CN114323644B

CN114323644B - Gear box fault diagnosis and signal acquisition method and device and electronic equipment

Info

Publication number: CN114323644B
Application number: CN202210244235.XA
Authority: CN
Inventors: 苏营; 秦玉文; 邹祖冰; 甘富航; 邓友汉; 王罗; 高远
Original assignee: China Three Gorges Corp
Current assignee: China Three Gorges Corp
Priority date: 2022-03-14
Filing date: 2022-03-14
Publication date: 2022-06-03
Anticipated expiration: 2042-03-14
Also published as: CN114323644A; DE112022000106T5; WO2023020388A1

Abstract

The invention discloses a method, a device and electronic equipment for diagnosing faults and acquiring signals of a gear box, wherein the diagnosing method comprises the steps of obtaining real signals of the gear box and analog signals generated by the acquiring method, and the real signals comprise real torque signals, real vibration signals and real rotation signals; respectively inputting the analog signal and the real signal into a preset fault diagnosis network to obtain a fault diagnosis result of the gearbox; the preset fault diagnosis network comprises a first neural network used for feature extraction and a second neural network with a hierarchical attention mechanism and used for outputting fault diagnosis results, and is generated based on analog signals and real signals in a training mode. According to the technical scheme provided by the invention, the data quality of the sensor is improved, and the accuracy of fault diagnosis of the gearbox is improved.

Description

Gear box fault diagnosis and signal acquisition method and device and electronic equipment

Technical Field

The invention relates to the field of algorithm design, in particular to a method and a device for diagnosing faults of a gearbox and acquiring signals and electronic equipment.

Background

The wind turbine generator works in severe environment conditions such as sand storm, thunder and rainstorm and the like all the year round, and the requirements on operation and maintenance of the wind turbine generator are more and more strict in order to ensure the reliability and safety of key components of the wind turbine generator. The gear box is a key transmission device of the wind turbine generator, long time is consumed for maintenance, replacement and repair, the cost is high, and meanwhile, the health state of the gear box is closely related to the stable operation of the wind turbine generator. In recent years, researchers widely adopt a signal analysis method to realize fault diagnosis of the gearbox, the method mostly utilizes time-domain statistical characteristics, wavelet transformation, fast Fourier transformation, empirical mode decomposition and the like to extract characteristics from fault signals of the gearbox, and further realizes fault diagnosis, and research results show that the method can obtain higher accuracy under a stable condition. In practical situations, however, the working conditions of the gearbox are complex and changeable, signal components are diversified, and multiple modes are mixed, so that the signals have non-stationarity. The problems of false component interference and low resolution of characteristic signals can occur in a common signal analysis method, and the fault state of the gearbox cannot be accurately judged. Meanwhile, a great deal of environmental noise is mixed in the final acquisition signal of the sensor, so that the signal-to-noise ratio is low. In addition, the existing sensor of the wind field can only collect characteristic signals of wind speed, voltage, current and the like, and cannot collect or accurately measure physical quantities of component stress, acceleration and the like, so that the fault state of the gearbox cannot be accurately reflected. Therefore, improving sensor data quality is critical to achieving accurate fault diagnosis of the gearbox.

The development of the digital twin technology provides an effective strategy for the real-time state evaluation of the gearbox in a changeable environment, and the fault evolution of the gearbox in the actual working condition operation is simulated by establishing a virtual simulation model, so that the high-precision simulation of the gearbox can be kept with the actual physical gearbox. By means of multi-physics and multi-scale comprehensive simulation of the gear box, the states of all systems of the gear box can be accurately monitored, and twin system collected data can be output in real time. And finally, model data and sensor data are fused, so that the defect that a physical sensor acquires data is effectively overcome. At present, developers establish a gearbox fault diagnosis model based on a digital twin technology and a deep neural network for fault detection (refer to a patent document CN 113505655A), and certain effects are achieved. However, most of the existing digital twin models are created based on a three-dimensional model platform, the models are too simple and convenient, the analyzed signal data are single, the characterization degree of the simulated fault data on the fault is not deep enough, the data quality is still to be improved, and the final accuracy of the fault diagnosis of the gearbox is influenced. Therefore, how to further improve the quality of the sensor data so as to improve the accuracy of the fault diagnosis of the gearbox is an urgent problem to be solved.

Disclosure of Invention

In view of this, embodiments of the present invention provide a method and an apparatus for diagnosing a fault of a gearbox and acquiring a signal, and an electronic device, so as to improve the quality of sensor data and improve the accuracy of diagnosing a fault of a gearbox.

According to a first aspect, the present invention provides a gearbox signal acquisition method, the method comprising: establishing a digital twin model based on material parameters and geometric dimension parameters of a gear in the gearbox; acquiring a real torque signal of the gearbox, wherein the real torque signal comprises first signal data used for correcting the parameters of the digital twin model and second signal data used for calculating the analog signal; inputting the first signal data into the digital twin model to correct parameters of the digital twin model; inputting the second signal data into the corrected digital twin model to generate an analog signal, the analog signal comprising an analog vibration signal and an analog rotation signal.

Optionally, the establishing of the digital twin model based on the material parameters and the geometric parameters of the gear in the gearbox comprises: respectively acquiring material parameters and geometric dimension parameters of a driving wheel and a driven wheel in the gearbox, wherein the material parameters comprise at least one of mass parameters, inertia parameters, rigidity parameters and damping parameters; creating a differential kinetic equation of the gearbox by utilizing the material parameters and the geometric dimension parameters based on a physical structure meshed between the driving wheel and the driven wheel; calculating a simulation matrix for outputting the simulation signal through the gearbox differential kinetic equation; establishing a mathematical model for representing a gear load relation by using the simulation matrix, the mass inertia matrix, the damping matrix and the rigidity matrix, and taking the mathematical model as the digital twin model; wherein the mass inertia matrix is constructed based on the mass parameters and the inertia parameters, the damping matrix is constructed based on the damping parameters and the geometric parameters, and the stiffness matrix is constructed based on the stiffness parameters and the geometric parameters.

According to a second aspect, the present invention provides a gearbox fault diagnosis method, the method comprising: acquiring real signals of the gearbox and the simulation signals generated according to any one of the alternative embodiments of the first aspect, wherein the real signals comprise real torque signals, real vibration signals and real rotation signals; respectively inputting the analog signal and the real signal into a preset fault diagnosis network to obtain a fault diagnosis result of the gearbox; the preset fault diagnosis network comprises a first neural network used for feature extraction and a second neural network with a hierarchical attention mechanism and used for outputting fault diagnosis results, and the preset fault diagnosis network is generated based on the simulation signal and the real signal training.

Optionally, the step of performing feature extraction on the target signal based on the first neural network by using the simulated signal and/or the real signal as the target signal includes: splicing the target signals into a two-dimensional plane vector by taking the number of sampling time points of the target signals and the number of sensors as two-dimensional plane axes; performing one-dimensional convolution calculation on the two-dimensional plane vector based on a plurality of one-dimensional convolution cores with the same dimension to obtain a plurality of one-dimensional original features; pooling each one-dimensional original feature to obtain a plurality of one-dimensional time sequence features; performing two-dimensional splicing on the plurality of one-dimensional time sequence characteristics to obtain a new two-dimensional plane vector, and adjusting the number and the dimension of the one-dimensional convolution kernels; taking the new two-dimensional plane vector as the two-dimensional plane vector, returning the two-dimensional plane vector to the one-dimensional convolution kernel based on the multiple same dimensions for one-dimensional convolution calculation based on the adjusted one-dimensional convolution kernel to obtain multiple one-dimensional original features, and obtaining multiple target time sequence features until the preset times are returned; splicing the target time sequence characteristics into a target one-dimensional vector, inputting the target one-dimensional vector into a preset full connection layer, and converting the target one-dimensional vector into a plurality of one-dimensional abstract characteristics with the same number as the sampling time points through the preset full connection layer, wherein the number of elements of the one-dimensional abstract characteristics is the same as the number of the sensors.

Optionally, the performing, based on a plurality of one-dimensional convolution cores with the same dimensionality, one-dimensional convolution calculation on the two-dimensional plane vector to obtain a plurality of one-dimensional original features includes: extracting multiple groups of elements with preset sampling time points from the two-dimensional plane vector in a sliding manner aiming at each sensor element, and forming the multiple groups of elements with the preset sampling time points into multiple one-dimensional vectors to be convolved; carrying out convolution operation on each one-dimensional vector to be convolved and the current one-dimensional convolution kernel in sequence to generate current one-dimensional original characteristics; and traversing all the one-dimensional convolution kernels until the one-dimensional original features corresponding to the one-dimensional convolution kernels are obtained based on the step of sequentially carrying out convolution operation on the one-dimensional vectors to be convolved and the current one-dimensional convolution kernel to generate the current one-dimensional original features.

Optionally, the step of outputting a fault diagnosis result based on the second neural network includes: performing level attention processing on the plurality of one-dimensional abstract features based on a linear perception machine, and fusing the plurality of one-dimensional abstract features to obtain an output vector; and inputting the output vector into a preset classifier so that the preset classifier outputs the fault diagnosis result.

Optionally, performing hierarchical attention processing on the plurality of one-dimensional abstract features based on a linear perceptron, and fusing the plurality of one-dimensional abstract features to obtain an output vector, including: inputting the plurality of one-dimensional abstract features into the linear perceptron in sequence to obtain a plurality of hidden features; calculating a plurality of weight coefficients corresponding to the plurality of hidden features through a softmax function, wherein the weight coefficients are used for representing the hierarchical attention of the hidden features; and carrying out weighted summation on the hidden features by utilizing the weighting coefficients to obtain the output vector.

According to a third aspect, the present invention provides a gearbox fault diagnosis device, the device comprising: a signal acquisition module for acquiring real signals of the gearbox and the simulated signals generated as in any one of the alternative embodiments of the first aspect, the real signals comprising real torque signals, real vibration signals and real rotation signals; the fault diagnosis module is used for respectively inputting the analog signal and the real signal into a preset fault diagnosis network to obtain a fault diagnosis result of the gearbox; the preset fault diagnosis network comprises a first neural network used for feature extraction and a second neural network with a hierarchical attention mechanism and used for outputting fault diagnosis results, and the preset fault diagnosis network is generated based on the simulation signal and the real signal training.

According to a fourth aspect, an embodiment of the present invention provides an electronic device, including: a memory and a processor, the memory and the processor being communicatively coupled to each other, the memory having stored therein computer instructions, and the processor performing the method of the first aspect, or any one of the optional embodiments of the first aspect, by executing the computer instructions.

According to a fifth aspect, an embodiment of the present invention provides a computer-readable storage medium, which stores computer instructions for causing a computer to thereby perform the method of the first aspect, or any one of the optional implementation manners of the first aspect.

The technical scheme provided by the application has the following advantages:

according to the technical scheme, a digital model of the digital twin model is established based on material parameters and geometric dimension parameters of a gear in the gear box, and then the parameters of the digital twin model are corrected through simulation in combination with a real torque signal of the gear box. Compared with a digital twin model in the prior art, the method better conforms to the actual working condition of the gearbox, so that the data quality of the analog signal better conforms to the real situation. Therefore, the real torque signal is used as the input of the model, the simulated vibration signal and the simulated rotation signal are output, the fault data type of the gear box is expanded, the simulated signal with higher data quality is obtained, fault diagnosis is carried out based on the real signal data and the simulated signal data of the gear box, and the fault diagnosis accuracy of subsequent fault diagnosis work of the gear box is improved.

In addition, the invention realizes fault diagnosis by utilizing the first neural network for feature extraction and the second neural network with a hierarchical attention mechanism for outputting fault diagnosis results during fault diagnosis, so that useful sample features are further strengthened, useless sample features are weakened, and the fault diagnosis accuracy is further improved. The method comprises a first neural network characteristic extraction step, a second neural network characteristic extraction step and a third neural network characteristic extraction step, wherein the first neural network characteristic extraction step is to convert fault signals acquired by a plurality of sensors from two-dimensional data into one-dimensional data, then carry out one-dimensional convolution, and then repeatedly splice the obtained one-dimensional characteristics into the two-dimensional data to carry out multi-level one-dimensional, two-dimensional and one-dimensional characteristic extraction. Finally, the signal features which are more prominent in features and have the same dimension as the original data are obtained. Therefore, the classification of subsequent fault signals is further easier, and the diagnosis accuracy is improved. The characteristic vectors are subjected to level attention processing based on a linear perceptron, and nonlinearity of a neural network is increased, so that degradation characteristics of the gearbox are fitted better, a fault diagnosis result is closer to a real situation, and the accuracy of fault diagnosis of the gearbox is improved.

Drawings

The features and advantages of the present invention will be more clearly understood by reference to the accompanying drawings, which are illustrative and not to be construed as limiting the invention in any way, and in which:

FIG. 1 is a schematic diagram illustrating the steps of a gearbox signal acquisition method according to an embodiment of the present invention;

FIG. 2 illustrates a block flow diagram of a gearbox signal acquisition method in one embodiment of the present invention;

FIG. 3 is a schematic diagram showing the meshing dynamics of a typical spur gear pair in the prior art;

FIG. 4 is a schematic diagram illustrating the steps of a gearbox fault diagnosis method in one embodiment of the present invention;

FIG. 5 is a schematic flow chart diagram illustrating a gearbox fault diagnosis method in accordance with an embodiment of the present invention;

FIG. 6 illustrates a schematic structural diagram of a first neural network in one embodiment of the present invention;

FIG. 7 is a schematic flow chart diagram illustrating another gearbox fault diagnosis method in accordance with one embodiment of the present invention;

FIG. 8 is a graph illustrating a depth-time diagnostic network loss function as a function of iteration number in one embodiment of the present invention;

FIG. 9 is a schematic structural diagram of a gearbox signal acquisition device according to an embodiment of the invention;

fig. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings of the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without inventive step based on the embodiments of the present invention, are within the scope of the present invention.

Referring to fig. 1 and fig. 2, in an embodiment, a method for acquiring a signal of a gearbox includes the following steps:

step S101: and establishing a digital twin model based on the material parameters and the geometric dimension parameters of the gear in the gearbox.

Step S102: acquiring a real torque signal of the gearbox, wherein the real torque signal comprises first signal data used for correcting parameters of the digital twin model and second signal data used for calculating an analog signal, and inputting the first signal data into the digital twin model to correct the parameters of the digital twin model.

Step S103: inputting the second signal data into the corrected digital twin model to generate an analog signal, the analog signal including an analog vibration signal and an analog rotation signal.

Specifically, in the present embodiment, in order to further improve the signal quality of the analog fault signal, compared to the conventional method of creating a digital twin model by simply using a three-dimensional model, the digital twin model of the gearbox is created based on the material parameters and the geometric parameters of the gears in the gearbox, including but not limited to the gear size, the gear material, the material stiffness, the inter-gear damping, and the gear mass. Therefore, the simulation working condition of the digital twin model is closer to the real working condition of the operation of the gearbox. The traditional digital twin model is usually an analog of a single vibration signal. In the embodiment, the geometric dimension parameters and the material parameters in the digital twin model are corrected by creating an equation relation of the vibration signal, the rotation signal and the torque signal and then using the real torque signal based on finite element simulation software. And then inputting the real torque signal as an excitation into the corrected digital twin model to obtain an excellent simulated vibration signal and an excellent simulated rotation signal. Therefore, the analog signal types output by the digital twin model are more, the quality of the analog signals is better, and the fault diagnosis accuracy of the subsequent gearbox is improved.

Specifically, in an embodiment, the step S101 specifically includes the following steps:

the method comprises the following steps: and respectively acquiring material parameters and geometric dimension parameters of a driving wheel and a driven wheel in the gearbox, wherein the material parameters comprise at least one of mass parameters, inertia parameters, rigidity parameters and damping parameters.

Step two: based on the physical structure of the engagement between the driving wheel and the driven wheel, a differential kinetic equation of the gearbox is created using the material parameters and the geometric dimension parameters.

Step three: a simulation matrix for outputting the simulation signal is calculated by a gearbox differential dynamics equation.

Step four: and creating a mathematical model for representing the gear load relationship by using the simulation matrix, the mass inertia matrix, the damping matrix and the rigidity matrix, and taking the mathematical model as a digital twin model.

The mass inertia matrix is constructed based on mass parameters and inertia parameters, the damping matrix is constructed based on damping parameters and geometric dimension parameters, and the rigidity matrix is constructed based on rigidity parameters and geometric dimension parameters.

Specifically, in the present embodiment, a mathematical equation model is created based on the gear load relationship, thereby achieving the creation of the digital twin model. Firstly, material parameters including mass parameters, inertia parameters, rigidity parameters and damping parameters are obtained, and then geometric dimension parameters representing the size of the gear are obtained. Then, a mass inertia matrix, a damping matrix and a rigidity matrix are constructed, and the gearbox in an ideal state is constructed by adopting the assumption in the construction of the gearbox simulation model in the embodiment:

1) assuming that all the components are rigid bodies, neglecting the integral elastic deformation of the gear and the tie bar;

2) the assumption is that no gear disengagement occurs between the gears which are meshed with each other;

3) the meshing reverse impact phenomenon is neglected.

The parameter matrices required to construct the model are as follows:

mass inertia matrix

，

Damping matrix

，

Rigidity matrix

，

Wherein the quality parameter is

And

（pand

representing driving and driven wheels, respectively) inertia parameters of

And

the stiffness parameter is

And

、

damping ginsengNumber is

And

、

wherein

Respectively the gear pair meshing comprehensive rigidity and comprehensive damping,

、

the rigidity of the translational vibration of the driving gear and the driven gear,

、

the damping is the translational vibration of the driving gear and the driven gear. And geometric dimensions

、

。

Then, according to the physical structure of the gear, as shown in fig. 3, a differential kinetic equation of the gearbox is constructed, and the expression is as follows:

wherein the content of the first and second substances,

、

respectively an elastic engagement force and a viscous engagement force,

、

the dynamic meshing force of the gear teeth on the driving gear and the driven gear respectively can be eliminated in the process of solving the equation.

、

Is a rotation variable used to characterize the rotation signal,

、

is a translational variable used to characterize the vibration signal,

、

for characterizing the torque signal. By means of the above-described kinetic model, the simulation matrix for outputting the simulation signal, i.e. the translational array rotation, can be solved

。

Then, the real torque signal obtained by monitoring in the physical sensor is used

、

The parameters of the digital twin model can be updated in real time by introducing the digital twin model

、

And simulating and correcting material parameters and geometric dimension parameters by using finite element simulation software. Obtaining an updated load matrix

After the parameter variables of the simulation model are determined, simulation solving can be carried out in Matlab, and the form of the digital twin model established based on the load relation is

Thereby keeping the digital twin model and the physical model highly consistent. Through the steps, the simulation working condition of the digital twin model is closer to the real working condition of the operation of the gearbox, so that the types of the simulation signals output by the digital twin model are more, the quality of the simulation signals is better, and the fault diagnosis accuracy of the subsequent gearbox is improved.

As shown in fig. 4 and 5, in one embodiment, a gearbox fault diagnosis method specifically includes the following steps:

step S201: acquiring a real signal of the gearbox and a simulated signal generated based on the method of the steps S101-S103, wherein the real signal comprises a real torque signal, a real vibration signal and a real rotation signal, and the simulated signal comprises a simulated vibration signal and a simulated rotation signal;

step S202: and respectively inputting the analog signal and the real signal into a preset fault diagnosis network to obtain a fault diagnosis result of the gearbox.

The preset fault diagnosis network comprises a first neural network used for feature extraction and a second neural network with a hierarchical attention mechanism and used for outputting fault diagnosis results, and is generated based on analog signals and real signals in a training mode.

Specifically, in this embodiment, the analog signals generated in steps S101 to S103 and the real signals are combined to be used as input data of a preset fault diagnosis network, so that the signal type and the signal quality are improved, and the classification accuracy of the fault diagnosis network on the fault signals is improved. In addition, in the present embodiment, the preset failure diagnosis network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting a failure diagnosis result. The second neural network is based on a hierarchical attention mechanism, and multi-level attention can be formed by dividing the features extracted by the first neural network into a plurality of levels in a divide-sum logic form. The attention is focused on the relevant part of the scene, the irrelevant part is ignored, the attention mechanism is essentially effective allocation of information processing resources, different weights are given to the input features of the model, the attention is focused on the information which is considered to be more critical at present, the influence of the critical features on the model is highlighted, the attention mechanism can solve the problem of information overload, and the efficiency and the prediction performance of the model are improved. Therefore, the fault diagnosis result with higher accuracy is further obtained based on the deeper and more critical characteristic information.

Specifically, in an embodiment, taking the analog signal and/or the real signal as a target signal, and performing feature extraction on the target signal based on the first neural network specifically includes the following steps:

step five: and splicing the target signals into a two-dimensional plane vector by taking the number of sampling time points of the target signals and the number of sensors as two-dimensional plane axes.

Step six: and performing one-dimensional convolution calculation on the two-dimensional plane vector based on a plurality of one-dimensional convolution cores with the same dimension to obtain a plurality of one-dimensional original features.

Step seven: and pooling each one-dimensional original feature to obtain a plurality of one-dimensional time sequence features.

Step eight: and performing two-dimensional splicing on the plurality of one-dimensional time sequence characteristics to obtain a new two-dimensional plane vector, and adjusting the number and the dimensionality of one-dimensional convolution kernels.

Step nine: and taking the new two-dimensional plane vector as a two-dimensional plane vector, and returning to the step six until the preset times are returned based on the adjusted one-dimensional convolution kernel to obtain a plurality of target time sequence characteristics.

Step ten: the target time sequence characteristics are spliced into a target one-dimensional vector, the target one-dimensional vector is input into a preset full connection layer, the target one-dimensional vector is converted into a plurality of one-dimensional abstract characteristics with the same number as the number of sampling time points through the preset full connection layer, and the number of elements of the one-dimensional abstract characteristics is the same as the number of sensors.

Specifically, in the present embodiment, by using a one-dimensional convolution method different from the related art, feature data having the same dimension as the input signal data and having more various features can be acquired. As shown in fig. 6, the convolutional neural network is a supervised learning neural network consisting of one or more convolutional layers, activation functions, and pooling layers. The network optimizes the objective function through a gradient descent method, and continuously iterates and trains to update network parameters. Convolutional layers and pooling layers are the core components of the convolutional neural network feature extractor. The input of each node in the convolutional layer is only the neurons in the local range in the neural network of the previous layer, and the convolutional layer carries out deeper analysis on the neurons so as to obtain the characteristics with higher abstraction degree. The pooling layer often takes the maximum of these windows as the result of the output and then continuously slides the window to process each depth slice of the input data volume separately, thereby reducing the network parameters.

Specifically, the number of sampling time points of the target signal and the number of sensors are taken as two-dimensional plane axes, and the target signal is spliced into a two-dimensional plane vector. The signal data initially input into the first neural network is two-dimensional plane data constructed by taking sampling time as an axis y and taking the serial number of the signal sensor as an axis x. Then, one-dimensional convolution calculation is performed on the two-dimensional plane data based on a plurality of one-dimensional convolution kernels with the same size, so that the signal features extracted for the first time are obtained. Compared with the prior art that the input one-dimensional signals are directly convolved by using the one-dimensional convolution kernel, the embodiment combines a plurality of groups of signals into two-dimensional data and then performs one-dimensional convolution, so that different signal sequences are mixed, the obtained characteristic information is higher in comprehensiveness and more representative. In addition, compared with the prior art in which a one-dimensional convolution kernel directly performs convolution calculation on two-dimensional plane data row by row or column by column, the specific operation steps of the embodiment different from the method are as follows:

1. extracting multiple groups of elements with preset sampling time points from the two-dimensional plane vector in a sliding manner aiming at each sensor element, and forming the multiple groups of elements with the preset sampling time points into multiple one-dimensional vectors to be convolved;

2. carrying out convolution operation on each one-dimensional vector to be convolved and the current one-dimensional convolution kernel in sequence to generate current one-dimensional original characteristics;

3. and traversing all the one-dimensional convolution kernels based on the step 2 until one-dimensional original features corresponding to the one-dimensional convolution kernels are obtained.

Specifically, for example, the x-axis of the two-dimensional plane data represents a sensor serial number, the y-axis represents the number of sampling time points, a preset number of consecutive sampling time points (for example, two) are determined on the y-axis, and then data of all sensors corresponding to the sampling time points are obtained in a sliding manner (for example, two rows of data, namely two sampling time points, on the two-dimensional plane data are obtained at a time from the first row until the data of all rows are traversed). And then splicing the acquired data of the preset sampling time points into one-dimensional sequence data in an end-to-end manner by taking the serial number of the sensor as a reference (for example, after two lines of data are acquired each time, two lines of data are spliced into one line of one-dimensional sequence data in an end-to-end manner). And performing convolution calculation on the one-dimensional sequence data obtained by each splicing until all the one-dimensional sequence data are subjected to convolution calculation to obtain a plurality of calculation results, and splicing the calculation results into a one-dimensional feature vector to obtain the one-dimensional original feature. And finally, traversing all the one-dimensional convolution kernels based on the step 2 until the one-dimensional original features corresponding to the one-dimensional convolution kernels are obtained.

I.e. for a given two-dimensional sensor data

Where n is the length of the sensor time series, the data corresponding to the ith time step is

And m represents the number of sensors. At the same time, let the vector

Represents the ith input sample as follows:

wherein

，

Representing the combination of different time windows. Vector quantity

With the jth convolution kernel

And performing operation, wherein the features extracted at the ith time step are represented as:

wherein

And

respectively bias terms and activation functions. When the convolution kernel moves from top to bottom in the input vector, the one-dimensional original feature extracted from the input vector can be obtained

，

Since one convolution kernel can only extract one feature map from the input vector, a plurality of convolution kernels are used to operate on the input vector, thereby extracting more complete fault features.

And then in step seven, the pooling layer is used for carrying out down-sampling operation on the plurality of one-dimensional original features output by the convolution layer to obtain a plurality of one-dimensional time sequence features. On one hand, the pooling layer can extract the most important part of each feature map, and on the other hand, the operation can remarkably reduce the feature dimension and is very suitable for processing high-dimensional data. The maximum pooling operation is expressed as:

wherein the timing characteristics

Is the result of a pooling operation, pool () is the maximum pooling function,pis the size of the pool, and the size of the pool,sis the step size.

Then, in order to further improve that the extracted features are more complete and representative, the embodiment further performs two-dimensional splicing on the obtained multiple one-dimensional time sequence features again to obtain a new two-dimensional plane vector, and adjusts the number and the dimensions of the original one-dimensional convolution kernels. And then, taking the new two-dimensional plane vector as a two-dimensional plane vector, and repeatedly performing the operations from the sixth step to the seventh step by using the adjusted one-dimensional convolution kernel until the preset times. In the present embodiment, the operation of feature extraction is performed 3 times, so that a plurality of target time series features with more excellent feature performance are obtained.

And finally, splicing the target time sequence characteristics into a target one-dimensional vector end to end, inputting the target one-dimensional vector into a preset full connection layer, adjusting the dimensionality and the quantity of the target one-dimensional vector through the preset full connection layer, converting the target one-dimensional vector into a plurality of one-dimensional abstract characteristics with the same quantity as the number of sampling time points of a signal, wherein the number of elements of each one-dimensional abstract characteristic is the same as the quantity of sensors. Therefore, the dimensionality of the signal input data is completely consistent with that of the feature data, and the problem that the accuracy is reduced due to the reduction of the data quantity in the feature extraction process is solved.

Specifically, in an embodiment, outputting the fault diagnosis result based on the second neural network specifically includes the following steps:

step eleven: and dimension expansion is carried out on the plurality of one-dimensional abstract features respectively.

Step twelve: and performing level attention processing on the plurality of expanded one-dimensional abstract features based on a linear perception machine, and fusing the plurality of one-dimensional abstract features to obtain an output vector.

Step thirteen: and inputting the output vector into a preset classifier so that the preset classifier outputs a fault diagnosis result.

In particular, in the present embodiment, a hierarchical attention mechanism is used to fuse the high-level feature information (i.e., one-dimensional abstract features) output by the first neural network. Before this, dimension expansion is first performed on one-dimensional abstract features. The attention network can adopt three modes to carry out dimension expansion on an input sequence so as to further improve feature accuracy, the first mode repeats the first moment data on the time dimension for preset times, and in a complex physical system, the value of the initial moment can greatly influence the development direction and the process of the physical system, so that the importance of the first moment is emphasized, and the accuracy of the one-dimensional abstract feature after dimension expansion can be higher. The other two processing modes are respectively that data at each moment and the first moment are subtracted or multiplied, degradation information is observed from multiple dimensions, and the attention network is helped to learn degradation characteristics more comprehensively and abundantly.

Then, the hidden state of the fault feature at each moment is calculated based on a linear perceptron to increase the nonlinearity of the network and better fit the degradation feature of the gearbox. Then, normalization processing is carried out on the hidden state, and the input sequence of the normalization processing is multiplied by the weight to realize feature fusion. Finally, inputting the feature vectors output by the hierarchical attention network into preset classifiers (including but not limited to softmax classifiers and logistic classifiers) to realize gearbox fault diagnosis.

Specifically, in an embodiment, the step twelve specifically includes the following steps:

fourteen steps: and sequentially inputting the plurality of one-dimensional abstract features into a linear perceptron to obtain a plurality of hidden features.

Step fifteen: and calculating a plurality of weight coefficients corresponding to the plurality of hidden features through a softmax function, wherein the weight coefficients are used for representing the hierarchical attention of the hidden features.

Sixthly, the steps are as follows: and carrying out weighted summation on the hidden features by utilizing the weight coefficients to obtain an output vector.

In particular, assume that the second neural network input is characterized as

Wherein

One-dimensional abstract features representing the ith time step will first be

Learning hidden feature expressions for input to a linear perceptron

The expression is as follows:

wherein

And

respectively representing the weight and bias of the linear perceptron.

Then, calculating a weight coefficient of each time step, wherein the larger the coefficient is, the more fault information is contained in the time step, and the weight coefficient is calculated through a softmax function and is expressed as:

wherein

Is a randomly initialized vector which is continuously updated along with the iteration number. Finally, weighting each time step fault information by using a weighting coefficient to obtain an output vector of a second neural network

：

Through the steps, the output vector containing more complete and accurate characteristics is accurately calculated and used for classifying the fault signals.

In this embodiment, the parameters of each layer of the time-series deep diagnostic network model composed of the first neural network and the second neural network are shown in table 1.

TABLE 1 time series deep diagnostic network model architecture description

Specifically, in one embodiment, an example of an application of experimental testing and simulation of a gearbox is as follows:

to monitor the real-time operating condition of the gearbox, the gearbox is first placed in a test system. The system consists of a gear box fault diagnosis platform, a vibration signal sensor, a rotation signal sensor, a torque signal sensor, a signal cable, a data acquisition instrument and a computer. The test system is convenient to operate, parts are convenient and fast to replace, and working conditions such as normal working conditions, tooth breaking faults, pitting faults and abrasion faults of the gear box can be simulated. The module of the big gear and the small gear in the gear box is 75, the number of the big gears is 55, and oil immersion type lubrication is adopted.

Firstly, measuring basic geometric dimension parameters and material parameters of the gear, and establishing initial input parameters of a dynamic model, including quality parameters

And with

Inertia parameter

And

stiffness parameter (c)

And

、

damping parameter

And with

、

And geometric dimensions

、

. Then a parameter matrix required by the model and a differential kinetic model of the gearbox can be constructed, and then the translation array rotation can be solved through the kinetic model

Variable of medium rotation

、

Variable of translation (vibration)

、

Then the rotation signal in the simulation model can be realized

、

Vibration signal

、

And torque signal

、

And (4) monitoring and collecting. Then the torque signal obtained by monitoring in the physical sensor

、

The digital twin model is introduced, namely the parameters of the digital twin model can be updated in real time

、

Then obtain the updated load matrix

After the parameter variables of the simulation model are determined, simulation solving can be carried out in Matlab, and the model is in the form of

Thereby keeping the digital twin model and the physical model highly consistent.

Finally, the digital twin model can output monitoring rotation signals

、

And vibration signal

、

Meanwhile, the output of the model can be compared with the rotation signal and the vibration signal measured by the physical sensor in real time to be used as index variables for reflecting the working condition of the gear, so that the real-time monitoring of the health state of the gear box is realized.

In the experiment, under the condition that the rotating speed is 880rpm and no load is applied, signals of sensors of normal gears, pitting gears, broken gears and worn gears during working are respectively collected by using a 5120Hz sampling rate, and meanwhile, the signals of the sensors are also output in a simulation modelSensor signals when the gear is working are always detected. Specific information is shown in table 2, and there are 3300 samples in total for all samples, wherein the normal state, the worn state are 1000 samples each, the pitting state is 800 samples, and the tooth breakage is 500 samples. Each sample comprises a physical sensor record when the gearbox continuously works and a rotation signal output by a digital twin model

、

And vibration signal

、

. They form a data set of a time series deep diagnostic model for training network parameters and testing network performance.

TABLE 2 Experimental sample quantity information

Because the data monitored by different sensors represent different physical characteristics, the dimensions and the magnitude of each parameter are different, and therefore, the fluctuation condition of each parameter cannot be directly reflected by the original data. In order to avoid the influence of the dimension and the value range of the predicted process parameters, the data needs to be uniformly normalized, and the values of all the attributes are reduced to a same value space. The normalization can reduce the difference between numerical values, avoid the problem of data bias, facilitate subsequent training and simultaneously contribute to improving the accuracy and the convergence of a prediction model. In view of this, a min-max normalization is used. Given a given input

Where N represents the time step and K represents the number of sensors. The ith vector isThe maximum-minimum normalization method is as follows:

wherein

Respectively represent vectorsx _iThe maximum and minimum values in (1).

Because the fault types of the gear box are four types, and simultaneously, in order to facilitate the direct input of the classification result by the network, the independent hot coding is adopted for each fault, and the result is shown in table 3.

TABLE 3 gearbox failure type one-hot code

During the training process, 80% of the samples are randomly selected as the training set for each complete training subset, and the remaining 20% are the validation set. Selecting and adjusting hyper-parameters of the time-series deep diagnosis network through the test effect of the model on the verification set, wherein the hyper-parameters are determined by considering the prediction precision and the calculation cost, and the final training flow chart of the model is shown in FIG. 7.

For each round of training, samples were randomly divided into small batches, each containing 32 samples, and placed into the training system. Firstly, extracting features from original data by using 3 one-dimensional convolution units, extracting local features of the data by using a plurality of filters in each convolution layer, wherein the size of each convolution kernel in each convolution layer is 2 x 1, moving one step at a time, and setting the filling value of each time sequence to be 1 in order to ensure that edge information is not lost. And then applying a nonlinear activation function to the convolutional layer output to enhance the network expression capability. And finally, removing redundant information from the output through a pooling layer, setting the window size in the pooling layer to be 2, and moving two step lengths each time to complete the feature mapping in one convolution unit. And after the signal fault features are extracted by the one-dimensional convolution network, inputting the fault features into a full connection layer to obtain higher-level fault features, wherein output data are m multiplied by n dimensions and have the same dimension as original input data. Next, a hierarchical attention machine is used to fuse the high-level feature information. The hidden state of the fault feature at each moment is calculated and normalized by using a softmax function, and then the input sequence is multiplied by the weight to realize feature fusion. And finally, inputting the characteristic vector output by the hierarchical attention network into a softmax classifier to realize the fault diagnosis of the gearbox. The time sequence depth diagnosis network uses a cross entropy loss function as an objective function, reverse propagation learning is used for updating weights in the network, and an Adam optimizer with self-adaptive adjustment capability is adopted for optimization. The above process completes one training of the model, and the learning rate is set to 0.0001 during the model training so that the model is stably converged. By default, the maximum number of training periods for the model is 700.

A deep time sequence diagnosis network is constructed by utilizing a pytorech deep learning framework, the iteration number of the network is set to be 700, 16 samples are input each time, and the learning rate is 0.0001. The network parameters were optimized using the gradient descent method, and the training curve of the network loss function is shown in fig. 7.

As can be seen from fig. 8, after 700 iterations, the network error has dropped sufficiently small to prove that the network can effectively fit the training samples. And applying the time sequence deep diagnosis to 660 test samples to evaluate the training effect of the network. The time sequence deep diagnosis network is used for diagnosing four faults of the gearbox as shown in the table. As can be seen from Table 4, the network has the highest recognition accuracy of 98% in the normal state of the gearbox, wherein 4 samples are misdiagnosed, and two samples are diagnosed as pitting and broken tooth faults. The recognition accuracy of the network on the pitting failure of the gearbox is the lowest, but still higher than 90%, wherein 10 samples are misdiagnosed as other states, 2 samples are diagnosed as a broken tooth failure, and 8 samples are diagnosed as a wear failure, which may be because the characteristics of the pitting failure and the wear failure are similar to each other, so that the network cannot effectively judge which failure belongs to. The network has good effects of respectively identifying 96% and 94% of gear breakage faults and wear faults. For 660 total test samples, 30 total test samples are misdiagnosed, and the overall recognition rate of the model is 95.5%, which shows that the time sequence deep diagnosis network can effectively recognize four faults of the gearbox.

TABLE 4 time series deep diagnosis network failure prediction results

Through the steps, according to the technical scheme provided by the application, a digital model of the digital twin model is established based on material parameters and geometric dimension parameters of a gear in the gearbox, and then the parameters of the digital twin model are corrected through simulation in combination with a real torque signal of the gearbox. Compared with a digital twin model in the prior art, the method better conforms to the actual working condition of the gearbox, so that the data quality of the analog signal better conforms to the real situation. Therefore, the real torque signal is used as the input of the model, the simulated vibration signal and the simulated rotation signal are output, the fault data type of the gear box is expanded, the simulated signal with higher data quality is obtained, fault diagnosis is carried out based on the real signal data and the simulated signal data of the gear box, and the fault diagnosis accuracy of subsequent fault diagnosis work of the gear box is improved.

As shown in fig. 9, the present embodiment also provides a gearbox fault diagnosis device, including:

the signal acquisition module 201 is configured to acquire a real signal of the gearbox and an analog signal generated by an embodiment of a method for acquiring a signal of the gearbox, where the real signal includes a real torque signal, a real vibration signal, and a real rotation signal. For details, refer to the related description of step S201 in the above method embodiment, and are not repeated herein.

And the fault diagnosis module 202 is used for inputting the analog signal and the real signal into a preset fault diagnosis network respectively to obtain a fault diagnosis result of the gearbox. For details, refer to the related description of step S201 in the above method embodiment, and no further description is provided here.

The gearbox fault diagnosis device provided by the embodiment of the invention is used for executing the gearbox fault diagnosis method provided by the embodiment, the implementation mode and the principle are the same, and the detailed content refers to the relevant description of the method embodiment and is not repeated.

Through the cooperation of all the components, the technical scheme provided by the application establishes a digital model of the digital twin model based on the material parameters and the geometric dimension parameters of the gear in the gear box, and then corrects the parameters of the digital twin model through simulation in combination with the real torque signal of the gear box. Compared with a digital twin model in the prior art, the method better conforms to the actual working condition of the gearbox, so that the data quality of the analog signal better conforms to the real situation. Therefore, the real torque signal is used as the input of the model, the simulated vibration signal and the simulated rotation signal are output, the fault data type of the gear box is expanded, the simulated signal with higher data quality is obtained, fault diagnosis is carried out based on the real signal data and the simulated signal data of the gear box, and the fault diagnosis accuracy of subsequent fault diagnosis work of the gear box is improved.

Fig. 10 shows an electronic device according to an embodiment of the present invention, where the device includes a processor 901 and a memory 902, which may be connected by a bus or by other means, and fig. 10 illustrates an example of a connection by a bus.

Processor 901 may be a Central Processing Unit (CPU). The Processor 901 may also be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components, or combinations thereof.

The memory 902, which is a non-transitory computer-readable storage medium, may be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as program instructions/modules corresponding to the methods in the above-described method embodiments. The processor 901 executes various functional applications and data processing of the processor by executing non-transitory software programs, instructions and modules stored in the memory 902, that is, implements the methods in the above-described method embodiments.

The memory 902 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required for at least one function; the storage data area may store data created by the processor 901, and the like. Further, the memory 902 may include high speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, the memory 902 may optionally include memory located remotely from the processor 901, which may be connected to the processor 901 via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

One or more modules are stored in the memory 902, which when executed by the processor 901 performs the methods in the above-described method embodiments.

The specific details of the electronic device may be understood by referring to the corresponding related descriptions and effects in the above method embodiments, and are not described herein again.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by a computer program, and the implemented program can be stored in a computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. The storage medium may be a magnetic Disk, an optical Disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a Flash Memory (Flash Memory), a Hard Disk (Hard Disk Drive, abbreviated as HDD) or a Solid State Drive (SSD), etc.; the storage medium may also comprise a combination of memories of the kind described above.

Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A gearbox signal acquisition method, the method comprising:

establishing a digital twin model based on material parameters and geometric dimension parameters of a gear in the gear box;

acquiring a real torque signal of the gearbox, wherein the real torque signal comprises first signal data used for correcting the parameters of the digital twin model and second signal data used for calculating the analog signal;

inputting the first signal data into the digital twin model to correct parameters of the digital twin model;

inputting the second signal data into a corrected digital twin model to generate an analog signal, the analog signal comprising an analog vibration signal and an analog rotation signal;

wherein the establishing of the digital twin model based on the material parameters and the geometric dimension parameters of the gear in the gear box comprises the following steps:

respectively acquiring material parameters and geometric dimension parameters of a driving wheel and a driven wheel in the gearbox, wherein the material parameters comprise at least one of mass parameters, inertia parameters, rigidity parameters and damping parameters;

creating a differential kinetic equation of the gearbox by utilizing the material parameters and the geometric dimension parameters based on a physical structure meshed between the driving wheel and the driven wheel;

calculating a simulation matrix for outputting the simulation signal through the gearbox differential kinetic equation;

establishing a mathematical model for representing a gear load relation by using the simulation matrix, the mass inertia matrix, the damping matrix and the rigidity matrix, and taking the mathematical model as the digital twin model;

wherein the mass-inertia matrix is constructed based on the mass parameter and the inertia parameter, the damping matrix is constructed based on the damping parameter and the geometric parameter, and the stiffness matrix is constructed based on the stiffness parameter and the geometric parameter.

2. A gearbox fault diagnosis method, characterized in that the method comprises:

acquiring real signals of a gearbox and simulated signals generated by the method of claim 1, the real signals comprising real torque signals, real vibration signals and real rotation signals;

respectively inputting the analog signal and the real signal into a preset fault diagnosis network to obtain a fault diagnosis result of the gearbox;

the preset fault diagnosis network comprises a first neural network used for feature extraction and a second neural network with a hierarchical attention mechanism and used for outputting fault diagnosis results, and the preset fault diagnosis network is generated based on the simulation signal and the real signal training.

3. The method according to claim 2, wherein the step of performing feature extraction on the target signal based on the first neural network by using the simulated signal and/or the real signal as the target signal comprises:

splicing the target signals into a two-dimensional plane vector by taking the number of sampling time points of the target signals and the number of sensors as two-dimensional plane axes;

performing one-dimensional convolution calculation on the two-dimensional plane vector based on a plurality of one-dimensional convolution cores with the same dimension to obtain a plurality of one-dimensional original features;

pooling each one-dimensional original feature to obtain a plurality of one-dimensional time sequence features;

performing two-dimensional splicing on the plurality of one-dimensional time sequence characteristics to obtain a new two-dimensional plane vector, and adjusting the number and the dimension of the one-dimensional convolution kernels;

taking the new two-dimensional plane vector as the two-dimensional plane vector, returning the two-dimensional plane vector to the one-dimensional convolution kernel based on the multiple same dimensions for one-dimensional convolution calculation based on the adjusted one-dimensional convolution kernel to obtain multiple one-dimensional original features, and obtaining multiple target time sequence features until the preset times are returned;

splicing the target time sequence characteristics into a target one-dimensional vector, inputting the target one-dimensional vector into a preset full connection layer, and converting the target one-dimensional vector into a plurality of one-dimensional abstract characteristics with the same number as the sampling time points through the preset full connection layer, wherein the number of elements of the one-dimensional abstract characteristics is the same as the number of the sensors.

4. The method of claim 3, wherein performing a one-dimensional convolution calculation on the two-dimensional plane vector based on a plurality of one-dimensional convolution kernels having the same dimension to obtain a plurality of one-dimensional original features comprises:

extracting multiple groups of elements with preset sampling time points from the two-dimensional plane vector in a sliding manner aiming at each sensor element, and forming the multiple groups of elements with the preset sampling time points into multiple one-dimensional vectors to be convolved;

carrying out convolution operation on each one-dimensional vector to be convolved and the current one-dimensional convolution kernel in sequence to generate current one-dimensional original characteristics;

and traversing all the one-dimensional convolution kernels until the one-dimensional original features corresponding to the one-dimensional convolution kernels are obtained based on the step of sequentially carrying out convolution operation on the one-dimensional vectors to be convolved and the current one-dimensional convolution kernel to generate the current one-dimensional original features.

5. The method of claim 3, wherein the step of outputting the fault diagnosis result based on the second neural network comprises:

respectively carrying out dimension expansion on the plurality of one-dimensional abstract features;

performing level attention processing on the plurality of expanded one-dimensional abstract features based on a linear perceptron, and fusing the plurality of one-dimensional abstract features to obtain an output vector;

and inputting the output vector into a preset classifier so that the preset classifier outputs the fault diagnosis result.

6. The method of claim 5, wherein performing hierarchical attention processing on the plurality of one-dimensional abstract features based on a linear perceptron and fusing the plurality of one-dimensional abstract features to obtain an output vector comprises:

inputting the plurality of one-dimensional abstract features into the linear perceptron in sequence to obtain a plurality of hidden features;

calculating a plurality of weight coefficients corresponding to the plurality of hidden features through a softmax function, wherein the weight coefficients are used for representing the hierarchical attention of the hidden features;

and carrying out weighted summation on the hidden features by utilizing the weighting coefficients to obtain the output vector.

7. A gearbox fault diagnosis device, characterized in that the device comprises:

a signal acquisition module for acquiring real signals of a gearbox and simulated signals generated by the method of claim 1, the real signals comprising real torque signals, real vibration signals and real rotation signals;

the fault diagnosis module is used for respectively inputting the analog signal and the real signal into a preset fault diagnosis network to obtain a fault diagnosis result of the gearbox;

8. An electronic device, comprising:

a memory and a processor communicatively coupled to each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the method of any of claims 1-6.

9. A computer-readable storage medium having stored thereon computer instructions for causing a computer to thereby perform the method of any one of claims 1-6.