WO2023020388A1 - 一种齿轮箱故障诊断、信号采集方法、装置和电子设备 - Google Patents

一种齿轮箱故障诊断、信号采集方法、装置和电子设备 Download PDF

Info

Publication number
WO2023020388A1
WO2023020388A1 PCT/CN2022/112143 CN2022112143W WO2023020388A1 WO 2023020388 A1 WO2023020388 A1 WO 2023020388A1 CN 2022112143 W CN2022112143 W CN 2022112143W WO 2023020388 A1 WO2023020388 A1 WO 2023020388A1
Authority
WO
WIPO (PCT)
Prior art keywords
dimensional
signal
gearbox
fault diagnosis
real
Prior art date
Application number
PCT/CN2022/112143
Other languages
English (en)
French (fr)
Inventor
苏营
秦玉文
邹祖冰
甘富航
邓友汉
王罗
高远
Original Assignee
中国长江三峡集团有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国长江三峡集团有限公司 filed Critical 中国长江三峡集团有限公司
Priority to DE112022000106.2T priority Critical patent/DE112022000106T5/de
Publication of WO2023020388A1 publication Critical patent/WO2023020388A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/021—Gearings
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/028—Acoustic or vibration analysis
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Definitions

  • the invention relates to the field of algorithm design, in particular to a gearbox fault diagnosis, signal collection method, device and electronic equipment.
  • Wind turbines work in harsh environments such as sandstorms, lightning, and rainstorms all year round.
  • the gearbox is the key transmission device of the wind turbine. Maintenance, replacement and repair are time-consuming and costly.
  • its health status is closely related to the stable operation of the wind turbine.
  • researchers have widely used signal analysis methods to realize gearbox fault diagnosis. The methods use time, frequency domain statistical features, wavelet transform, fast Fourier transform, empirical mode decomposition, etc. to extract features from gearbox fault signals. Then the fault diagnosis is realized, and the research results show that this kind of method can achieve higher accuracy in the steady state.
  • the working conditions of the gearbox are complex and changeable, the signal components are diverse, and multiple modes are mixed, resulting in non-stationary signals.
  • Commonly used signal analysis methods will have the problems of false component interference and low resolution of characteristic signals, and cannot accurately judge the fault status of the gearbox.
  • the final signal collected by the sensor is mixed with a lot of environmental noise, resulting in a low signal-to-noise ratio.
  • the existing sensors in the wind field can only collect characteristic signals such as wind speed, voltage, and current, but cannot collect or accurately measure physical quantities such as component stress and acceleration, so they cannot accurately reflect the fault status of the gearbox. Therefore, improving the quality of sensor data is crucial to achieving accurate gearbox fault diagnosis.
  • the embodiments of the present invention provide a gearbox fault diagnosis, a signal acquisition method, a device and electronic equipment, thereby improving the quality of sensor data and improving the accuracy of gearbox fault diagnosis.
  • the present invention provides a gearbox signal acquisition method, the method comprising: establishing a digital twin model based on the material parameters and geometric dimension parameters of the gear inside the gearbox; acquiring the real torque signal of the gearbox, the real The torque signal includes first signal data for correcting parameters of the digital twin model and second signal data for calculating an analog signal; the first signal data is input into the digital twin model to modify the parameters of the digital twin model performing correction; 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.
  • the establishment of a digital twin model based on the material parameters and geometric dimension parameters of the gears in the gearbox includes: respectively obtaining the material parameters and geometric dimension parameters of the driving wheel and the driven wheel in the gearbox, and the material parameters include At least one of mass parameters, inertia parameters, stiffness parameters and damping parameters; based on the physical structure of the mesh between the driving wheel and the driven wheel, using the material parameters and geometric size parameters to create a gearbox differential dynamics equation; by The differential dynamic equation of the gearbox is used to calculate the simulation matrix used to output the simulation signal; the mathematical model used to characterize the gear load relationship is created by using the simulation matrix, the mass inertia matrix, the damping matrix and the stiffness matrix, and the A mathematical model is used 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 dimension parameter, and the stiffness matrix is constructed based on The stiffness parameters and the geometric
  • the present invention provides a gearbox fault diagnosis method, the method comprising: acquiring the real signal of the gearbox and the analog signal generated according to any optional implementation manner of the first aspect, the real signal includes The real torque signal, the real vibration signal and the real rotation signal; respectively input the analog signal and the real signal into the preset fault diagnosis network to obtain the fault diagnosis result of the gearbox; wherein, the preset fault diagnosis The network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results, and the preset fault diagnosis network is trained based on the simulated signal and the real signal generate.
  • the step of performing feature extraction on the target signal based on the first neural network includes: using the sampling time points of the target signal and The number of sensors is a two-dimensional plane axis, and the target signal is spliced into a two-dimensional plane vector; based on multiple one-dimensional convolution kernels with the same dimension, one-dimensional convolution calculation is performed on the two-dimensional plane vector to obtain multiple one-dimensional original feature; each of the one-dimensional original features is pooled to obtain a plurality of one-dimensional time series features; two-dimensional splicing is performed on a plurality of the one-dimensional time series features to obtain a new two-dimensional plane vector, and the one-dimensional convolution The number and dimension of the kernel are adjusted; the new two-dimensional plane vector is used as the two-dimensional plane vector, and based on the adjusted one-dimensional convolution kernel, the one-dimensional convolution based on the same multiple dimensions is returned The step of performing one-dimensional convolution calculation on the
  • performing one-dimensional convolution calculation on the two-dimensional plane vector based on multiple one-dimensional convolution kernels with the same dimension to obtain multiple one-dimensional original features including: for each sensor element, from the two-dimensional plane Slidingly extract elements of multiple sets of preset sampling time points from the vector, and form multiple one-dimensional vectors to be convoluted with the elements of the multiple sets of preset sampling time points;
  • the convolution kernel performs a convolution operation to generate the current one-dimensional original feature; based on the steps of performing convolution operations on each one-dimensional vector to be convolved with the current one-dimensional convolution kernel in turn to generate the current one-dimensional original feature, traverse all One-dimensional convolution kernels until the one-dimensional original features corresponding to each one-dimensional convolution kernel are obtained.
  • the step of outputting the fault diagnosis result based on the second neural network includes: performing hierarchical attention processing on the multiple one-dimensional abstract features based on the linear perceptron, and fusing the multiple one-dimensional abstract features to obtain an output vector; inputting the output vector into a preset classifier, so that the preset classifier outputs the fault diagnosis result.
  • performing hierarchical attention processing on the multiple one-dimensional abstract features based on a linear perceptron, and fusing the multiple one-dimensional abstract features to obtain an output vector including: sequentially inputting the multiple one-dimensional abstract features The linear perceptron obtains multiple hidden features; calculates multiple weight coefficients corresponding to the multiple hidden features through a softmax function, and the weight coefficients are used to characterize the hierarchical attention of the hidden features; using the multiple The weight coefficient performs weighted summation on the plurality of hidden features to obtain the output vector.
  • the present invention provides a gearbox fault diagnosis device, the device includes: a signal acquisition module, used to acquire the real signal of the gearbox and the analog signal generated according to any optional implementation manner of the first aspect , the real signal includes a real torque signal, a real vibration signal and a real rotation signal; a fault diagnosis module is configured to input the analog signal and the real signal into a preset fault diagnosis network to obtain a fault of the gearbox Diagnosis results; wherein, the preset fault diagnosis network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results, and the preset fault diagnosis A network is trained based on the simulated signal and the real signal.
  • an embodiment of the present invention provides an electronic device, including: a memory and a processor, the memory and the processor are connected to each other in communication, the memory stores computer instructions, and the processor By executing the computer instructions, the first aspect or the method described in any optional implementation manner of the first aspect is executed.
  • an embodiment of the present invention provides a computer-readable storage medium, the computer-readable storage medium stores computer instructions, and the computer instructions are used to cause the computer to execute the first aspect, or the first The method described in any one of the alternative embodiments of the aspect.
  • the digital model of the digital twin model is established based on the material parameters and geometric size parameters of the internal gear of the gearbox, and then combined with the real torque signal of the gearbox, the parameters of the digital twin model are corrected through simulation.
  • the digital twin model of traditional technology it is more in line with the actual working conditions of the gearbox, so that the data quality of the analog signal is more in line with the real situation.
  • the real torque signal is used as the input of the model, and the simulated vibration signal and simulated rotation signal are output, which not only expands the types of gearbox fault data, but also obtains simulated signals with higher data quality, so that based on the real signal data and simulated signals of the gearbox
  • the data is used for fault diagnosis, which improves the fault diagnosis accuracy of the subsequent gearbox fault diagnosis work.
  • the present invention utilizes the first neural network for feature extraction and the second neural network with a hierarchical attention mechanism for outputting fault diagnosis results to realize fault diagnosis during fault diagnosis, so that useful sample features are further strengthened, and useless The characteristics of the sample are weakened, thereby further improving the accuracy of fault diagnosis.
  • the feature extraction step of the first neural network the fault signals collected by multiple sensors are converted from two-dimensional data to one-dimensional data, and then one-dimensional convolution is performed, and then the obtained one-dimensional features are repeatedly spliced into two-dimensional data Perform multi-level one-dimensional, two-dimensional, and one-dimensional feature extraction. Finally, the signal features with more prominent features and the same dimensions as the original data are obtained.
  • the classification of subsequent fault signals is further facilitated, and the accuracy of diagnosis is improved.
  • the feature vector is processed with hierarchical attention, which increases the nonlinearity of the neural network, thereby better fitting the degradation characteristics of the gearbox, making the fault diagnosis result closer to the real situation, and improving the accuracy of gearbox fault diagnosis. .
  • Fig. 1 shows a schematic diagram of the steps of a gearbox signal acquisition method in an embodiment of the present invention
  • Fig. 2 shows a block flow diagram of a gearbox signal acquisition method in an embodiment of the present invention
  • Fig. 3 shows a schematic diagram of the meshing dynamics structure of a typical spur gear pair in the prior art
  • Fig. 4 shows a schematic diagram of the steps of a gearbox fault diagnosis method in an embodiment of the present invention
  • Fig. 5 shows a schematic flow chart of a gearbox fault diagnosis method in an embodiment of the present invention
  • Fig. 6 shows a schematic structural diagram of the first neural network in one embodiment of the present invention
  • Fig. 7 shows a schematic flow chart of another gearbox fault diagnosis method in an embodiment of the present invention.
  • FIG. 8 shows a graph of the loss function of the deep timing diagnosis network changing with the number of iterations in an embodiment of the present invention
  • Fig. 9 shows a schematic structural diagram of a gear box signal acquisition device in an embodiment of the present invention.
  • Fig. 10 shows a schematic structural diagram of an electronic device in an embodiment of the present invention.
  • a kind of gearbox signal collection method specifically includes the following steps:
  • Step S101 Establish a digital twin model based on the material parameters and geometric size parameters of the gear box's internal gears.
  • Step S102 Obtain the real torque signal of the gearbox, the real torque signal includes the first signal data for correcting the parameters of the digital twin model and the second signal data for calculating the analog signal, and input the first signal data into the digital twin model, To correct the parameters of the digital twin model.
  • Step S103 Input the second signal data into the corrected digital twin model to generate an analog signal, the analog signal includes an analog vibration signal and an analog rotation signal.
  • the gearbox is established based on the material parameters and geometric size parameters of the gears in the gearbox.
  • the digital twin model of , including but not limited to gear size, gear material, material stiffness, inter-gear damping, gear mass.
  • the simulated working conditions of the digital twin model are closer to the real working conditions of the gearbox.
  • Traditional digital twin models are usually simulations of a single vibration signal.
  • step S101 specifically includes the following steps:
  • Step 1 Obtain the material parameters and geometric dimension parameters of the driving wheel and the driven wheel in the gearbox respectively.
  • the material parameters include at least one of mass parameters, inertia parameters, stiffness parameters and damping parameters.
  • Step 2 Based on the physical structure of the mesh between the driving wheel and the driven wheel, the differential dynamic equation of the gearbox is created using material parameters and geometric size parameters.
  • Step 3 Calculating the simulation matrix used to output the simulation signal through the differential dynamic equation of the gearbox.
  • Step 4 Use the simulation matrix, mass inertia matrix, damping matrix, and stiffness matrix to create a mathematical model for characterizing the gear load relationship, and use 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
  • the stiffness matrix is constructed based on stiffness parameters and geometric dimension parameters.
  • a mathematical equation model is created based on the gear load relationship, thereby realizing the creation of a digital twin model. Firstly, the material parameters including mass parameters, inertia parameters, stiffness parameters and damping parameters are obtained, and then the geometric dimension parameters representing the gear size are obtained. Then construct the mass inertia matrix, damping matrix and stiffness matrix. In this embodiment, when establishing the simulation model of the gearbox, it is assumed to construct the ideal gearbox:
  • the parameter matrix required to form the model is as follows:
  • the mass parameters are m p and m g (p and g represent the driving wheel and the driven wheel respectively)
  • the inertia parameters are I p and I g
  • the stiffness parameters are km m and k py , k gy
  • the damping parameters are c m and c py , c gy , where k m , c m are the comprehensive stiffness and comprehensive damping of gear pair meshing respectively
  • k py , k gy are the main force
  • the stiffness of the translational vibration of the driven gear and c py , c gy are the main force, the damping of the translational vibration of the driven gear .
  • geometric dimensions R p , R g are geometric dimensions.
  • F k , F c are the elastic meshing force and viscous meshing force respectively
  • F p , F g are the dynamic meshing forces of the primary and driven gears respectively, which will be eliminated in the process of solving the equation.
  • ⁇ p and ⁇ g are rotation variables used to represent rotation signals
  • y p and y g are translation variables used to represent vibration signals
  • T p0 and T g0 are used to represent torque signals.
  • the real torque signals T p0 and T g0 monitored by the physical sensor are transmitted to the digital twin model, and the parameters T p and T g of the digital twin model can be updated in real time, and the material parameters and geometry can be simulated and corrected by using the finite element simulation software. Size parameters. Get the updated loading matrix
  • the parameter variables of the simulation model After the parameter variables of the simulation model are determined, they can be simulated and solved in Matlab.
  • the form of the digital twin model established based on the load relationship is In this way, the digital twin model maintains a high degree of consistency with the physical model.
  • the simulated working conditions of the digital twin model are closer to the real working conditions of the gearbox, and the digital twin model outputs more types of analog signals with better quality, thereby improving the accuracy of subsequent gearbox fault diagnosis. .
  • a gearbox fault diagnosis method specifically includes the following steps:
  • Step S201 Obtain the real signal of the gearbox and the analog signal generated based on the above steps S101-S103.
  • the real signal includes a real torque signal, a real vibration signal and a real rotation signal
  • the analog signal includes a simulated vibration signal and a simulated rotation signal;
  • Step S202 input the analog signal and the real signal into the preset fault diagnosis network respectively, and obtain the fault diagnosis result of the gearbox.
  • the preset fault diagnosis network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results, and the preset fault diagnosis network is based on analog signals and real signals training generated.
  • the analog signals generated in steps S101 to S103 are combined with the real signals as the input data of the preset fault diagnosis network, which improves the signal type and signal quality, thereby improving the fault diagnosis network's ability to detect fault signals. classification accuracy.
  • the preset fault diagnosis network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results.
  • the second neural network is based on a hierarchical attention mechanism, and can form multi-level attention by dividing the features extracted by the first neural network into multiple levels with a sub-total logic form. Focus on the relevant parts of the scene while ignoring the irrelevant parts.
  • the attention mechanism is essentially an effective allocation of information processing resources.
  • the attention is focused on what is currently considered to be more critical.
  • the influence of key features on the model is highlighted, and the attention mechanism can solve the problem of information overload and improve model efficiency and prediction performance. In this way, based on deeper and more critical feature information, more accurate fault diagnosis results can be obtained.
  • using the analog signal and/or the real signal as the target signal, and performing feature extraction on the target signal based on the first neural network specifically includes the following steps:
  • Step 5 Take the sampling time points of the target signal and the number of sensors as the two-dimensional plane axis, and stitch the target signal into a two-dimensional plane vector.
  • Step 6 Perform one-dimensional convolution calculation on the two-dimensional plane vector based on multiple one-dimensional convolution kernels with the same dimension to obtain multiple one-dimensional original features.
  • Step 7 Pool each one-dimensional original feature to obtain multiple one-dimensional time series features.
  • Step 8 Perform two-dimensional splicing of multiple one-dimensional time series features to obtain a new two-dimensional plane vector, and adjust the number and dimension of one-dimensional convolution kernels.
  • Step 9 Use the new two-dimensional plane vector as a two-dimensional plane vector, and return to step six based on the adjusted one-dimensional convolution kernel until the preset number of times is returned to obtain multiple target time series features.
  • Step 10 Concatenate multiple target time-series features into a target one-dimensional vector, and input the target one-dimensional vector into the preset fully connected layer, so as to convert the target one-dimensional vector into the number of sampling time points through the preset fully connected layer
  • a 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 the gradient descent method, and continuously iteratively trains and updates the network parameters.
  • Convolutional and pooling layers are the core components of convolutional neural network feature extractors.
  • the input of each node in the convolutional layer is only the neurons in the local range of the previous layer of neural network, and the convolutional layer analyzes them more deeply to obtain features with a higher degree of abstraction.
  • the pooling layer often takes the maximum value of these windows as the output result, and then continuously slides the window to process each depth slice of the input data body separately, thereby reducing network parameters.
  • the target signal is spliced into a two-dimensional plane vector. That is, the signal data initially input to the first neural network is two-dimensional plane data constructed with the sampling time as the y-axis and the serial number of the signal sensor as the x-axis. Then, one-dimensional convolution calculation is performed on the two-dimensional plane data based on multiple one-dimensional convolution kernels of the same size, so as to obtain the signal features extracted for the first time.
  • this embodiment Compared with the method of using a one-dimensional convolution kernel to directly convolve the input one-dimensional signal in the prior art, this embodiment combines multiple sets of signals into two-dimensional data and then performs one-dimensional convolution, so that different signal sequences are aliased to obtain The feature information is more comprehensive, and the feature information is more representative.
  • the specific operation steps of this embodiment are as follows:
  • step 2 Based on step 2, traverse all one-dimensional convolution kernels until the one-dimensional original features corresponding to each one-dimensional convolution kernel are obtained.
  • the x-axis of two-dimensional plane data represents the sensor serial number
  • the y-axis represents the number of sampling time points.
  • determine a preset number of continuous sampling time points for example, two
  • slide to obtain the corresponding The data of all sensors for example, starting from the first row, two rows of data on the two-dimensional plane data are taken each time, that is, two sampling time points, until all rows of data are traversed).
  • the acquired data of preset sampling time points will be based on the sensor serial number, and the acquired data will be spliced end to end into one-dimensional sequence data (for example, after acquiring two rows of data each time, connect the two rows of data end to end into one row 1D sequence data).
  • traverse all one-dimensional convolution kernels until the one-dimensional original features corresponding to each one-dimensional convolution kernel are obtained.
  • one convolution kernel can only extract one feature map from the input vector, multiple convolution kernels are used to operate on the input vector to extract more complete fault features.
  • the pooling layer is used to down-sample the multiple one-dimensional original features output by the convolutional layer to obtain multiple one-dimensional time series features.
  • the pooling layer can extract the most important part of each feature map.
  • this operation can significantly reduce the feature dimension, which is very suitable for processing high-dimensional data.
  • the max pooling operation is expressed as:
  • y j pool(z j , p, s)
  • the timing feature y j is the result of the pooling operation
  • pool() is the maximum pooling function
  • p is the pooling size
  • s is the step size.
  • this embodiment also performs two-dimensional splicing on the multiple one-dimensional time series features obtained again to obtain a new two-dimensional plane vector, and the original one-dimensional convolution kernel Adjust the number and dimensions. Then use the new two-dimensional plane vector as a two-dimensional plane vector, and use the adjusted one-dimensional convolution kernel to repeat the operations from step six to step seven until the preset number of times. In this embodiment, three feature extraction operations are performed, so as to obtain multiple target time series features with better feature performance.
  • multiple target time series features are spliced end to end into a target one-dimensional vector, and then the target one-dimensional vector is input into the preset fully connected layer to adjust the dimension and quantity of the target one-dimensional vector through the preset fully connected layer , transform it into multiple one-dimensional abstract features with the same number of complete sampling time points of the signal, and the number of elements of each one-dimensional abstract feature is the same as the number of sensors. This ensures that the signal input data is completely consistent with the dimension of the feature data, and avoids the problem of reduced accuracy due to the reduction in the amount of data in the feature extraction process.
  • outputting the fault diagnosis result based on the second neural network specifically includes the following steps:
  • Step 11 Dimensionally expand multiple one-dimensional abstract features respectively.
  • Step 12 Perform hierarchical attention processing on the expanded multiple one-dimensional abstract features based on the linear perceptron, and fuse multiple one-dimensional abstract features to obtain an output vector.
  • Step thirteen Input the output vector into a preset classifier, so that the preset classifier outputs a fault diagnosis result.
  • a hierarchical attention mechanism is used to fuse the high-level feature information (ie, one-dimensional abstract feature) output by the first neural network.
  • dimensionality augmentation is first performed on 1D abstract features.
  • the attention network can adopt three methods to expand the dimension of the input sequence, so as to further improve the accuracy of features.
  • the first method repeats the first time data in the time dimension for a preset number of times. In a complex physical system, the initial time The value of will greatly affect the development direction and process of the physical system, so emphasizing the importance of the first moment can make the one-dimensional abstract features after dimension expansion more accurate.
  • the other two processing methods are to subtract or multiply the data at each moment and the first moment, and observe the degradation information from multiple dimensions to help the attention network learn the degradation features more comprehensively and richly.
  • the hidden state of the fault characteristics at each moment is calculated based on the linear perceptron to increase the nonlinearity of the network and better fit the degradation characteristics of the gearbox. Then the hidden state is normalized, and then the normalized input sequence is multiplied by the weight to achieve feature fusion. Finally, the feature vector output by the hierarchical attention network is input into the preset classifier (including but not limited to softmax classifier, logistic classifier) to realize gearbox fault diagnosis.
  • the preset classifier including but not limited to softmax classifier, logistic classifier
  • step 12 specifically includes the following steps:
  • Step 14 Input multiple one-dimensional abstract features into the linear perceptron in sequence to obtain multiple hidden features.
  • Step fifteen Calculate multiple weight coefficients corresponding to multiple hidden features through the softmax function, and the weight coefficients are used to represent the hierarchical attention of hidden features.
  • Step 16 Using multiple weight coefficients to perform weighted summation of multiple hidden features to obtain an output vector.
  • the perceptron learns hidden feature representation u i , whose expression is:
  • W and b denote the weights and biases of the linear perceptron, respectively.
  • the weight coefficient is calculated by the softmax function, expressed as:
  • u z is a randomly initialized vector that is continuously updated with the number of iterations.
  • the weight coefficient is used to weight the fault information of each time step to obtain the output vector s of the second neural network:
  • the parameters of each layer of the time series deep diagnosis network model composed of the first neural network and the second neural network are shown in Table 1.
  • gearbox experimental testing and simulation are as follows:
  • the gearbox In order to monitor the real-time working status of the gearbox, the gearbox is first placed in the test system.
  • the system consists of gearbox fault diagnosis platform, vibration signal sensor, rotation signal sensor, torque signal sensor, signal cable, data acquisition instrument and computer.
  • the test system is easy to operate, easy to replace parts, and can simulate normal working conditions of gearboxes, broken teeth, pitting, wear and other working conditions.
  • the modulus of the large and small gears in the gearbox is 75, the number of large gears is 55, and the number of small gears is 55. Oil-immersed lubrication is used.
  • the simulation model can be simulated and solved in Matlab after the parameter variables are determined.
  • the form of the model is In this way, the digital twin model maintains a high degree of consistency with the physical model.
  • the digital twin model can output the monitoring rotation signals ⁇ p0 , ⁇ g0 and vibration signals y p0 , y g0 , and at the same time, the model output can be compared with the rotation signals and vibration signals measured by physical sensors in real time, as an index variable reflecting the gear working condition , to achieve real-time monitoring of the health status of the gearbox.
  • min( xi ), max( xi ) respectively represent the maximum and minimum values in the vector xi .
  • each complete training subset randomly selects 80% of the samples as the training set and the remaining 20% as the validation set.
  • the hyperparameters of the time series deep diagnosis network are selected and adjusted through the test effect of the model on the verification set. These hyperparameters are determined after weighing the prediction accuracy and computing cost.
  • the final training flow chart of the model is shown in Figure 7.
  • each mini-batch contains 32 samples, and put into the training system.
  • the size of each convolution kernel in the convolution layer is 2*1, each time To move a step, in order to ensure that the edge information is not lost, the filling value of each time series is set to 1.
  • a non-linear activation function is then applied to the output of the convolutional layer to enhance the expressiveness of the network. Finally, the output is removed through the pooling layer to remove redundant information.
  • the window size is set to 2, and two steps are moved each time, that is, the feature map in one convolution unit is completed.
  • the one-dimensional convolutional network extracts the signal fault features
  • the fault features are input to the fully connected layer to obtain higher-level fault features.
  • the output data is m ⁇ n dimensional, which is the same as the original input data dimension.
  • a hierarchical attention mechanism is used to fuse high-level feature information. It calculates the hidden state of the fault feature at each moment and normalizes it using the softmax function, and then multiplies the input sequence with the weight to achieve feature fusion.
  • the feature vectors output by the hierarchical attention network are fed into a softmax classifier to realize gearbox fault diagnosis.
  • the time series deep diagnosis network uses the cross-entropy loss function as the objective function, and the backpropagation learning is used to update the weights in the network, and the Adam optimizer with adaptive adjustment ability is used for optimization.
  • the above process completes one training of the model.
  • the learning rate is set to 0.0001 to make the model converge stably.
  • the maximum number of training epochs for the model is 700.
  • the timing depth diagnosis is applied to 660 test samples to evaluate the training effect of the network.
  • the time series deep diagnosis network diagnoses the four kinds of gearbox faults, as shown in the table. It can be seen from Table 4 that the network has the highest recognition accuracy rate of 98% in the normal state of the gearbox. Among them, 4 samples were misdiagnosed, and two of them were diagnosed as pitting corrosion and broken teeth. The network has the lowest recognition accuracy rate for gear box pitting faults, but it is still higher than 90%.
  • the technical solution provided by this application establishes a digital model of the digital twin model based on the material parameters and geometric size parameters of the gear box's internal gear, and then combines the real torque signal of the gearbox to correct the parameters of the digital twin model through simulation.
  • the digital twin model of traditional technology it is more in line with the actual working conditions of the gearbox, so that the data quality of the analog signal is more in line with the real situation.
  • the real torque signal is used as the input of the model, and the simulated vibration signal and simulated rotation signal are output, which not only expands the types of gearbox fault data, but also obtains simulated signals with higher data quality, so that based on the real signal data and simulated signals of the gearbox
  • the data is used for fault diagnosis, which improves the fault diagnosis accuracy of the subsequent gearbox fault diagnosis work.
  • the present invention utilizes the first neural network for feature extraction and the second neural network with a hierarchical attention mechanism for outputting fault diagnosis results to realize fault diagnosis during fault diagnosis, so that useful sample features are further strengthened, and useless The characteristics of the sample are weakened, thereby further improving the accuracy of fault diagnosis.
  • the feature extraction step of the first neural network the fault signals collected by multiple sensors are converted from two-dimensional data to one-dimensional data, and then one-dimensional convolution is performed, and then the obtained one-dimensional features are repeatedly spliced into two-dimensional data Perform multi-level one-dimensional, two-dimensional, and one-dimensional feature extraction. Finally, the signal features with more prominent features and the same dimensions as the original data are obtained.
  • the classification of subsequent fault signals is further facilitated, and the accuracy of diagnosis is improved.
  • the feature vector is processed with hierarchical attention, which increases the nonlinearity of the neural network, thereby better fitting the degradation characteristics of the gearbox, making the fault diagnosis result closer to the real situation, and improving the accuracy of gearbox fault diagnosis. .
  • this embodiment also provides a gearbox fault diagnosis device, which includes:
  • the signal acquisition module 201 is used to acquire the real signal of the gearbox and the analog signal generated by an embodiment of the gearbox signal acquisition method.
  • the real signal includes a real torque signal, a real vibration signal and a real rotation signal.
  • the fault diagnosis module 202 is used to input the analog signal and the real signal into the preset fault diagnosis network to obtain the fault diagnosis result of the gearbox. For details, refer to the relevant description of step S201 in the above method embodiment, and details are not repeated here.
  • the preset fault diagnosis network includes a first neural network for feature extraction and a second neural network with a hierarchical attention mechanism for outputting fault diagnosis results, and the preset fault diagnosis network is based on analog signals and real signals training generated.
  • the gear box fault diagnosis device provided in the embodiment of the present invention is used to implement the gear box fault diagnosis method provided in the above embodiment. repeat.
  • the technical solution provided by this application establishes a digital model of the digital twin model based on the material parameters and geometric size parameters of the gear inside the gearbox, and then combines the real torque signal of the gearbox to correct the digital twin through simulation.
  • Model parameters Compared with the digital twin model of traditional technology, it is more in line with the actual working conditions of the gearbox, so that the data quality of the analog signal is more in line with the real situation.
  • the real torque signal is used as the input of the model, and the simulated vibration signal and simulated rotation signal are output, which not only expands the types of gearbox fault data, but also obtains simulated signals with higher data quality, so that based on the real signal data and simulated signals of the gearbox
  • the data is used for fault diagnosis, which improves the fault diagnosis accuracy of the subsequent gearbox fault diagnosis work.
  • the present invention utilizes the first neural network for feature extraction and the second neural network with a hierarchical attention mechanism for outputting fault diagnosis results to realize fault diagnosis during fault diagnosis, so that useful sample features are further strengthened, and useless The characteristics of the sample are weakened, thereby further improving the accuracy of fault diagnosis.
  • the feature extraction step of the first neural network the fault signals collected by multiple sensors are converted from two-dimensional data to one-dimensional data, and then one-dimensional convolution is performed, and then the obtained one-dimensional features are repeatedly spliced into two-dimensional data Perform multi-level one-dimensional, two-dimensional, and one-dimensional feature extraction. Finally, the signal features with more prominent features and the same dimensions as the original data are obtained.
  • the classification of subsequent fault signals is further facilitated, and the accuracy of diagnosis is improved.
  • the feature vector is processed with hierarchical attention, which increases the nonlinearity of the neural network, thereby better fitting the degradation characteristics of the gearbox, making the fault diagnosis result closer to the real situation, and improving the accuracy of gearbox fault diagnosis. .
  • FIG. 10 shows an electronic device according to an embodiment of the present invention.
  • the device includes a processor 901 and a memory 902, which may be connected via a bus or in other ways.
  • connection via a bus is taken as an example.
  • the processor 901 may be a central processing unit (Central Processing Unit, CPU).
  • the processor 901 can also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application-specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate array (Field-Programmable Gate Array, FPGA) or Other chips such as programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above-mentioned types of chips.
  • DSP Digital Signal Processor
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • Other chips such as programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above-mentioned types of chips.
  • the memory 902 as a non-transitory computer-readable storage medium, can 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 method embodiments.
  • the processor 901 executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory 902, that is, implements the methods in the above 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 and an application program required by at least one function; the data storage area may store data created by the processor 901 and the like.
  • the memory 902 may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices.
  • the memory 902 may optionally include memory located remotely relative to the processor 901, and these remote memories may be connected to the processor 901 through a network. Examples of the aforementioned 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, and when executed by the processor 901, the methods in the foregoing method embodiments are executed.
  • the storage medium can be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a flash memory (Flash Memory), a hard disk (Hard Disk Drive) , abbreviation: HDD) or solid-state drive (Solid-State Drive, SSD), etc.; the storage medium may also include a combination of the above-mentioned types of memory.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

一种齿轮箱故障诊断、信号采集方法、装置和电子设备，其中诊断方法包括获取齿轮箱的真实信号和采集方法生成的模拟信号，真实信号包括真实扭矩信号、真实振动信号和真实旋转信号（S201）；将模拟信号和真实信号分别输入预设的故障诊断网络，得到齿轮箱的故障诊断结果（S202）；其中，预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，预设的故障诊断网络基于模拟信号和真实信号训练生成。提高了传感器数据质量，并提高了齿轮箱故障诊断的准确率。

Description

一种齿轮箱故障诊断、信号采集方法、装置和电子设备 技术领域
本发明涉及算法设计领域，具体涉及一种齿轮箱故障诊断、信号采集方法、装置和电子设备。
背景技术
风电机组常年工作在沙尘暴、雷电、暴雨等恶劣环境状况下，为保证机组关键部件可靠性和安全性，对其运维检修的要求越来越严格。齿轮箱是风电机组关键传动装置，维护更换维修耗时长且成本高，同时其健康状态与风电机组稳定运行密切相关。近年来，研究人员广泛采用信号分析方法实现齿轮箱故障诊断，其方法多利用时、频域统计特征、小波变换、快速傅里叶变换、经验模态分解等从齿轮箱故障信号中提取特征，进而实现故障诊断，研究结果表明此类方法在平稳状况能取得较高准确度。但在实际情况中，齿轮箱工作状况复杂多变，信号成分多样化，多种模态混叠，导致信号具有非平稳性。常用信号分析方法会出现虚假成分干扰、特征信号分辨率低的问题，无法准确判断齿轮箱故障状态。同时，传感器最终采集信号中混有大量环境噪声，信噪比低下。此外，风场现有传感器只能采集风速、电压、电流等特征信号，无法采集或者准确测量部件应力、加速度等物理量，故不能准确反映齿轮箱故障状态。因此，提升传感器数据质量对实现齿轮箱准确故障诊断至关重要。
数字孪生技术的发展为齿轮箱在多变环境中实时状态评估提供了有效策略，它通过建立虚拟仿真模型模拟齿轮箱实际工况运行时故障演变，可与实际物理齿轮箱保持高精度仿真性。通过对齿轮箱多物理、多尺度综合仿真，能准确监测齿轮箱各个系统状态，实时输出孪生系统采集数据。最终将模型数据和传感器数据进行融合，有效弥补物理传感器采集数据的缺陷。目前，有开发人员基于数字孪生技术和用于故障检测的深度神经网络建立了齿轮箱故障诊断模型(可参考专利文件CN113505655A)，且取得了一定效果。但是目前的数字孪生模型多是基于三维模型平台进行UI创建，模型过于简便且进行分析的信号数据单一，其模拟出的故障数据对故障的表征程度还不够深入，数据质量还有待提高，从而影响齿轮箱故障诊断的最终准确率。因此，如何进一步提高传感器数据质量，从而提高齿轮箱故障诊断的准确率，是亟待解决的问题。
发明内容
有鉴于此，本发明实施方式提供了一种齿轮箱故障诊断、信号采集方法、装置和电子设备，从而提高了传感器数据质量，并提高了齿轮箱故障诊断的准确率。
根据第一方面，本发明提供了一种齿轮箱信号采集方法，所述方法包括：基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型；获取齿轮箱的真实扭矩信号，所述真实扭矩信号包括用于校正数字孪生模型参数的第一信号数据和用于计算模拟信号的第二信号数据；将所述第一信号数据输入所述数字孪生模型，以对所述数字孪生模型的参数进行校正；将所述第二信号数据输入校正后的数字孪生模型以生成模拟信号，所述模拟信号包括模拟振动信号和模拟旋转信号。
可选地，所述基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型，包括：分别获取所述齿轮箱内主动轮和从动轮的材料参数和几何尺寸参数，所述材料参数包括质量参数、惯量参数、刚度参数和阻尼参数中的至少一种；基于所述主动轮和从动轮之间啮合的物理结构，利用所述材料参数和几何尺寸参数创建齿轮箱微分动力学方程；通过所述齿轮箱微分动力学方程计算用于输出所述模拟信号的模拟矩阵；利用所述模拟矩阵、质量惯量矩阵、阻尼矩阵和刚度矩阵创建用于表征齿轮载荷关系的数学模型，并将所述数学模型作为所述数字孪生模型；其中，所述质量惯量矩阵基于所述质量参数和所述惯量参数构建，所述阻尼矩阵基于所述阻尼参数和所述几何尺寸参数构建，所述刚度矩阵基于所述刚度参数和所述几何尺寸参数构建。
根据第二方面，本发明提供了一种齿轮箱故障诊断方法，所述方法包括：获取齿轮箱的真实信号和如第一方面任意一种可选实施方式生成的模拟信号，所述真实信号包括真实扭矩信号、真实振动信号和真实旋转信号；将所述模拟信号和所述真实信号分别输入预设的故障诊断网络，得到所述齿轮箱的故障诊断结果；其中，所述预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，所述预设的故障诊断网络基于所述模拟信号和所述真实信号训练生成。
可选地，将所述模拟信号和/或所述真实信号作为目标信号，基于所述第一神经网络对所述目标信号进行特征提取的步骤，包括：以所述目标信号的采样时间点数和传感器数量为二维平面轴，将所述目标信号拼接为二维平面向量；基于多个维度相同的一维卷积核对所述二维平面向量进行一维卷积计算，得到多个一维原始特征；分别对各个所述一维原始特征进行池化得到多个一维时序特征；将多个所述一维时序特征进行二维拼接得到新二维平面向量，并对所述一维卷积核的个数和维度进行调整；将所述新二维平面向量作为所述二维平面向量，基于调整后的所述一维卷积核，返回所述基于多个维度相同的一维卷积核对所述二维平面向量进行一维卷积计算，得到多个一维原始特征的步骤，直至返回预设次数为止，得到多个目标时序特征；将所述多个目标时序特征拼接为目标一维向量，并将所述目标一维向量输入预设的全连接层，以通过所述预设的全连接层将所述目标一维向量转化为与所述采样时间点数数量相同的多个一维抽象特征，其中所述一维抽象特征的元素个数与所述传感器数量相同。
可选地，所述基于多个维度相同的一维卷积核对所述二维平面向量进行一维卷积计算，得到多个一维原始特征，包括：针对各个传感器元素从所述二维平面向量中滑动提取多组预设采样时间点数的元素，并将所述多组预设采样时间点数的元素组成多个一维待卷积向量；将各个一维待卷积向量依次与当前一维卷积核进行卷积运算，生成当前一维原始特征；基于所述将各个一维待卷积向量依次与当前一维卷积核进行卷积运算，生成当前一维原始特征的步骤，遍历全部一维卷积核，直至得到各个一维卷积核对应的一维原始特征为止。
可选地，基于所述第二神经网络输出故障诊断结果的步骤，包括：基于线性感知机对所述多个一维抽象特征进行层次注意力处理，并融合所述多个一维抽象特征得到输出向量；将所述输出向量输入预设的分类器，以使所述预设的分类器输出所述故障诊断结果。
可选地，基于线性感知机对所述多个一维抽象特征进行层次注意力处理，并融合所述多个一维抽象特征得到输出向量，包括：将所述多个一维抽象特征依次输入所述线性 感知机，得到多个隐藏特征；通过softmax函数计算所述多个隐藏特征对应的多个权重系数，所述权重系数用于表征所述隐藏特征的层次注意力；利用所述多个权重系数对所述多个隐藏特征进行加权求和，得到所述输出向量。
根据第三方面，本发明提供了一种齿轮箱故障诊断装置，所述装置包括：信号采集模块，用于获取齿轮箱的真实信号和如第一方面任意一种可选实施方式生成的模拟信号，所述真实信号包括真实扭矩信号、真实振动信号和真实旋转信号；故障诊断模块，用于将所述模拟信号和所述真实信号分别输入预设的故障诊断网络，得到所述齿轮箱的故障诊断结果；其中，所述预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，所述预设的故障诊断网络基于所述模拟信号和所述真实信号训练生成。
根据第四方面，本发明实施例提供了一种电子设备，包括：存储器和处理器，所述存储器和所述处理器之间互相通信连接，所述存储器中存储有计算机指令，所述处理器通过执行所述计算机指令，从而执行第一方面，或者第一方面任意一种可选实施方式中所述的方法。
根据第五方面，本发明实施例提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机指令，所述计算机指令用于使所述计算机从而执行第一方面，或者第一方面任意一种可选实施方式中所述的方法。
本申请提供的技术方案，具有如下优点：
本申请提供的技术方案，基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型的数字模型，之后结合齿轮箱的真实扭矩信号，通过仿真校正数字孪生模型参数。相比传统技术的数字孪生模型，更符合齿轮箱实际工况，以使模拟信号的数据质量更符合真实情况。从而以真实扭矩信号作为模型的输入，输出模拟振动信号和模拟旋转信号，不仅扩充了齿轮箱故障数据种类，还得到了数据质量更高的模拟信号，从而基于齿轮箱的真实信号数据和模拟信号数据进行故障诊断，提高了后续的齿轮箱故障诊断工作的故障诊断准确性。
此外，本发明在故障诊断时利用用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络实现故障诊断，使有用的样本特征进一步强化，无用的样本特征弱化，从而进一步提高故障诊断的准确率。其中，第一神经网络的特征提取步骤，将多个传感器采集的故障信号从二维数据转换为一维数据，再进行一维卷积，然后将得到的一维特征再反复拼接为二维数据进行多层次的一维、二维、一维的特征提取。最终得到特征更突出、且维度与原始数据维度相同的信号特征。从而进一步使后续故障信号的分类的更加容易，提高诊断准确率。基于线性感知机对特征向量进行层次注意力处理，增加了神经网络的非线性，从而更好地拟合齿轮箱的退化特征，使得故障诊断结果更加接近真实情况，提高齿轮箱故障诊断的准确率。
附图说明
通过参考附图会更加清楚的理解本发明的特征和优点，附图是示意性的而不应理解为对本发明进行任何限制，在附图中：
图1示出了本发明一个实施方式中一种齿轮箱信号采集方法的步骤示意图；
图2示出了本发明一个实施方式中一种齿轮箱信号采集方法的流程框图；
图3示出了现有技术中典型直齿圆柱齿轮副啮合动力学结构示意图；
图4示出了本发明一个实施方式中一种齿轮箱故障诊断方法的步骤示意图；
图5示出了本发明一个实施方式中一种齿轮箱故障诊断方法的流程示意图；
图6示出了本发明一个实施方式中第一神经网络的结构示意图；
图7示出了本发明一个实施方式中另一个一种齿轮箱故障诊断方法的流程示意图；
图8示出了本发明一个实施方式中深度时序诊断网络损失函数随迭代次数变化的曲线图；
图9示出了本发明一个实施方式中一种齿轮箱信号采集装置的结构示意图；
图10示出了本发明一个实施方式中一种电子设备的结构示意图。
具体实施方式
为使本发明实施方式的目的、技术方案和优点更加清楚，下面将结合本发明实施方式中的附图，对本发明实施方式中的技术方案进行清楚、完整地描述，显然，所描述的实施方式是本发明一部分实施方式，而不是全部的实施方式。基于本发明中的实施方式，本领域技术人员在没有作出创造性劳动前提下所获得的所有其他实施方式，都属于本发明保护的范围。
请参阅图1和图2，在一个实施方式中，一种齿轮箱信号采集方法，具体包括以下步骤：
步骤S101：基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型。
步骤S102：获取齿轮箱的真实扭矩信号，真实扭矩信号包括用于校正数字孪生模型参数的第一信号数据和用于计算模拟信号的第二信号数据，并将第一信号数据输入数字孪生模型，以对数字孪生模型的参数进行校正。
步骤S103：将第二信号数据输入校正后的数字孪生模型以生成模拟信号，模拟信号包括模拟振动信号和模拟旋转信号。
具体地，在本实施例中，为了进一步提高模拟故障信号的信号质量，从而相较传统技术单纯使用三维模型创建数字孪生模型的方法，基于齿轮箱内齿轮的材料参数和几何尺寸参数建立齿轮箱的数字孪生模型，包括但不限于齿轮大小、齿轮材料、材料刚度、齿轮间阻尼、齿轮质量。从而使数字孪生模型的模拟工况更加贴近齿轮箱运行的真实工况。传统数字孪生模型，通常是对单一振动信号的模拟。而在本实施例中，通过创建振动信号、旋转信号与扭矩信号的等式关系，然后先基于有限元仿真软件使用真实扭矩信号对数字孪生模型中的几何尺寸参数和材料参数进行校正。再将真实扭矩信号作为激励输入矫正后的数字孪生模型，得到表现优秀的模拟振动信号和模拟旋转信号。从而使数字孪生模型输出的模拟信号类型更多、模拟信号质量更好，从而提高后续齿轮箱的故障诊断准确率。
具体地，在一实施例中，上述步骤S101，具体包括如下步骤：
步骤一：分别获取齿轮箱内主动轮和从动轮的材料参数和几何尺寸参数，材料参数包括质量参数、惯量参数、刚度参数和阻尼参数中的至少一种。
步骤二：基于主动轮和从动轮之间啮合的物理结构，利用材料参数和几何尺寸参 数创建齿轮箱微分动力学方程。
步骤三：通过齿轮箱微分动力学方程计算用于输出模拟信号的模拟矩阵。
步骤四：利用模拟矩阵、质量惯量矩阵、阻尼矩阵和刚度矩阵创建用于表征齿轮载荷关系的数学模型，并将数学模型作为数字孪生模型。
其中，质量惯量矩阵基于质量参数和惯量参数构建，阻尼矩阵基于阻尼参数和几何尺寸参数构建，刚度矩阵基于刚度参数和几何尺寸参数构建。
具体地，在本实施例中，基于齿轮载荷关系创建数学等式模型，从而实现数字孪生模型的创建。首先获取包括质量参数、惯量参数、刚度参数和阻尼参数的材料参数，再获取表征齿轮大小的几何尺寸参数。然后构建质量惯量矩阵、阻尼矩阵和刚度矩阵，本实施例在建立齿轮箱仿真模型时采用假设构建理想状态的齿轮箱：
1)假设各构件均为刚体，忽略齿轮和系杆的整体弹性变形；
2)假设相互啮合的齿轮间不发生脱齿；
3)忽略啮合反向冲击现象。
构成模型所需的参数矩阵如下：
质量惯量矩阵
Figure PCTCN2022112143-appb-000001
阻尼矩阵
Figure PCTCN2022112143-appb-000002
刚度矩阵
Figure PCTCN2022112143-appb-000003
其中质量参数为m p与m g(p和g分别表示主动轮和从动轮)惯量参数为I p与I g，刚度参数为k m与k py、k gy，阻尼参数为c m与c py、c gy，其中k m、c m分别是齿轮副啮合综合刚度和综合阻尼，k py、k gy为主、被动齿轮平移振动的刚度，c py、c gy为主、被动齿轮平移振动的阻尼。以及几何尺寸R p、R g。
之后，根据齿轮物理结构，如图3所示，构建齿轮箱微分动力学方程，其表达式为：
Figure PCTCN2022112143-appb-000004
其中，F k、F c分别为弹性啮合力和粘性啮合力，F p、F g分别为主、被动齿轮上的轮齿动态啮合力，在方程化解过程中会被消去。θ p、θ g是用于表征旋转信号的旋转变量，y p、y g是用于表征振动信号的平移变量，T p0、T g0用于表征扭矩信号。通过上述动力学模型可求解用于输出模拟信号的模拟矩阵，即平移阵列旋转{δ}＝{y p θ p y g θ g} T。
之后，将物理传感器中监测得到的真实扭矩信号T p0、T g0传入数字孪生模型，即可实时更新数字孪生模型参数T p、T g，并利用有限元仿真软件模拟仿真校正材料参数和几何尺寸参数。得到更新后的载荷矩阵
Figure PCTCN2022112143-appb-000005
仿真模型在参数变量确定后，即可在Matlab中进行仿真求解，基于载荷关系建立的数字孪生模型的形式为
Figure PCTCN2022112143-appb-000006
从而使数字孪生模型与物理模型保持高度一致性。通过上述步骤，使数字孪生模型的模拟工况更加贴近齿轮箱运行的真实工况，使数字孪生模型输出的模拟信号类型更多、模拟信号质量更好，从而提高后续齿轮箱的故障诊断准确率。
如图4和图5所示，在一个实施方式中，一种齿轮箱故障诊断方法，具体包括以下步骤：
步骤S201：获取齿轮箱的真实信号和基于上述步骤S101～S103方法生成的模拟信号，真实信号包括真实扭矩信号、真实振动信号和真实旋转信号，模拟信号包括模拟振动信号和模拟旋转信号；
步骤S202：将模拟信号和真实信号分别输入预设的故障诊断网络，得到齿轮箱的故障诊断结果。
其中，预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，预设的故障诊断网络基于模拟信号和真实信号训练生成。
具体地，在本实施例中，将步骤S101～S103生成的模拟信号与真实信号联合作为预设的故障诊断网络的输入数据，提高了信号类型和信号质量，从而提高了故障诊断网络对故障信号的分类准确率。此外，在本实施例中，预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络。第二神经网络基于层次注意力机制，可通过将第一神经网络提取的特征划分到多个具有分-总逻辑形式的层次上，形成多层次注意力。将注意力集中在场景相关的部分，而忽略不相关部分，注意力机制本质上是信息处理资源的有效分配，通过对模型输入特征赋予不同的权重，将注意力集中在对当前认为更为关键的信息上，突出关键特征对模型的影响，注意力机制可解决信息过载问题，提高模型效率和预测性能。从而基于更深度更关键的特征信息，进一步取得准确度更高的故障诊断结果。
具体地，在一实施例中，将模拟信号和/或真实信号作为目标信号，基于第一神经网络对目标信号进行特征提取，具体包括如下步骤：
步骤五：以目标信号的采样时间点数和传感器数量为二维平面轴，将目标信号拼接为二维平面向量。
步骤六：基于多个维度相同的一维卷积核对二维平面向量进行一维卷积计算，得到多个一维原始特征。
步骤七：分别对各个一维原始特征进行池化得到多个一维时序特征。
步骤八：将多个一维时序特征进行二维拼接得到新二维平面向量，并对一维卷积核的个数和维度进行调整。
步骤九：将新二维平面向量作为二维平面向量，基于调整后的一维卷积核，返回步骤六，直至返回预设次数为止，得到多个目标时序特征。
步骤十：将多个目标时序特征拼接为目标一维向量，并将目标一维向量输入预设的全连接层，以通过预设的全连接层将目标一维向量转化为与采样时间点数数量相同的多个一维抽象特征，其中一维抽象特征的元素个数与传感器数量相同。
具体地，在本实施例中，采用与现有技术不同的一维卷积方法，可获取与输入的信号数据维度相同且特征更加多样的特征数据。如图6所示，卷积神经网络由一个或多个卷积层、激活函数和池化层组成的监督学习神经网络。网络通过梯度下降法优化目标函数，不断迭代训练更新网络参数。卷积层和池化层是卷积神经网络特征提取器的核心组成部分。卷积层中每个节点的输入只是上一层神经网络中局部范围内的神经元，卷积层将它们进行更加深入的分析从而得到抽象程度更高的特征。池化层常取这些窗口中的最大值作为输出的结果，然后不断滑动窗口，对输入数据体的每一个深度切片单独处理，从而减少网络参数。
具体地，以目标信号的采样时间点数和传感器数量为二维平面轴，将目标信号拼接为二维平面向量。即初始输入第一神经网络的信号数据，是以采样时间为y轴，以信号传感器的序号为x轴构建的二维平面数据。然后，基于多个大小相同的一维卷积核对二维平面数据进行一维卷积计算，从而得到第一次提取的信号特征。相较现有技术利用一维卷积核对输入的一维信号直接进行卷积方式，本实施例将多组信号拼成二维数据再进行一维卷积，使得不同信号序列进行混叠，得到的特征信息综合性更强，特征信息更具有代表性。此外，相比现有技术一维卷积核直接在二维平面数据上逐行或者逐列进行卷积计算的方法，本实施例与其区别的具体操作步骤如下：
1.针对各个传感器元素从二维平面向量中滑动提取多组预设采样时间点数的元素，并将多组预设采样时间点数的元素组成多个一维待卷积向量；
2.将各个一维待卷积向量依次与当前一维卷积核进行卷积运算，生成当前一维原始特征；
3.基于步骤2，遍历全部一维卷积核，直至得到各个一维卷积核对应的一维原始特征为止。
具体地，例如二维平面数据x轴表示传感器序号，y轴表示采样时间点数，在y轴上确定预设个数的连续采样时间点(例如两个)，然后滑动获取与采样时间点对应的所有传感器的数据(例如从第一行开始，每次取二维平面数据上的两行数据，即两个采样时间点，直至把所有行的数据均遍历到为止)。之后将获取的预设采样时间点数的数据以传感器序号为基准，将获取的数据首尾相接拼接为一维序列数据(例如每次在获取两行数据之后，将两行数据首尾相接为一行一维序列数据)。然后将每次拼接的得到的一维序列数据进行一次卷积计算，直至将全部一维序列数据进行卷积计算，得到的多个计算结果，再将各个计算结果拼接为一维的特征向量，即得到了一维原始特征。最后，基于步骤2，遍历全部一维卷积核，直至得到各个一维卷积核对应的一维原始特征为止。
即，对于给定二维传感器数据
Figure PCTCN2022112143-appb-000007
其中n是传感器时间序列的长度，则第i个时间步对应的数据为x i＝[x 1，i x 2，i L x m，i] T，m表示传感器的个数。同时，让向量X i：i+t-1代表第i输入样本，为如下所示：
Figure PCTCN2022112143-appb-000008
其中i＝1，2，L，n-t+1，
Figure PCTCN2022112143-appb-000009
代表将不同的时间窗口组合起来。向量X i：i+t-1与第j个卷积核
Figure PCTCN2022112143-appb-000010
进行运算，第i个时间步提取到的特征表示为：
Figure PCTCN2022112143-appb-000011
其中b j和f分别为偏置项和激活函数。当卷积核在输入向量中从上向下运动时，即可得到从输入向量中提取的一维原始特征z j，
Figure PCTCN2022112143-appb-000012
由于一个卷积核只能从输入向量中提取一个特征图，故使用多个卷积核对输入向量进行操作，从而提取到更完整的故障特征。
然后在步骤七，使用池化层对卷积层输出的多个一维原始特征进行下采样操作，得到多个一维时序特征。一方面池化层能提取每个特征映射中最重要部分，另一方面该操作可显著降低特征维数，非常适合处理高维数据。最大池化操作表示为：
y j＝pool(z j，p，s)
其中时序特征y j是池化操作的结果，pool()是最大池化函数，p是池化大小，s是步长。
之后，为了进一步提高提取的特征更加完整、具有代表性，本实施例还将得到的多个一维时序特征再次进行二维拼接得到新二维平面向量，并对原先的一维卷积核的个数和维度进行调整。然后将新二维平面向量作为二维平面向量，利用调整后的一维卷积核，反复进行步骤六到步骤七的操作，直至预设次数为止。在本实施例中进行了3次特征提取的操作，从而得到特征表现更加优秀的多个目标时序特征。
最后，将多个目标时序特征首尾拼接为一个目标一维向量，然后将目标一维向量输入预设的全连接层，以通过预设的全连接层对目标一维向量的维度和数量进行调整，将其转化为与信号完整的采样时间点数数量相同的多个一维抽象特征，并且其中每个一维抽象特征的元素个数与传感器数量相同。从而保证了信号输入数据与特征数据维度完全一致，避免因为特征提取过程数据量的减少导致准确率降低的问题。
具体地，在一实施例中，基于第二神经网络输出故障诊断结果，具体包括如下步骤：
步骤十一：分别将多个一维抽象特征进行维度扩充。
步骤十二：基于线性感知机对扩充后的多个一维抽象特征进行层次注意力处理，并融合多个一维抽象特征得到输出向量。
步骤十三：将输出向量输入预设的分类器，以使预设的分类器输出故障诊断结果。
具体地，在本实施例中，层次注意力机制被用来融合第一神经网络输出的高级特征信息(即一维抽象特征)。在此之前，首先对一维抽象特征进行维度扩充。注意力网络可采取三种方式对输入序列进行维度扩充，从而进一步提高特征准确度，第一种方式将第一个时刻数据在时间维度上重复预设次数，在一个复杂物理系统中，初始时刻的取值会极大地影响物理系统的发展方向和进程，故强调第一个时刻的重要性可以使维度扩充后的一维抽象特征准确度更高。其余两种处理方式分别是每个时刻与第一个时刻数据相减或相乘，从多维度观察退化信息，帮助注意力网络更全面、更丰富地学习退化特征。
然后，基于线性感知机计算每个时刻故障特征的隐藏状态，以增加网络的非线性，更好拟合齿轮箱的退化特征。然后对隐藏状态进行归一化处理，再将归一化处理的输入序列与权重相乘实现特征融合。最后将层次注意力网络输出的特征向量输入预设分类器(包括但不限于softmax分类器、logistic分类器)，实现齿轮箱故障诊断。
具体地，在一实施例中，上述步骤十二，具体包括如下步骤：
步骤十四：将多个一维抽象特征依次输入线性感知机，得到多个隐藏特征。
步骤十五：通过softmax函数计算多个隐藏特征对应的多个权重系数，权重系数用于表征隐藏特征的层次注意力。
步骤十六：利用多个权重系数对多个隐藏特征进行加权求和，得到输出向量。
具体地，假设第二神经网络输入的特征表示为H＝[h 1 h 2 L h i L h n]，其中h i代表第i个时间步的一维抽象特征，首先将h i输入至线性感知机学习隐藏特征表达u i，其表达式为：
u i＝tanh(Wh i+b)
其中W和b分别表示线性感知机的权重和偏置。
然后计算每一个时间步的权重系数，系数越大表明该时间步包含更多的故障信息，权重系数通过softmax函数计算得到，表示为：
Figure PCTCN2022112143-appb-000013
其中u z是一个随机初始化的向量，随着迭代次数不断更新。最终使用权重系数对每个时间步故障信息进行加权，得到第二神经网络的输出向量s：
Figure PCTCN2022112143-appb-000014
通过上述步骤，从而准确计算出包含更完整、更准确特征的输出向量，用于故障信号的分类。
在本实施例中，第一神经网络和第二神经网络组成的时序深度诊断网络模型每一层参数如表1所示。
表1时序深度诊断网络模型结构描述
Figure PCTCN2022112143-appb-000015
具体地，在一实施例中，齿轮箱实验测试与仿真的应用实例如下：
为监测齿轮箱实时工作状态，首先将齿轮箱置于测试系统中。该系统由齿轮箱故障诊断平台、振动信号传感器、旋转信号传感器、扭矩信号传感器、信号电缆、数据采集仪及计算机组成。该测试系统操作方便，便捷更换零部件，可模拟齿轮箱正常工况、断齿故障、点蚀故障、磨损故障等工况。齿轮箱中大小齿轮模数为，大齿轮数为75，小齿轮数为55，采用浸油式润滑。
首先测量齿轮基本几何尺寸参数、材料参数，建立动力学模型的初始输入参数，包括质量参数m p与m g，惯量参数I p与I g，刚度参数k m与k py、k gy，阻尼参数c m与c py、c gy，以及几何尺寸R p、R g。接着可构建模型所需的参数矩阵和齿轮箱微分动力学模型，然后通过动力学模型可求解平移阵列旋转{δ}＝{y p θ p y g θ g} T中旋转变量θ p、θ g，平移(振动)变量y p、y g，即可实现对仿真模型中旋转信号θ p0、θ g0，振动信号y p0、y g0以及扭矩信号T p0、T g0的监测采集。然后将物理传感器中监测得到的扭矩信号 T p0、T g0传入数字孪生模型，即可实时更新数字孪生模型参数T p、T g，则得到更新后的载荷矩阵
Figure PCTCN2022112143-appb-000016
仿真模型在参数变量确定后，即可在Matlab中进行仿真求解，模型的形式为
Figure PCTCN2022112143-appb-000017
从而使数字孪生模型与物理模型保持高度一致性。
最后数字孪生模型可输出监测旋转信号θ p0、θ g0与振动信号y p0、y g0，同时可将模型输出与物理传感器测量旋转信号和振动信号进行实时比对，作为反映齿轮工况的指标变量，实现齿轮箱健康状态实时监测。
本实验在转速为880rpm不带载情况下，使用5120Hz采样率分别采集了正常、点蚀、断齿、磨损四种齿轮工作时的传感器信号，同时在仿真模型中也输出了始终齿轮工作时的传感器信号。具体信息见表2所示，所有样本总共有3300个，其中正常状态、磨损状态各是1000个样本、点蚀状态是800个样本，断齿样本500个。每个样本中均包含齿轮箱连续工作时的物理传感器记录和数字孪生模型输出的旋转信号θ p0、θ g0与振动信号y p0、y g0。它们构成了时序深度诊断模型的数据集，用于训练网络参数并测试网络性能。
表2实验采集样本数量信息
状态 样本数
正常 1000
点蚀 800
断齿 500
磨损 1000
由于不同传感器监测的数据表示不同的物理特性，导致各参数的量纲和数量级不同，因此原始数据无法直接反映各个参数的波动情况。为了避免预测的过程参数的量纲和取值范围影响，需对数据统一进行归一化处理，把全部属性的值规约到一个相同的取值空间。归一化可以缩小数值之间的差异，避免数据偏置问题，方便后续训练，同时也有利于提高预测模型的准确性和收敛性。考虑到这一点，使用最小-最大方式归一化。假设给定输入
Figure PCTCN2022112143-appb-000018
其中N代表时间步长，K代表传感器数目。第i个向量使用最大最小归一化方法如下：
Figure PCTCN2022112143-appb-000019
其中min(x i)，max(x i)分别代表向量x i中的最大最小值。
由于齿轮箱故障类型共有四类，同时为方便网络直接输入分类结果，对每种故障采用独热编码，其结果如表3所示。
表3齿轮箱故障类型独热编码
Figure PCTCN2022112143-appb-000020
在训练过程中，每个完整的训练子集随机选择80％的样本作为训练集，其余20％为验证集。通过模型在验证集上的测试效果对时序深度诊断网络的超参数进行选择和调整，这些超参数是通过权衡考虑预测精度与计算成本后确定，模型最终的训练流程图如图7所示。
对于每轮训练，将样本随机分为多个小批量，每个小批量包含32个样本，并放入训练系统。首先使用3个一维卷积单元从原始数据中提取特征，在每个卷积层使用多个滤波器提取数据局部特征，卷积层中每个卷积核的大小为2*1，每次移动一个步长，为保证边缘信息不丢失，对每个时间序列的填充值设置为1。然后对卷积层输出应用非线性激活函数，增强网络表达能力。最后将输出通过池化层去除冗余信息，池化层中设置窗口大小为2，每次移动两个步长，即完成一个卷积单元内的特征映射。当一维卷积网络提取信号故障特征后，将故障特征输入至全连接层得到更高级的故障特征，输出数据为m×n维，与原始输入数据维度相同。接着，层次注意力机制被用来融合高级特征信息。它计算每个时刻故障特征的隐藏状态并使用softmax函数归一化，然后将输入序列与权重相乘实现特征融合。最后将层次注意力网络输出特征向量输入softmax分类器，实现齿轮箱故障诊断。时序深度诊断网络使用交叉熵损失函数作为目标函数，反向传播学习用于网络中权值的更新，同时采用具有自适应调整能力的Adam优化器进行优化。上述过程即完成模型一次训练，模型训练时将学习率设置为0.0001使模型稳定收敛。默认情况下，模型最大训练时段数为700。
利用pytorch深度学习框架构建深度时序诊断网络，设定网络的迭代次数为700，每次输入16个样本，学习率0.0001。使用梯度下降法对网络参数进行优化，网络损失函数训练曲线如图7所示。
由图8可知，经过700次迭代后，网络误差已经下降到足够小，证明网络能有效地拟合训练样本。将时序深度诊断应用于660个测试样本，评估网络的训练效果。时序深度诊断网络对齿轮箱四种故障诊断结果如表所示。由表4可知，网络对齿轮箱正常状态下识别准确 率最高，为98％，其中有4个样本被误诊，各有两个被诊断为点蚀和断齿故障。网络对齿轮箱点蚀故障下的识别准确率最低，但仍高于90％，其中有10个样本被误诊为其他状态，其中2个样本诊断为断齿故障，8个样本被诊断为磨损故障，这可能是因为点蚀故障和磨损故障的特征有一些相似，从而导致网络无法有效判别属于哪一种故障。网络对于齿轮断齿故障和磨损故障的识别准确率分别为96％和94％，均取得了较好的效果。对于总计660个测试样本，一共有30个样本误诊，模型的整体识别率为95.5％，表明时序深度诊断网络能够有效地识别齿轮箱四种故障。
表4时序深度诊断网络故障预测结果
Figure PCTCN2022112143-appb-000021
通过上述步骤，本申请提供的技术方案，基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型的数字模型，之后结合齿轮箱的真实扭矩信号，通过仿真校正数字孪生模型参数。相比传统技术的数字孪生模型，更符合齿轮箱实际工况，以使模拟信号的数据质量更符合真实情况。从而以真实扭矩信号作为模型的输入，输出模拟振动信号和模拟旋转信号，不仅扩充了齿轮箱故障数据种类，还得到了数据质量更高的模拟信号，从而基于齿轮箱的真实信号数据和模拟信号数据进行故障诊断，提高了后续的齿轮箱故障诊断工作的故障诊断准确性。
此外，本发明在故障诊断时利用用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络实现故障诊断，使有用的样本特征进一步强化，无用的样本特征弱化，从而进一步提高故障诊断的准确率。其中，第一神经网络的特征提取步骤，将多个传感器采集的故障信号从二维数据转换为一维数据，再进行一维卷积，然后将得到的一维特征再反复拼接为二维数据进行多层次的一维、二维、一维的特征提取。最终得到特征更突出、且维度与原始数据维度相同的信号特征。从而进一步使后续故障信号的分类的更加容易，提高诊断准确率。基于线性感知机对特征向量进行层次注意力处理，增加了神经网络的非线性，从而更好地拟合齿轮箱的退化特征，使得故障诊断结果更加接近真实情况，提高齿轮箱故障诊断的准确率。
如图9所示，本实施例还提供了一种齿轮箱故障诊断装置，该装置包括：
信号采集模块201，用于获取齿轮箱的真实信号和一种齿轮箱信号采集方法实施例生成的模拟信号，真实信号包括真实扭矩信号、真实振动信号和真实旋转信号。详细内容参见上述方法实施例中步骤S201的相关描述，在此不再进行赘述。
故障诊断模块202，用于将模拟信号和真实信号分别输入预设的故障诊断网络，得 到齿轮箱的故障诊断结果。详细内容参见上述方法实施例中步骤S201的相关描述，在此不再进行赘述。
其中，预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，预设的故障诊断网络基于模拟信号和真实信号训练生成。
本发明实施例提供的一种齿轮箱故障诊断装置，用于执行上述实施例提供的一种齿轮箱故障诊断方法，其实现方式与原理相同，详细内容参见上述方法实施例的相关描述，不再赘述。
通过上述各个组成部分的协同合作，本申请提供的技术方案，基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型的数字模型，之后结合齿轮箱的真实扭矩信号，通过仿真校正数字孪生模型参数。相比传统技术的数字孪生模型，更符合齿轮箱实际工况，以使模拟信号的数据质量更符合真实情况。从而以真实扭矩信号作为模型的输入，输出模拟振动信号和模拟旋转信号，不仅扩充了齿轮箱故障数据种类，还得到了数据质量更高的模拟信号，从而基于齿轮箱的真实信号数据和模拟信号数据进行故障诊断，提高了后续的齿轮箱故障诊断工作的故障诊断准确性。
此外，本发明在故障诊断时利用用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络实现故障诊断，使有用的样本特征进一步强化，无用的样本特征弱化，从而进一步提高故障诊断的准确率。其中，第一神经网络的特征提取步骤，将多个传感器采集的故障信号从二维数据转换为一维数据，再进行一维卷积，然后将得到的一维特征再反复拼接为二维数据进行多层次的一维、二维、一维的特征提取。最终得到特征更突出、且维度与原始数据维度相同的信号特征。从而进一步使后续故障信号的分类的更加容易，提高诊断准确率。基于线性感知机对特征向量进行层次注意力处理，增加了神经网络的非线性，从而更好地拟合齿轮箱的退化特征，使得故障诊断结果更加接近真实情况，提高齿轮箱故障诊断的准确率。
图10示出了本发明实施例的一种电子设备，该设备包括处理器901和存储器902，可以通过总线或者其他方式连接，图10中以通过总线连接为例。
处理器901可以为中央处理器(Central Processing Unit，CPU)。处理器901还可以为其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等芯片，或者上述各类芯片的组合。
存储器902作为一种非暂态计算机可读存储介质，可用于存储非暂态软件程序、非暂态计算机可执行程序以及模块，如上述方法实施例中的方法所对应的程序指令/模块。处理器901通过运行存储在存储器902中的非暂态软件程序、指令以及模块，从而执行处理器的各种功能应用以及数据处理，即实现上述方法实施例中的方法。
存储器902可以包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需要的应用程序；存储数据区可存储处理器901所创建的数据等。此外，存储器902可以包括高速随机存取存储器，还可以包括非暂态存储器，例如至少一个磁盘存储器件、闪存器件、或其他非暂态固态存储器件。在一些实施例中，存储器902可选包括相对 于处理器901远程设置的存储器，这些远程存储器可以通过网络连接至处理器901。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
一个或者多个模块存储在存储器902中，当被处理器901执行时，执行上述方法实施例中的方法。
上述电子设备具体细节可以对应参阅上述方法实施例中对应的相关描述和效果进行理解，此处不再赘述。
本领域技术人员可以理解，实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，实现的程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory，ROM)、随机存储记忆体(Random Access Memory，RAM)、快闪存储器(Flash Memory)、硬盘(Hard Disk Drive，缩写：HDD)或固态硬盘(Solid-State Drive，SSD)等；存储介质还可以包括上述种类的存储器的组合。
虽然结合附图描述了本发明的实施例，但是本领域技术人员可以在不脱离本发明的精神和范围的情况下作出各种修改和变型，这样的修改和变型均落入由所附权利要求所限定的范围之内。

Claims (9)

  1. 一种齿轮箱信号采集方法，其特征在于，所述方法包括：
    基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型；
    获取齿轮箱的真实扭矩信号，所述真实扭矩信号包括用于校正数字孪生模型参数的第一信号数据和用于计算模拟信号的第二信号数据；
    将所述第一信号数据输入所述数字孪生模型，以对所述数字孪生模型的参数进行校正；
    将所述第二信号数据输入校正后的数字孪生模型以生成模拟信号，所述模拟信号包括模拟振动信号和模拟旋转信号；
    其中，所述基于齿轮箱内齿轮的材料参数和几何尺寸参数建立数字孪生模型，包括：
    分别获取所述齿轮箱内主动轮和从动轮的材料参数和几何尺寸参数，所述材料参数包括质量参数、惯量参数、刚度参数和阻尼参数中的至少一种；
    基于所述主动轮和从动轮之间啮合的物理结构，利用所述材料参数和几何尺寸参数创建齿轮箱微分动力学方程；
    通过所述齿轮箱微分动力学方程计算用于输出所述模拟信号的模拟矩阵；
    利用所述模拟矩阵、质量惯量矩阵、阻尼矩阵和刚度矩阵创建用于表征齿轮载荷关系的数学模型，并将所述数学模型作为所述数字孪生模型；
    其中，所述质量惯量矩阵基于所述质量参数和所述惯量参数构建，所述阻尼矩阵基于所述阻尼参数和所述几何尺寸参数构建，所述刚度矩阵基于所述刚度参数和所述几何尺寸参数构建。
  2. 一种齿轮箱故障诊断方法，其特征在于，所述方法包括：
    获取齿轮箱的真实信号和如权利要求1所述方法生成的模拟信号，所述真实信号包括真实扭矩信号、真实振动信号和真实旋转信号；
    将所述模拟信号和所述真实信号分别输入预设的故障诊断网络，得到所述齿轮箱的故障诊断结果；
    其中，所述预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，所述预设的故障诊断网络基于所述模拟信号和所述真实信号训练生成。
  3. 根据权利要求2所述的方法，其特征在于，将所述模拟信号和/或所述真实信号作为目标信号，基于所述第一神经网络对所述目标信号进行特征提取的步骤，包括：
    以所述目标信号的采样时间点数和传感器数量为二维平面轴，将所述目标信号拼接为二维平面向量；
    基于多个维度相同的一维卷积核对所述二维平面向量进行一维卷积计算，得到多个一维原始特征；
    分别对各个所述一维原始特征进行池化得到多个一维时序特征；
    将多个所述一维时序特征进行二维拼接得到新二维平面向量，并对所述一维卷积核的个数和维度进行调整；
    将所述新二维平面向量作为所述二维平面向量，基于调整后的所述一维卷积核，返回所述基于多个维度相同的一维卷积核对所述二维平面向量进行一维卷积计算，得到多个一维原始特征的步骤，直至返回预设次数为止，得到多个目标时序特征；
    将所述多个目标时序特征拼接为目标一维向量，并将所述目标一维向量输入预设的全连接层，以通过所述预设的全连接层将所述目标一维向量转化为与所述采样时间点数数量相同的多个一维抽象特征，其中所述一维抽象特征的元素个数与所述传感器数量相同。
  4. 根据权利要求3所述的方法，其特征在于，所述基于多个维度相同的一维卷积核对所述二维平面向量进行一维卷积计算，得到多个一维原始特征，包括：
    针对各个传感器元素从所述二维平面向量中滑动提取多组预设采样时间点数的元素，并将所述多组预设采样时间点数的元素组成多个一维待卷积向量；
    将各个一维待卷积向量依次与当前一维卷积核进行卷积运算，生成当前一维原始特征；
    基于所述将各个一维待卷积向量依次与当前一维卷积核进行卷积运算，生成当前一维原始特征的步骤，遍历全部一维卷积核，直至得到各个一维卷积核对应的一维原始特征为止。
  5. 根据权利要求3所述的方法，其特征在于，基于所述第二神经网络输出故障诊断结果的步骤，包括：
    分别将多个一维抽象特征进行维度扩充；
    基于线性感知机对扩充后的所述多个一维抽象特征进行层次注意力处理，并融合所述多个一维抽象特征得到输出向量；
    将所述输出向量输入预设的分类器，以使所述预设的分类器输出所述故障诊断结果。
  6. 根据权利要求5所述的方法，其特征在于，基于线性感知机对所述多个一维抽象特征进行层次注意力处理，并融合所述多个一维抽象特征得到输出向量，包括：
    将所述多个一维抽象特征依次输入所述线性感知机，得到多个隐藏特征；
    通过softmax函数计算所述多个隐藏特征对应的多个权重系数，所述权重系数用于表征所述隐藏特征的层次注意力；
    利用所述多个权重系数对所述多个隐藏特征进行加权求和，得到所述输出向量。
  7. 一种齿轮箱故障诊断装置，其特征在于，所述装置包括：
    信号采集模块，用于获取齿轮箱的真实信号和如权利要求1所述方法生成的模拟信号，所述真实信号包括真实扭矩信号、真实振动信号和真实旋转信号；
    故障诊断模块，用于将所述模拟信号和所述真实信号分别输入预设的故障诊断网络，得到所述齿轮箱的故障诊断结果；
    其中，所述预设的故障诊断网络包括用于特征提取的第一神经网络和带有层次注意力机制的用于输出故障诊断结果的第二神经网络，所述预设的故障诊断网络基于所述模拟信号和所述真实信号训练生成。
  8. 一种电子设备，其特征在于，包括：
    存储器和处理器，所述存储器和所述处理器之间互相通信连接，所述存储器中存储有计算机指令，所述处理器通过执行所述计算机指令，从而执行如权利要求1-6任一项所述的方法。
  9. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机指令，所述计算机指令用于使所述计算机从而执行如权利要求1-6任一项所述的方法。
PCT/CN2022/112143 2022-03-14 2022-08-12 一种齿轮箱故障诊断、信号采集方法、装置和电子设备 WO2023020388A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
DE112022000106.2T DE112022000106T5 (de) 2022-03-14 2022-08-12 Verfahren zur Getriebefehlerdiagnose und Signalerfassung, eine Vorrichtung und ein elektronisches Gerät

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210244235.XA CN114323644B (zh) 2022-03-14 2022-03-14 一种齿轮箱故障诊断、信号采集方法、装置和电子设备
CN202210244235.X 2022-03-14

Publications (1)

Publication Number Publication Date
WO2023020388A1 true WO2023020388A1 (zh) 2023-02-23

Family

ID=81033235

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/112143 WO2023020388A1 (zh) 2022-03-14 2022-08-12 一种齿轮箱故障诊断、信号采集方法、装置和电子设备

Country Status (3)

Country Link
CN (1) CN114323644B (zh)
DE (1) DE112022000106T5 (zh)
WO (1) WO2023020388A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115839846A (zh) * 2023-02-27 2023-03-24 济南嘉宏科技有限责任公司 一种基于无线传感器的设备故障预警诊断方法
CN116150676A (zh) * 2023-04-19 2023-05-23 山东能源数智云科技有限公司 基于人工智能的设备故障诊断与识别方法及装置
CN116341396A (zh) * 2023-05-30 2023-06-27 青岛理工大学 一种基于多源数据融合的复杂装备数字孪生建模方法
CN116428129A (zh) * 2023-06-13 2023-07-14 山东大学 基于注意力混合神经网络的风机叶片冲击定位方法及系统
CN116449717A (zh) * 2023-06-20 2023-07-18 中机生产力促进中心有限公司 一种基于数字孪生的挤压机减速箱状态监控系统
CN116551466A (zh) * 2023-05-24 2023-08-08 深圳市捷辉创科技有限公司 Cnc加工过程中的智能监控系统及其方法
CN116894190A (zh) * 2023-09-11 2023-10-17 江西南昌济生制药有限责任公司 轴承故障诊断方法、装置、电子设备和存储介质
CN116930741A (zh) * 2023-07-19 2023-10-24 中国电子产品可靠性与环境试验研究所（（工业和信息化部电子第五研究所）（中国赛宝实验室）） 开关器件故障程度诊断方法、装置和计算机设备
CN116957361A (zh) * 2023-07-27 2023-10-27 中国舰船研究设计中心 一种基于虚实结合的船舶任务系统健康状态检测方法
CN117006002A (zh) * 2023-09-27 2023-11-07 广东海洋大学 基于数字孪生的海上风力发电机监测方法及系统
CN117249996A (zh) * 2023-11-10 2023-12-19 太原理工大学 矿用刮板机齿轮箱轴承的故障诊断方法
CN118152894A (zh) * 2024-05-09 2024-06-07 国网浙江省电力有限公司经济技术研究院 一种电力系统宽频振荡识别方法、装置、终端及介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114323644B (zh) * 2022-03-14 2022-06-03 中国长江三峡集团有限公司 一种齿轮箱故障诊断、信号采集方法、装置和电子设备
CN114526910A (zh) * 2022-04-21 2022-05-24 杭州杰牌传动科技有限公司 一种基于数字孪生驱动的传动系统故障定位方法

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110442936A (zh) * 2019-07-24 2019-11-12 中国石油大学(北京) 基于数字孪生模型的设备故障诊断方法、装置及系统
AU2020102863A4 (en) * 2020-10-19 2020-12-17 Beihang University Digital-twin-driven fault prognosis method and system for subsea production system of offshore oil
CN112100874A (zh) * 2020-07-24 2020-12-18 西安交通大学 基于数字孪生的转子叶片健康监测方法和监测系统
CN112765748A (zh) * 2021-01-25 2021-05-07 长安大学 机理-数据异构信息融合的旋转机械数字孪生建模方法
CN113236491A (zh) * 2021-05-27 2021-08-10 华北电力大学 一种风力发电数字孪生系统
CN113379196A (zh) * 2021-05-17 2021-09-10 国网浙江省电力有限公司宁波供电公司 一种基于数字孪生技术的变电设备管理评价系统
CN113505655A (zh) * 2021-06-17 2021-10-15 电子科技大学 面向数字孪生系统的轴承故障智能诊断方法
KR20220011461A (ko) * 2020-07-21 2022-01-28 한국전력공사 디지털 트윈 가상 모델을 이용한 발전 설비 고장 진단 장치 및 방법
CN114323644A (zh) * 2022-03-14 2022-04-12 中国长江三峡集团有限公司 一种齿轮箱故障诊断、信号采集方法、装置和电子设备

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101823746B1 (ko) * 2016-02-05 2018-01-30 울산대학교 산학협력단 베어링 고장 진단 방법
US20190005826A1 (en) * 2017-06-28 2019-01-03 Ge Aviation Systems, Llc Engine load model systems and methods
CN108896296A (zh) * 2018-04-18 2018-11-27 北京信息科技大学 一种基于卷积神经网络的风电齿轮箱故障诊断方法
CN109932174A (zh) * 2018-12-28 2019-06-25 南京信息工程大学 一种基于多任务深度学习的齿轮箱故障诊断方法
CN112417742B (zh) * 2021-01-22 2021-04-23 浙江中自庆安新能源技术有限公司 一种基于数字孪生模型的齿轮箱寿命动态评估方法及系统
CN112861787A (zh) * 2021-03-09 2021-05-28 上海电力大学 一种风电机组行星齿轮箱故障诊断方法
CN113742855B (zh) * 2021-07-27 2022-03-18 清华大学 一种故障预测方法、系统、电子设备与可读存储介质

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110442936A (zh) * 2019-07-24 2019-11-12 中国石油大学(北京) 基于数字孪生模型的设备故障诊断方法、装置及系统
KR20220011461A (ko) * 2020-07-21 2022-01-28 한국전력공사 디지털 트윈 가상 모델을 이용한 발전 설비 고장 진단 장치 및 방법
CN112100874A (zh) * 2020-07-24 2020-12-18 西安交通大学 基于数字孪生的转子叶片健康监测方法和监测系统
AU2020102863A4 (en) * 2020-10-19 2020-12-17 Beihang University Digital-twin-driven fault prognosis method and system for subsea production system of offshore oil
CN112765748A (zh) * 2021-01-25 2021-05-07 长安大学 机理-数据异构信息融合的旋转机械数字孪生建模方法
CN113379196A (zh) * 2021-05-17 2021-09-10 国网浙江省电力有限公司宁波供电公司 一种基于数字孪生技术的变电设备管理评价系统
CN113236491A (zh) * 2021-05-27 2021-08-10 华北电力大学 一种风力发电数字孪生系统
CN113505655A (zh) * 2021-06-17 2021-10-15 电子科技大学 面向数字孪生系统的轴承故障智能诊断方法
CN114323644A (zh) * 2022-03-14 2022-04-12 中国长江三峡集团有限公司 一种齿轮箱故障诊断、信号采集方法、装置和电子设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
PANG YU: "Simulation Study on Dynamics and Fault Diagnosis of Digital Twin in Planetary Gearbox", CHINA MASTER’S THESES FULL-TEXT DATABASE, no. 9, 15 September 2022 (2022-09-15), XP093036510 *

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115839846B (zh) * 2023-02-27 2023-06-20 济南嘉宏科技有限责任公司 一种基于无线传感器的设备故障预警诊断方法
CN115839846A (zh) * 2023-02-27 2023-03-24 济南嘉宏科技有限责任公司 一种基于无线传感器的设备故障预警诊断方法
CN116150676B (zh) * 2023-04-19 2023-09-26 山东能源数智云科技有限公司 基于人工智能的设备故障诊断与识别方法及装置
CN116150676A (zh) * 2023-04-19 2023-05-23 山东能源数智云科技有限公司 基于人工智能的设备故障诊断与识别方法及装置
CN116551466A (zh) * 2023-05-24 2023-08-08 深圳市捷辉创科技有限公司 Cnc加工过程中的智能监控系统及其方法
CN116551466B (zh) * 2023-05-24 2024-05-14 深圳市捷辉创科技有限公司 Cnc加工过程中的智能监控系统及其方法
CN116341396A (zh) * 2023-05-30 2023-06-27 青岛理工大学 一种基于多源数据融合的复杂装备数字孪生建模方法
CN116341396B (zh) * 2023-05-30 2023-08-11 青岛理工大学 一种基于多源数据融合的复杂装备数字孪生建模方法
CN116428129A (zh) * 2023-06-13 2023-07-14 山东大学 基于注意力混合神经网络的风机叶片冲击定位方法及系统
CN116428129B (zh) * 2023-06-13 2023-09-01 山东大学 基于注意力混合神经网络的风机叶片冲击定位方法及系统
CN116449717A (zh) * 2023-06-20 2023-07-18 中机生产力促进中心有限公司 一种基于数字孪生的挤压机减速箱状态监控系统
CN116449717B (zh) * 2023-06-20 2023-09-22 中机生产力促进中心有限公司 一种基于数字孪生的挤压机减速箱状态监控系统
CN116930741A (zh) * 2023-07-19 2023-10-24 中国电子产品可靠性与环境试验研究所（（工业和信息化部电子第五研究所）（中国赛宝实验室）） 开关器件故障程度诊断方法、装置和计算机设备
CN116957361A (zh) * 2023-07-27 2023-10-27 中国舰船研究设计中心 一种基于虚实结合的船舶任务系统健康状态检测方法
CN116957361B (zh) * 2023-07-27 2024-02-06 中国舰船研究设计中心 一种基于虚实结合的船舶任务系统健康状态检测方法
CN116894190A (zh) * 2023-09-11 2023-10-17 江西南昌济生制药有限责任公司 轴承故障诊断方法、装置、电子设备和存储介质
CN116894190B (zh) * 2023-09-11 2023-11-28 江西南昌济生制药有限责任公司 轴承故障诊断方法、装置、电子设备和存储介质
CN117006002A (zh) * 2023-09-27 2023-11-07 广东海洋大学 基于数字孪生的海上风力发电机监测方法及系统
CN117006002B (zh) * 2023-09-27 2024-02-09 广东海洋大学 基于数字孪生的海上风力发电机监测方法及系统
CN117249996A (zh) * 2023-11-10 2023-12-19 太原理工大学 矿用刮板机齿轮箱轴承的故障诊断方法
CN117249996B (zh) * 2023-11-10 2024-02-13 太原理工大学 矿用刮板机齿轮箱轴承的故障诊断方法
CN118152894A (zh) * 2024-05-09 2024-06-07 国网浙江省电力有限公司经济技术研究院 一种电力系统宽频振荡识别方法、装置、终端及介质

Also Published As

Publication number Publication date
DE112022000106T5 (de) 2023-05-25
CN114323644B (zh) 2022-06-03
CN114323644A (zh) 2022-04-12

Similar Documents

Publication Publication Date Title
WO2023020388A1 (zh) 一种齿轮箱故障诊断、信号采集方法、装置和电子设备
Li et al. Intelligent fault diagnosis of rolling bearings under imbalanced data conditions using attention-based deep learning method
Pan et al. A hybrid DBN-SOM-PF-based prognostic approach of remaining useful life for wind turbine gearbox
CN104792522B (zh) 一种基于分数阶小波变换和bp神经网络的齿轮缺陷智能分析方法
WO2020176295A1 (en) Artificial neural network compression via iterative hybrid reinforcement learning approach
CN110543860B (zh) 基于tjm迁移学习的机械故障诊断方法及系统
Gai et al. A Parameter‐Optimized DBN Using GOA and Its Application in Fault Diagnosis of Gearbox
Xu et al. RUL prediction for rolling bearings based on Convolutional Autoencoder and status degradation model
CN109543905A (zh) 基于改进的二维cnn模型的滚动轴承剩余寿命预测方法
CN113743016B (zh) 基于自编码器和回声状态网络的发动机剩余寿命预测方法
CN111784061B (zh) 一种电网工程造价预测模型的训练方法、装置和设备
CN111340282A (zh) 基于da-tcn的设备剩余使用寿命的估计方法及系统
Huang et al. Deep residual networks-based intelligent fault diagnosis method of planetary gearboxes in cloud environments
CN116316591A (zh) 基于混合双向门控循环的短期光伏功率预测方法及系统
Sun et al. An adaptive anti-noise gear fault diagnosis method based on attention residual prototypical network under limited samples
CN116226646A (zh) 轴承健康状态及剩余寿命的预测方法、系统、设备及介质
Wu et al. Distillation-enhanced fast neural architecture search method for edge-side fault diagnosis of wind turbine gearboxes
Wang et al. Spatiotemporal non-negative projected convolutional network with bidirectional NMF and 3DCNN for remaining useful life estimation of bearings
CN115290326A (zh) 一种滚动轴承故障智能诊断方法
Du et al. A remaining useful life prediction method based on time–frequency images of the mechanical vibration signals
Ge et al. A deep condition feature learning approach for rotating machinery based on MMSDE and optimized SAEs
CN114295967A (zh) 一种基于迁移神经网络的模拟电路故障诊断方法
Zhang et al. Graph neural network-based bearing fault diagnosis using Granger causality test
CN112101482A (zh) 一种对有缺失卫星数据进行参数异常模式检测的方法
Handikherkar et al. Gear fault detection using machine learning techniques-a simulation-driven approach

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22857700

Country of ref document: EP

Kind code of ref document: A1