagnosis.

3. The invention can be applied to and expanded to different mechanical transmission devices such as a fan gearbox, a heading machine slewing bearing, an aircraft engine electromechanical actuator and the like, thereby solving the problem of fault diagnosis under the condition of scarce fault data.

Drawings

FIG. 1 is a flow chart of a method of the present invention;

FIG. 2 is a schematic diagram of Fast Fourier Transform (FFT) processing data of the method of the present invention;

FIG. 3 is a schematic view of the test stand of the present invention;

FIG. 4 is a schematic diagram of a data generation process of the method of the present invention;

FIG. 5 is a schematic diagram of a network architecture of multiple generators of the method of the present invention;

FIG. 6 is a schematic diagram of a data selection process of the method of the present invention;

FIG. 7 is a schematic diagram of a fault classification process of the method of the present invention;

FIG. 8 is a schematic diagram of the data fusion transformation of the multi-scale signal to each bitmap input according to the present invention;

fig. 9 is a schematic diagram of the system structure of the invention.

Detailed Description

The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited thereto.

Examples

As shown in fig. 1, the method for diagnosing a fault of a harmonic reducer based on a generative countermeasure network of the present embodiment mainly includes the following steps:

s1, preprocessing data, collecting vibration acceleration signals of the harmonic reducer through three vibration acceleration sensors, processing the data through Fast Fourier Transform (FFT) to extract characteristics of original signals, and constructing an original data set by using normalized data for subsequent data generation;

s2, generating data, enhancing various scarce fault data by using a generating type countermeasure network GAN composed of a convolution layer and a full connection layer, and generating multi-type fault data by using a plurality of generating type countermeasure networks GAN;

s3, selecting data, and screening the generated data by filtering and purifying the generated data by using a data selection module consisting of data filtering and data purifying to improve the quality of the generated data;

and S4, fault classification is carried out, a new balanced data set is formed by utilizing the real data and various generated fault data, and the multi-scale convolutional neural network MSCNN formed by multi-scale decomposition and convolutional neural network CNN is used as a classifier to carry out multi-classification of harmonic reducer faults.

As shown in fig. 2, in this embodiment, the specific implementation process of step S1 is as follows:

s11, collecting vibration acceleration signals of the harmonic reducer through three vibration acceleration sensors, and resampling the vibration acceleration signals to N times of the original number by using a movable and fixed sampling window, wherein N is greater than 1, and N is an integer;

s12, extracting the characteristic information of the resampled vibration acceleration signal by using Fast Fourier Transform (FFT) to form an original data set for subsequent data generation and fault diagnosis;

s13, normalizing each data sample according to the formulas (1) and (2), and converting the data samples into 0-1 intervals to accelerate the convergence rate of the training network, wherein the specific formula is as follows:

wherein Result is a pixel Value, Value is each data Value on the signal, Range is a maximum Range Value of the signal, and scaled Value is a normalized Value.

As shown in fig. 3, in the present embodiment, in step S11, three vibration acceleration sensors are all installed at the end of the harmonic reducer, and the sensors are orthogonal in pairs and are respectively installed in X, Y, Z three mutually orthogonal directions.

As shown in fig. 4 and 5, in the present embodiment, each generative countermeasure network GAN in step S2 includes a generator and a discriminator for generating one-dimensional fault data of one type for subsequent data selection, and the generated fault data of various fault types and the number of different fault types are adjusted by using data rolling and averaging operations.

As shown in fig. 6, in this embodiment, the specific implementation process of step S3 is as follows:

s31, filtering original data outside the real data distribution through the Mahalanobis distance, and setting the sample of the gamma small category in the actual data as:

S_γ＝(S_{γ_1}，S_{γ_2}，…，S_{γ_m})

γ＝(1,2，…，N)

wherein m is the number of samples;

sample S_{γ_i}(i ═ (1,2, …, m)) and dataset S_γThe mahalanobis distance between can be calculated as follows:

wherein the content of the first and second substances,representing the centroid of the real data space as the average value of each dimension; covariance matrix cov (S)_γ) It needs to be reversible.

Using Principal Component Analysis (PCA) to reduce the dimension of the variable and setting a minimum sample space for actual data to determine whether the generated data should be discarded;

s32, further purifying the filtered data by using Euclidean distance, and obtaining two points y₁And y₂The expression of the euclidean distance between is as follows:

wherein u is the total dimension of a certain point, k is the kth dimension of the point, 1< k < u, and k is an integer;

s33, building a balanced data set by using the generated data which is close to the true distribution and is reserved in the steps S1 and S2 for subsequent training.

As shown in fig. 7, in this embodiment, a specific implementation process of the multi-scale decomposition in step S4 is as follows:

s41, extracting complementary and rich characteristic information of signals with different scales by utilizing multi-scale decomposition, and setting x_iIs the original signal x ═ x₁，x₂，…，x_NH value of, where N is the total number of samples; obtaining a continuous signal { y (S) } from the raw signal by using a multiscale factor S; with a four-layer scale decomposition, each element of the different scale signals is represented as:

when the value of the multi-scale factor is S, the total number of samples in the scale is N/S; j represents the jth sample out of the total number of samples at the scale; js represents the abbreviation for the multiplication of both j and s; y is_j ^(s)When the scale factor is S, j point data in the total sample under the scale is represented;

s42, according to the input unit size and the multi-scale factor coefficient of the convolutional network, as shown in fig. 8, respectively intercepting 3072, 1536, 1024, 716 consecutive data of the four-scale signals to form each unit, and then sequentially connecting the four units to construct each two-dimensional convolutional neural network 2D-CNN input unit for subsequent convolutional neural network CNN training.

In this embodiment, additional gaussian noise based on different SNR is added to the test data for evaluation to check the robustness of the method of the present invention to environmental noise, which is defined as:

wherein, P_signalIs the power of the original signal; p_noiseThe power of the additive gaussian noise.

As shown in fig. 9, the present invention also provides a harmonic reducer fault diagnosis system based on a generative countermeasure network, which can be applied to off-line and on-line diagnosis, and comprises: the system comprises a generating type confrontation network GAN model, a fault data generating module, an industrial scene module, a real data collecting module, a local database module, a remote server, a balanced data set module, a multi-scale convolutional neural network MSCNN model and a fault diagnosis result module;

the method comprises the following steps of off-line diagnosis, wherein a generative confrontation network GAN model is connected with a fault data generation module, and the generative confrontation network GAN model transmits generated fault data into the fault data generation module;

the industrial scene module is connected with the real data collecting module, real data collected through an industrial scene are transmitted to the real data collecting module, the real data collecting module transmits the real data to the local database through wireless connection, a new balance data set is formed by the generated fault data and the real data in the local database and is input into the balance data set module, and a fault diagnosis result is obtained through the multi-scale convolutional neural network MSCNN model;

and in on-line diagnosis, the industrial scene module is connected with the real data collecting module, real data are uploaded to a remote server for storage through wired network connection, and then a fault diagnosis result is obtained through a multi-scale convolutional neural network MSCNN model.

Specifically, the online diagnosis operation can share the collected real data with a local database through a remote server, update the local database and optimize the multi-scale convolutional neural network MSCNN model, so that the online diagnosis operation is more suitable for the diagnosis of equipment. Therefore, the diagnostic system of the embodiment can collect and display the operation data of the first-out equipment in real time, realize the fault diagnosis of harmonic transmission by depending on a network model, and provide adjustment feedback for the equipment.

The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.