Time-frequency multi-scale multi-task fault diagnosis and residual life prediction combined method

Abstract

The application provides a time-frequency multi-scale multi-task fault diagnosis and residual life prediction combined method, comprising the following steps: reading a bearing vibration acceleration signal; performing short-time Fourier transform calculation on a time domain signal of the bearing vibration acceleration, and obtaining a vibration acceleration power spectrum density diagram; according to a root mean square of the bearing vibration acceleration signal, data labeling is performed on bearing full life cycle degradation data; the bearing full life cycle degradation data, a health state value and a normalized residual life value are taken as a complete data set; the training set is substituted into a gated multi-task multi-scale convolutional neural network; after network parameters are optimized, a convergence condition is judged; and a multi-scale multi-task learning model fault diagnosis and residual life prediction combined model is obtained. The method can more intuitively depict the bearing degradation process, capture different scale features, and dynamically adjust the relationship between different scale features.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical fault diagnosis and residual life prediction, in particular to a time-frequency multi-scale multi-task fault diagnosis and residual life prediction joint method. BACKGROUND

[0002] Bearings are widely used in industrial systems, and their reliability is one of the important elements of the safety of many mechanical and electrical systems. Due to the harsh working conditions, bearing failures often occur, which may cause serious consequences. Therefore, bearing fault diagnosis and residual life prediction are very important for improving the reliability of mechanical and electrical systems.

[0003] Fault diagnosis is a process of identifying the occurrence of a fault as early as possible, identifying the type and location of the fault, and the degree of the fault. Residual life prediction is a process of estimating the time left before a component fails. Most fault diagnosis includes two key environments, feature extraction and fault identification. Typical feature extraction methods include hidden Markov models, and in recent years, artificial intelligence-based methods such as support vector machines.

[0004] In practice, the safety of industrial systems depends not only on fault diagnosis, because it is too late once a fault occurs. Therefore, residual life estimation is also very important, which can be used to determine whether a component needs to be repaired or replaced. There are three main methods of residual life estimation, including model-based methods, data-based methods, and hybrid methods. The main problem of the model method is the uncertainty of the prior knowledge in the model and the problem of parameter estimation caused by simplification. Therefore, data-driven methods can be used to estimate these parameters or ignore the simplified model to directly obtain an input-output relationship.

[0005] There are also existing technologies that integrate fault diagnosis and residual life estimation into a model, but existing multi-task models lack the capture of different scale features and lack a means to dynamically adjust the relationship between different scales. Moreover, simply using time-domain representation cannot well depict the bearing degradation process. SUMMARY

[0006] The main purpose of the present application is to overcome the above-mentioned defects in the prior art, and to propose a time-frequency multi-scale multi-task fault diagnosis and residual life prediction joint method, which can more intuitively depict the bearing degradation process, capture different scale features, and dynamically adjust the relationship between different scale features.

[0007] The present application adopts the following technical solutions:

[0008] The time-frequency multi-scale multi-task fault diagnosis and residual life prediction joint method comprises the following steps:

[0009] 1) Read the vibration acceleration signal of the bearing, calculate the time domain signal of the bearing vibration acceleration by short-time Fourier transform, and the frequency domain signal of the vibration acceleration, wherein the calculation formula of the short-time Fourier transform is:

[0010]

[0011] Wherein is the length of the window function, represents the time domain discrete sequence of bearing vibration acceleration, the frequency domain signal of vibration acceleration obtained by short-time Fourier transform, is the window function, is the sequence point, is the imaginary unit, and e is the natural constant;

[0012] 2) According to the frequency domain signal of the vibration acceleration, the power spectrum density diagram of the vibration acceleration is calculated, wherein the calculation formula of the power spectrum density is:

[0013]

[0014] 3) Preprocess the degradation data of the bearing throughout the life cycle, set several threshold values according to the root mean square, determine the health state and residual life value according to the threshold value, and complete the labeling of the degradation data;

[0015] 4) The degradation data of the bearing throughout the life cycle converted into the power spectrum density diagram in step 2), and the health state value and the normalized residual life value are taken as a complete data set; The data set is randomly divided into training set and test set two parts, wherein the training set accounts for 70% of the complete data set, and the test set accounts for 30% of the complete data set;

[0016] 5) Substitute the training set into the gated multi-task multi-scale convolutional neural network, and use the Adam optimization algorithm to optimize the network parameters to obtain new network parameters; wherein the basic formula of forward calculation of the neural network is:

[0017]

[0018]

[0019] Wherein is the intermediate quantity of convolution calculation, and the size of the convolution kernel is , is the input of the neural network, is the neural network kernel, and selecting several different size neural network kernels can obtain different scale information of the input extracted by the multi-scale network, is the output of the neural network;

[0020] ​The different scale weight allocation methods used in the gating network are as follows:

[0021]

[0022] in It is the second to last layer of the gated network. One output value, represent The value can be any one of the outputs of the second-to-last layer of the gating network. Let i be the i-th weight output by the gating network. This is the i-th output value obtained from the last layer of the gated network;

[0023] Gated networks are added to handle multi-task scenarios. The process of combining the output weights of the gated network with the multi-scale network follows the formula:

[0024]

[0025] in Represents the number of branches in a multi-scale network. The eigenvalues ​​represent the weighted summation of multi-scale network features by the gated network. Let i be the i-th output value in the multi-scale network. The weight value assigned to the i-th output value in the gating network;

[0026] 6) Convergence condition judgment, and joint model of fault diagnosis and remaining lifetime prediction of multi-scale multi-task learning model.

[0027] Specifically, in step 5), the method for optimizing the neural network model parameters is as follows: minimize the objective function described below:

[0028]

[0029] in For the parameters of the neural network, It involves adjusting the weights of the two tasks: fault diagnosis and remaining life prediction.

[0030] Specifically,

[0031]

[0032] in This represents the model's prediction of the remaining lifetime. This is the normalized remaining lifetime value calibrated in step three, and this item represents the total loss of the remaining lifetime task.

[0033]

[0034] in This is the predicted probability value of the model for the category of the true health status value of the i-th sample, as labeled in step three, in a classification task. This term represents the probability prediction of the model for each health status value of the i-th sample in step three, and is the total loss of the fault diagnosis task.

[0035] Specifically, in step 6), the method for determining the convergence condition can be as follows: If The algorithm converges when the previously set maximum value of the optimization cycle for the algorithm parameters is reached, and the current prediction model is output; otherwise, the model parameters continue to be optimized.

[0036] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0037] This invention provides a joint method for time-frequency multi-scale, multi-task fault diagnosis and remaining life prediction. Utilizing time-frequency domain representation, it better represents the characteristics of bearing vibration signals, treating fault diagnosis and remaining life prediction as a unified model in multi-task learning, making it easier to implement in industrial measurement systems. Furthermore, it extracts feature information at different scales from the time-frequency domain representation of bearing vibration signals, enabling the system to better perform different tasks. Simultaneously, this invention utilizes a gating network to assign different weights to the differences in features at different scales, thereby better adapting to the variations in features at different scales. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the present invention.

[0039] Figure 2 Figure (a) shows the label allocation diagram of the dataset in the embodiment of the present invention, and Figure (b) shows the label allocation diagram of another dataset.

[0040] Figure 3 This is a schematic diagram of a gated multi-scale multi-task neural network used in an embodiment of the present invention.

[0041] Figure 4 The remaining lifetime prediction results are shown after the implementation of the present invention. Figure (a) shows the method of the present invention, and Figure (b) shows the GRU network.

[0042] Figure 5 The results are the fault diagnosis results after the implementation of the present invention. Figure (a) shows the present invention, and Figure (b) shows the GRU network.

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0044] The present invention will be further described below through specific embodiments.

[0045] 1) Read the bearing vibration acceleration signal and perform a short-time Fourier transform on the time-domain signal of the bearing vibration acceleration. The formula for calculating the short-time Fourier transform is as follows:

[0046] (1)

[0047] in It is the length of the window function. Represents the time-domain discrete sequence of bearing vibration acceleration. It is a window function.

[0048] The window function can be expressed by the following formula.

[0049] (2)

[0050] Where n is the length of the vibration signal, and N is the length of the window function, which is selected manually;

[0051] 2) Based on the frequency domain signal of the vibration acceleration obtained from the short-time Fourier transform, the power spectral density of the vibration acceleration is calculated. The formula for calculating the power spectral density is:

[0052] (3);

[0053] 3) Preprocess the degradation data of the bearing throughout its entire life cycle, set several thresholds based on the root mean square, determine the health status and remaining life value according to the thresholds, and complete the labeling of the degradation data;

[0054] 4) The bearing lifecycle degradation data converted to time-frequency representation in step two, along with the corresponding health status values ​​and normalized remaining life values ​​obtained in step three, are combined into a complete dataset. This dataset is randomly divided into a training set and a test set, with the training set comprising 70% of the complete dataset and the test set comprising 30%.

[0055] 5) Substitute the training set from step four into the inductively controlled multi-task multi-scale convolutional neural network, and use the Adam optimization algorithm to optimize the network parameters to obtain new network parameters. The basic formula for the forward computation of the neural network is:

[0056] (4)

[0057] (5)

[0058] in It is an intermediate quantity in the convolution calculation, and the size of the convolution kernel is... , As input to the neural network, By selecting several neural network kernels of different sizes, a multi-scale network can be obtained to extract information from the input at different scales. For the output of the neural network, It is a nonlinear function.

[0059] Nonlinear functions can be:

[0060] (6)

[0061] The different scale weight allocation methods used in the gating network are as follows:

[0062] (7)

[0063] in It is the second to last layer of the gated network. One output value, represent The value can be any one of the outputs of the second-to-last layer of the gating network. Let i be the i-th weight output by the gating network. This is the i-th output value obtained from the last layer of the gated network.

[0064] Gated networks can be added to accommodate multi-task scenarios. The process of combining the output weights of a gated network with that of a multi-scale network is given by the following formula:

[0065] (8)

[0066] in Represents the number of branches in a multi-scale network. The eigenvalues ​​represent the weighted summation of multi-scale network features by the gated network. Let i be the i-th output value in the multi-scale network. The weight value assigned to the i-th output value in the gating network.

[0067] The parameters of a neural network can be optimized by minimizing the objective function described below:

[0068] (9)

[0069] in For the parameters of the neural network, This is used to adjust the weights of the two tasks: fault diagnosis and remaining life prediction. This value is selected manually.

[0070] Specifically,

[0071] (10)

[0072] in the predicted value of the remaining useful life by the model, the normalized remaining useful life value calibrated in step three, which is the total loss of the remaining useful life task.

[0073] (11)

[0074] where is the predicted probability value of the class of the true health state value calibrated in step three by the model for the i-th sample in the classification task, is the probability predicted value of each health state value in step three for the i-th sample by the model, which is the total loss of the fault diagnosis task.

[0075] For the optimization process, we first calculate the first-order gradient of the loss function:

[0076] (12)

[0077] where is the objective function, is the parameter of the neural network.

[0078] Then calculate the first-order momentum and the second-order momentum:

[0079] (13)

[0080] (14)

[0081] where and are artificially selected parameters, and the initial parameters are set to and , represent element-wise multiplication.

[0082] Then the corrected first-order momentum and the second-order momentum are calculated:

[0083] (15)

[0084] (16)

[0085] where and are and raised to the power of .

[0086] Finally, the parameters in the neural network are optimized by the following formula:

[0087] (17)

[0088] is a number slightly larger than 0, to avoid the calculation problem when the second order momentum is 0. is a number slightly larger than 0, to avoid the calculation problem when the second order momentum is 0.

[0089] 6) Convergence condition judgment, get the multi-scale multi-task learning model of fault diagnosis and residual life prediction joint model.

[0090] In step 6), the specific method of the convergence condition judgment can be: if , the algorithm converges, and the current prediction model is output, otherwise the model parameters are continuously optimized, or the maximum value of the previously set algorithm parameter optimization period is reached, and the current prediction model is output.

[0091] Now combined with the PRONOSITIA data set embodiment is described:

[0092] 1) Read the bearing vibration acceleration signal, and perform short-time Fourier transform calculation on the time domain signal of the bearing vibration acceleration, wherein the calculation formula of the short-time Fourier transform is:

[0093] (1)

[0094] wherein represents the time domain discrete sequence of bearing vibration acceleration, the length is 2560, is a window function.

[0095] The window function can be expressed by the following formula

[0096] (2)

[0097] wherein n is the length of the vibration signal, and N is the length of the window function, which is artificially selected as 256;

[0098] 2) According to the frequency domain signal of the vibration acceleration obtained by the short-time Fourier transform, the power spectrum density diagram of the vibration acceleration is calculated, wherein the calculation formula of the power spectrum density is:

[0099] (3);

[0100] 3) The degradation data of the bearing full life cycle is preprocessed, and the bearing full life cycle degradation data is labeled according to the root mean square of the bearing vibration acceleration signal, that is, the health state value and the normalized residual life value of each data, as shown in Figure 2 , specifically, the bearing full life cycle degradation data is divided into three categories and labeled with residual life;

[0101] 4) The bearing full life cycle degradation data converted into time-frequency representation in step two, and its corresponding health state value and normalized remaining useful life value obtained in step three, are taken as a complete data set. The data set is randomly divided into a training set and a test set, wherein the training set accounts for 70% of the complete data set, and the test set accounts for 30% of the complete data set.

[0102] 5) The training set in step four is substituted into the gated multi-task multi-scale convolutional neural network, the general framework of which is shown in Figure 3 After the network parameters are optimized by using the Adam optimization algorithm, new network parameters are obtained. The basic formula for forward calculation of the neural network is as follows:

[0103] (4)

[0104] (5)

[0105] wherein x is the input of the neural network, is the neural network kernel, and a plurality of neural network kernels of different sizes are selected to obtain information of different scales extracted from the input by the multi-scale network, is the output of the neural network, is a nonlinear function. The nonlinear function can be:

[0106] (6) The different scale weight distribution mode adopted in the gated network is as follows:

[0107]

[0108] (7) wherein wi is the i-th weight of the output of the gated network,

[0109] is the i-th output value obtained by the last layer in the gated network. The gated network can be increased according to the multi-task condition, and the process of integrating the output weight of the gated network and the multi-scale network is as follows:

[0110] (8)

[0111] wherein xi is the i-th output value in the multi-scale network, is the weight value allocated to the i-th output value in the gated network. The multi-task part is realized by two networks.

[0112] The optimization of the neural network parameters can be obtained by minimizing the following objective function:

[0113]

[0114] ​​​(9)

[0115] where are the parameters of the neural network, is the weight that regulates the two tasks of fault diagnosis and remaining useful life prediction, which is manually selected with a step of 0.05, and there are 21 values in total. Specifically,

[0116] (10)

[0117] where is the predicted value of the remaining useful life by the model, is the normalized remaining useful life value calibrated in step three, which is the total loss of the remaining useful life task.

[0118] (11)

[0119] where is the predicted probability value of the class of the true health state value of the model for the i-th sample in step three, is the probability prediction value of each health state value of the model for the i-th sample in step three, which is the total loss of the fault diagnosis task.

[0120] For the optimization process, we first calculate the first-order gradient of the loss function:

[0121] (12)

[0122] where is the objective function, are the parameters of the neural network.

[0123] Then calculate the first-order momentum and the second-order momentum:

[0124] (13)

[0125] (14)

[0126] where and are artificially selected parameters, which are 0.9 and 0.999 respectively, and the initial parameters are set to and , represent element-wise multiplication.

[0127] Then the corrected first-order momentum and the second-order momentum are calculated:

[0128] (15)

[0129] (16)

[0130] wherein and are and are powers.

[0131] Finally, the parameters in the neural network are optimized by the following formula:

[0132] (17)

[0133] is a learning rate, is set to 0.001, is a number slightly larger than 0, to avoid the calculation problem when the second order momentum is 0.

[0134] 6) Convergence condition judgment, get the multi-scale multi-task learning model of the fault diagnosis and residual life prediction joint model.

[0135] In step 6), the specific method of the convergence condition judgment can be: if , the algorithm converges, and the current prediction model is output, otherwise, go to step 5 to continue optimizing the model parameters, or go to the maximum value 20 of the previously set algorithm parameter optimization period, and output the current prediction model.

[0136] The result graphs of residual life and fault diagnosis are shown in Figures 4-5 It can be seen from Figure 4 and Figure 5 that the method of the present application is better than the classic gated recurrent unit (GRU), the results of the present application are closer to the true value, and the comprehensive accuracy on the two tasks of residual life prediction and fault diagnosis is better.

[0137] In summary, the present application is completely applicable to the field of mechanical fault diagnosis and life prediction, and has the advantages of utilizing, etc. The present application can provide more accurate data support for advanced applications such as mechanical equipment maintenance and mechanical equipment motion control, and has a good application prospect

[0138] The above is only a specific embodiment of the present application, but the design concept of the present application is not limited thereto, and any non-essential modification of the present application using this concept shall be regarded as an act of infringing the protection scope of the present application.

Claims

1. A time-frequency multi-scale multi-task fault diagnosis and residual life prediction combined method based on, characterized in that, Comprising the following steps: 1) Read the vibration acceleration signal of the bearing, calculate the time domain signal of the bearing vibration acceleration, and the frequency domain signal of the vibration acceleration by short-time Fourier transform, wherein the calculation formula of the short-time Fourier transform is: wherein is the length of the window function, denotes a time domain discrete sequence of bearing vibration acceleration, is a frequency domain signal of the vibration acceleration obtained by performing a short-time Fourier transform, is the window function, is a sequence point, is the imaginary unit, and e is the natural constant; 2) According to the frequency domain signal of the vibration acceleration, the power spectrum density diagram of the vibration acceleration is calculated, wherein the calculation formula of the power spectrum density is: ； 3) Preprocess the degradation data of the bearing throughout the life cycle, set several threshold values according to the root mean square, and determine the health state and residual life value according to the threshold value to complete the labeling of the degradation data; 4) The degradation data of the bearing throughout the life cycle converted into the power spectrum density diagram in step 2), and the health state value and the normalized residual life value are taken as a complete data set; The data set is randomly divided into two parts of training set and test set, wherein the training set accounts for 70% of the complete data set, and the test set accounts for 30% of the complete data set; 5) Put the training set into the gated multi-task multi-scale convolutional neural network, and obtain new network parameters after optimizing the network parameters by using the Adam optimization algorithm; The basic formula for forward calculation of the neural network is: wherein is an intermediate quantity of convolution calculation, the size of the convolution kernel is , is the neural network input, is the neural network kernel, and selecting several neural network kernels of different sizes can obtain a multi-scale network to extract different scale information of the input, is the neural network output; The different scale weight distribution method used in the gated network is: wherein is the (i-1)th output value of the second-to-last layer of the gating network, is the ith output value of the last layer of the gating network, represents is any one of all the outputs of the second-to-last layer of the gating network, is the ith weight of the output of the gating network, is the ith output value obtained by the last layer of the gating network; The process of the gated network output weight and the multi-scale network integration according to the multi-task condition has the following formula: wherein represents the number of branches of the multi-scale network, represents the feature value of the multi-scale network after the gated network weights and sums the feature values of the multi-scale network, is the i-th output value in the multi-scale network, is the weight value assigned to the i-th output value in the gated network; 6) Convergence condition judgment, get the joint model of fault diagnosis and residual life prediction of multi-scale multi-task learning model.

2. The method of claim 1, wherein the method is characterized by, In step 5), the specific method for optimizing the parameters of the neural network model is to minimize the following objective function: wherein are parameters of the neural network, is a weight that regulates both the tasks of fault diagnosis and residual life prediction; Specifically, wherein is the predicted value of the remaining lifetime by the model, is the calibrated normalized remaining lifetime value in step three, is the total loss of the remaining lifetime task; wherein is the predicted probability value of the class of the true health state value that the model pair labels the model pair's prediction of the i-th sample in step three for the classification task, is the probability prediction value of each health state value of the i-th sample in step three for the model pair, is the total loss for the fault diagnosis task. 3.The method of claim 1, wherein, In step 6), the specific method of the convergence condition judgment can be: if the number of iterations reaches a preset maximum number of iterations or reaches a preset maximum value of the period of algorithm parameter optimization, the algorithm converges, and the current prediction model is output, otherwise the model parameters continue to be optimized. or reaches a preset maximum value of the period of algorithm parameter optimization, the algorithm converges, and the current prediction model is output, otherwise the model parameters continue to be optimized.

Images