Bearing fault diagnosis method of self-optimization deep mixed kernel extreme learning machine

Abstract

The invention provides a bearing fault diagnosis method of a self-optimization deep mixed kernel extreme learning machine, and belongs to the technical field of bearing fault diagnosis. The technical problems that the bearing fault diagnosis accuracy is low, and fault classification is prone to errors are solved. According to the technical scheme, the method comprises the following steps of S1, dividing a data set after fault information extraction into a training set and a test set; s2, setting the number of hidden layers and an activation function, and constructing a mixed kernel extreme learning machine model; s3, obtaining various parameters of the model by using an improved hawk algorithm; and S4, judging whether the bearing has a fault or not in the operation process and determining the fault type. The method has the advantages that the advantages of the improved optimization algorithm and the deep mixed kernel extreme learning machine are fused, the fault type of the bearing can be accurately judged, and the fault features of various faults can be effectively obtained for subsequent state judgment and fault type classification.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method of a self-optimizing deep hybrid kernel extreme learning machine. Background Art

[0002] Rotating machinery is widely used in aerospace, metallurgy, electric power, railway and other industrial fields. Bearings are very important parts in rotating machinery and are called "the joints of industry". Bearings are basic parts that play the role of motion transmission, friction reduction and force bearing, and are very common in rotating machinery. Studies have shown that once a bearing fails during operation, it is easy to cause safety hazards in the equipment and lead to accidents. Therefore, it is of great significance to obtain the failure data of rolling bearings and conduct research and analysis to determine the type and cause of the failure.

[0003] The existing bearing fault diagnosis models mainly include: bearing fault diagnosis models constructed using convolutional neural networks and extreme learning machines, etc. However, the bearing fault diagnosis model constructed using convolutional neural networks is very sensitive to the initial weight input data size, and its black box model has difficulty explaining the decision-making process of the model, and is prone to converge to local minima, resulting in low fault classification accuracy; another bearing fault diagnosis model constructed using extreme learning machines cannot determine the number of hidden layer nodes of the extreme learning machine, resulting in low construction efficiency of the bearing fault diagnosis model, and this method cannot quickly determine the target bearing state characteristics from the fault sample set, resulting in low fault classification accuracy for bearing fault diagnosis using this bearing fault diagnosis model. For example, in CN202210433009.6, a bearing fault diagnosis method based on the Gray Wolf Algorithm to optimize the extreme learning machine is used. Although the Gray Wolf Algorithm is used to optimize the extreme learning machine parameters, the Gray Wolf Algorithm is prone to stagnation at the local optimal solution and cannot continue to search for the global optimal solution. In addition, the convergence accuracy is not high and the convergence speed is not fast enough. CN202010760962.2 uses a brushless DC motor bearing fault diagnosis method based on the extreme learning machine model. This method does not process the weights and hidden layer thresholds of the input layer and hidden layer of the ELM. The weights and hidden layer thresholds of the input layer and hidden layer are randomly generated, which has certain limitations on the accuracy of fault diagnosis, and its generalization ability and robustness are also limited. Therefore, there are still problems in the diagnosis of bearing fault types, such as low fault type discrimination rate and low fault diagnosis accuracy. An efficient and accurate bearing fault diagnosis method is urgently needed. Summary of the invention

[0004] In view of the technical problems of low accuracy in bearing fault diagnosis and easy error in fault classification in the prior art, the purpose of the present invention is to provide a bearing fault diagnosis method of a self-optimizing deep hybrid kernel extreme learning machine. The method combines the advantages of the improved Sky Eagle optimization algorithm and the deep hybrid kernel extreme learning machine to accurately determine the fault type of the bearing and effectively obtain the fault characteristics of various faults for subsequent state judgment and fault type classification. The diagnostic method can be used to diagnose the health status of motor bearings, effectively investigate the type of fault, reduce economic losses, and resume production operations as soon as possible.

[0005] The inventive idea of ​​the present invention is: the present invention provides a bearing fault diagnosis method of a self-optimizing deep hybrid kernel extreme learning machine, uses empirical mode decomposition to process the collected vibration signal to obtain characteristic variables, sets the number of hidden layers and the activation function, and constructs a hybrid kernel extreme learning machine model based on an extreme learning machine and an autoencoder. In the model, the relevant parameters are optimized by the improved Sky Eagle algorithm and the training set is used for training to obtain a deep hybrid kernel extreme learning machine model. Finally, the test set is input into the trained model to determine whether the bearing has a fault during operation and determine the type of fault.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: comprising the following steps:

[0007] Step S1, perform empirical mode decomposition on the collected bearing vibration signal to obtain the inherent mode components that conform to the inherent mode principle, input the decomposed signal into the Hilbert transform model, perform envelope mediation analysis to extract fault information, and divide the data set after extracting the fault information into a training set and a test set.

[0008] In step S2, the number of hidden layers and activation functions are set. A multi-layer extreme learning machine (ML-ELM) is created by stacking an extreme learning machine (ELM)-based autoencoder. A hybrid kernel extreme learning machine model (HKELM) is established by constructing a hybrid kernel extreme learning machine as an ML-ELM classifier combined with a kernel method.

[0009] Step S3, using the Improved Aquila Optimizer (IAO) to obtain various parameters, including the number of hidden layer nodes, kernel function type and weight, etc., to modify the hybrid kernel extreme learning machine model.

[0010] Step S4, putting the obtained parameters into the hybrid kernel extreme learning machine model and loading the training set data for training to obtain a deep hybrid kernel extreme learning machine (DHKELM), calling the model to perform classification prediction on the test set to determine whether the bearing has a fault during operation and determine the fault type.

[0011] Furthermore, in step S1, the collected bearing vibration signal is subjected to empirical mode decomposition (EMD) to obtain intrinsic mode components that comply with the intrinsic mode principle, until the original signal cannot be decomposed or a given termination condition is met.

[0012] The collected bearing vibration signal is subjected to empirical mode decomposition to obtain the average value m1(t) of the upper and lower envelopes, which is expressed as follows:

[0013]

[0014] In the formula, e max (t) and e min (t) are the upper envelope value and the lower envelope value, respectively.

[0015] Screen the original signal x(t) to get a new signal C 1,1 (t) is called the intermediate signal, and the expression of this signal is:

[0016] x(t)-m1(t)=C 1,1 (t) (2)

[0017] Determine the intermediate signal C 1,1 (t) Whether the two conditions of the Intrinsic Mode Function (IMF) are met, namely:

[0018] (1) The number of local extreme points and zero-crossing points of the function must be equal or differ by at most one in the entire time range;

[0019] (2) At any point in time, the average of the envelope of the local maximum (upper envelope) and the envelope of the local minimum (lower envelope) must be zero.

[0020] If the intermediate signal C 1,1 (t) satisfies these two conditions, then the signal is an IMF component; if not, based on the signal, repeat the analysis of (1) and (2), that is, continue to screen until the signal obtained after k decompositions satisfies the IMF condition and obtains the first IMF component I1(t) of the original signal. The expression of this component is:

[0021] I1(t)=C 1,k-1 (t)-m k (t) (3)

[0022] At this time, I1(t) represents the highest frequency IMF component in the original signal x(t), and the expression of the residual component r1(t) is:

[0023] r1(t)=x(t)-I1(t) (4)

[0024] The acquisition of IMF components usually requires several iterations.

[0025] Performing the above screening process on r1(t) can obtain the second IMF component I2(t), and then subtracting I2(t) from r1(t) to obtain the residual component r2(t). This decomposition continues until the last residual signal r n (t) until it cannot be further decomposed, the empirical mode decomposition process of the signal x(t) is completed. The expression of this process is as follows:

[0026] r1(t)-I2(t)=r2(t)

[0027] … (5)

[0028] r n-1 (t)-I n (t) = r n (t)

[0029] When n (t) is a monotonic function, the remaining r n (t) is called the residual component, and the sum of all IMF components and the residual component is the original signal x(t), that is:

[0030]

[0031] After obtaining a series of IMF components, the corresponding instantaneous frequency and instantaneous amplitude of each IMF component can also be calculated. n (t) is transformed by Hilbert, and we can get

[0032]

[0033] Among them, I i (τ) is the value of the input signal at time τ, t is the time at which the Hilbert transform is to be calculated, and π is a constant.

[0034] Then, the analytical signal z can be constructed i (t), its expression is:

[0035]

[0036] Among them, the amplitude function a i (t) is expressed as:

[0037]

[0038] Phase function φ i (t) is expressed as:

[0039]

[0040] Instantaneous frequency f i (t) is expressed as

[0041]

[0042] This gives the Hilbert spectrum, denoted as Its expression is:

[0043]

[0044] In the formula, RP represents the real part. The residual component r is omitted here. n (t). The fault feature information is obtained by extracting it.

[0045] Further, in step S2, the number of hidden layers and activation functions are set, an ML-ELM is created by stacking extreme learning machine-based autoencoders, and a hybrid kernel extreme learning machine model is established by constructing a hybrid kernel extreme learning machine as an ML-ELM classifier combined with a kernel method.

[0046] The learning process of the extreme learning machine can be expressed by the least squares method, and its expression is:

[0047] β=H + T (13)

[0048] Where H + is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H, T is the target output matrix, and β is the output weight matrix of the hidden layer.

[0049] The kernel method has a strong nonlinear mapping capability. In order to make the extreme learning machine obtain stronger learning and generalization capabilities, different types of kernel functions are fused. Introducing the kernel parameter HH T , in order to improve the generalization ability of the ELM model, the output function of the kernel extreme learning machine (KELM) can be expressed as

[0050]

[0051] Where: the positive constant C is the penalty parameter; I is the identity matrix.

[0052] The selection of kernel functions and their related parameters largely determines the performance of kernel extreme learning machines, so it is of great significance to find a suitable kernel function for kernel extreme learning machines. In addition, a single kernel function may not be universal, so two kernel functions can be used as basis functions to establish a new hybrid kernel function through multi-kernel learning methods such as linear or nonlinear combination. The kernel extreme learning machine model with polynomial kernel function has good generalization performance, but its learning ability is slightly weaker. The wavelet kernel function constructed based on the wavelet equation can significantly improve the approximation efficiency of the kernel extreme learning machine and has strong classification performance. The expressions of Gaussian radial kernel function, polynomial and wavelet kernel function are shown below:

[0053]

[0054] K poly (x,x′)=((x i ·x i ′)+m) n (16)

[0055]

[0056] Where x represents the normalized training set sample, x T represents the transpose of the normalized test set samples, δ is the bandwidth used to control the local range of the Gaussian kernel function, m and n are the kernel parameters of the polynomial kernel function, and a, c, and d are the kernel parameters of the Morlet wavelet kernel function.

[0057] Therefore, in order to improve the versatility and approximation ability of the kernel extreme learning machine, this paper combines the above two functions to form a new hybrid function as the kernel function of the kernel extreme learning machine. A new hybrid kernel learning machine model is established, and its calculation method is as follows:

[0058]

[0059] Among them, ω is the weight coefficient, ω∈[0,1].

[0060] Thanks to the good capabilities of the hybrid kernel function, the resulting hybrid kernel learning machine model has stronger learning and generalization capabilities. However, since it is still a shallow network architecture, in some practical work, the application effect may not matter at all. Here, an ML-ELM is created by stacking extreme learning machine-based autoencoders (ELM-AEs). With the help of the multi-hidden layer architecture of ML-ELM, they achieved better performance while still having the fast learning rate inherited from ELM. Therefore, a deep neural network can be established by stacking multiple ELM-AEs. The output weight β of the i-th ELM-AE is iIt is regarded as the weight of the 4th layer of ML-ELM, and multiple repeatedly stacked ELM-AEs are used to construct a deep extreme learning machine (DELM). The calculation method between adjacent hidden layers in ML-ELM is as follows:

[0061] H i =h((β i ) T H i-1 ) (19)

[0062] Among them, the output weight β i It can be calculated according to formula (13).

[0063] By constructing the ML-ELM as described above, a deep network architecture based on the extreme learning machine can be established. In this paper, the proposed hybrid kernel extreme learning machine is used as the top classifier of ML-ELM, and the hybrid kernel extreme learning machine model is established in combination with the kernel method.

[0064] Furthermore, in step S3, the improved Sky Eagle algorithm is used to obtain various parameters, including the number of hidden layer nodes, kernel function type and weight, etc., to modify the hybrid kernel extreme learning machine model.

[0065] Initialize the parameters of the Eagle algorithm, including setting the current number of iterations t = 1, setting the maximum number of iterations T; and generating the fitness value of the Eagle algorithm and the position X of the randomly initialized population within the search range. i,j , the formula is as follows:

[0066] X ij =rand×(UB j -LB j )+LB j ,i=1,2,...,Nj=1,2,...,Dim (20)

[0067] Where rand is a random number, LB j represents the jth lower bound of the given problem, UB j represents the j-th upper bound for the given problem.

[0068] Secondly, different Sky Eagle optimization algorithms are selected according to the relationship between the current number of iterations t and the maximum number of iterations T, as well as the size of the random vector rand. The Sky Eagle optimization algorithms include: extended exploration (X1): selecting a hunting area by soaring at high altitudes with vertical dives, and narrowing the exploration range (X2): exploring in space through short-time gliding attacks, and extended development (X3): approaching prey at low altitudes through low flight and slow descent attacks, and narrowing the mining range (X4): grabbing prey by swooping down based on random movements.

[0069] The specific settings are as follows:

[0070] (1) In the traditional Skyhawk algorithm, if The algorithm will enter the global search phase, otherwise it will enter the local search phase, t is the number of iterations, and T is the number of allowed iterations. In order to better balance the global and local phases, an oscillating parameter that changes with the number of iterations is proposed, denoted as E, and its expression is:

[0071]

[0072] When E>0.2, the algorithm performs a global search, otherwise it performs a local search. After this improvement, the algorithm can also perform a local search in the middle of the iteration, making the algorithm search more thoroughly in the search space; while retaining the local search method in the late iteration, the global search is added, so that the algorithm still has the ability to search extensively in the late iteration, helping the algorithm avoid local optimality.

[0073] (2) If When E>0.2 and rand<0.5, extended exploration (X1) is started to identify the prey area and select the best hunting area by soaring vertically. The mathematical model of this behavior is as follows:

[0074]

[0075] Where X1(t+1) is the solution of the next iteration of t, generated by the first search method (extended exploration). best (t) is the th The best solution obtained before iterations, which reflects the approximate location of the prey. This equation Used to control the extended search (exploration) by the number of iterations. rand is a random value between 0 and 1. T and t represent the current number of iterations and the maximum number of iterations, respectively. M (t) represents the tth th The position average of the current solution of the connection at the iteration is expressed as follows.

[0076]

[0077] Where Dim is the dimension of the problem and N is the number of candidate solutions (total size).

[0078] (3) If When E>0.2 and rand>0.5, the exploration is reduced (X2). When the prey area is discovered from high altitude, the eagle will hover above the target prey, prepare to land, and then attack. In this behavior, the eagle reduces the exploration of the target prey's selection area to prepare for the attack. The mathematical model of this behavior is as follows:

[0079] X2(t+1)=X best (t)×Levy(D)+X R (t)+(yx)*rand (24)

[0080] where X2(t+1) is the solution of the next iteration of t, generated by the second search method (reduced exploration). D is the dimensional space, and Levy(D) is the Levy flight distribution function, calculated using equation (26). X R (t) is the i-th th A random solution in the range [1,N] is taken at the iteration.

[0081]

[0082] Where s is a constant value fixed to 0.01, and u and v are random numbers between 0 and 1. σ is calculated using equation (27).

[0083]

[0084] Where β is a constant value fixed to 1.5. In equation (25), y and x are used to represent the spiral shape during the search process, which is calculated as follows.

[0085] y=r×cos(θ) (27)

[0086] x=r×sin(θ) (28)

[0087] here

[0088] r=r1+U×D1 (29)

[0089] θ=-ω×D1+θ1 (30)

[0090]

[0091] R1 has a value between 1 and 20, while U is a small value fixed to 0.00565. D1 is an integer from 1 to the length of the search space (Dim), while ω is a small value fixed to 0.005.

[0092] (4) If When E<0.2 and rand<0.5, the third search method expansion development (X3) is started. When the prey area is accurately determined and ready for landing and attack, the eagle will descend vertically and make a preliminary attack to find the reaction of the prey. This method is called low flight and slow descent attack. In this behavior, the eagle uses the selected area to approach the prey and prepare for the attack. The representation of this behavior in the mathematical model is as follows:

[0093] X3(t+1)=(X best (t)-X M (t))×α-rand+((UB-LB)×rand+LB)×δ (32)

[0094] where X3(t+1) is the solution for the next iteration of t, generated by the third search method (X3). best (t) represents the approximate position of the prey before the i-th iteration (the best solution is obtained), X M (t) represents the i-th th The average value of the current solution at the iteration is calculated using equation (23). rand is a random value between 0 and 1. α, δ are development tuning parameters that are fixed to a small value (0.1) in this paper. LB represents the lower bound of the problem, and UB represents the upper bound of the given problem.

[0095] (5) If When E<0.2 and rand>0.5, the fourth search method is started to narrow down the development (X4). When the eagle approaches the prey, it will attack the prey at the last position according to their random movement. The mathematical model of this behavior is as follows:

[0096] X4(t+1)=X best (t)×QF-(G1×X(t)×rand)-G2×Levy(D)+rand×G1 (33)

[0097] Among them, X4(t+1) is the solution of the next iteration of t, generated by the fourth search method (X4). QF represents the quality function used to balance the search strategy, and its mathematical expression is as follows:

[0098]

[0099] G1 represents the various movements of the eagle used to track its prey during the process, and its mathematical expression is as follows:

[0100] G1=2×rand-1 (35)

[0101] QF(t) is the tth th The quality function value at the iteration, rand is a random value between 0 and 1. t and T represent the current iteration and the maximum number of iterations. Levy(D) is the taxed flight distribution function calculated using equation (25).

[0102] G2 in the traditional Skyhawk algorithm represents a decreasing value from 2 to 0, representing the flight slope of the AO from the first position (1) to the last position (t) during the period, generated using equation (36). X(t) is the tth th The current solution at iteration .

[0103]

[0104] In the improved Sky Eagle algorithm, a nonlinear decreasing parameter is used to replace the linear decreasing parameter in the traditional Sky Eagle algorithm. Its calculation formula is:

[0105]

[0106] Based on the new balance method, the algorithm performs local search in the middle of the iteration. At this time, the improved G2 has a larger value, which makes the algorithm more dependent on Levy flight, thereby searching the solution space more thoroughly. In the later stage of the iteration, the improved G2 is smaller and is less affected by the random large-scale search of Levy flight. Therefore, the optimal solution X obtained in the previous step best (t) and historical solutions X(t) will have a greater impact on the search results of this step, improving the local search capability in the later stages of iteration.

[0107] Furthermore, in step S4, the obtained parameters are put into a hybrid kernel extreme learning machine model and the training set data is loaded for training to obtain a deep kernel extreme learning machine. The model is called to perform classification prediction on the test set to determine whether the bearing has a fault during operation and determine the type of fault.

[0108] The number of hidden layers of the deep hybrid kernel extreme learning machine model is defined as n, and the kernel matrix of the model is expressed as follows:

[0109]

[0110] Therefore, the output classification equation of DHKELM given the input data x can be expressed as:

[0111]

[0112]

[0113] After the result is output, the predicted fault label is obtained, and the obtained fault label is its fault type. The predicted fault type is used to determine whether the bearing has a fault and determine the fault type.

[0114] Compared with the prior art, the present invention has the following beneficial effects:

[0115] (1) The advantage of empirical mode decomposition is that it can adaptively generate intrinsic mode functions based on the analyzed signal, and has a high signal-to-noise ratio and good time-frequency focusing.

[0116] (2) The use of deep hybrid kernel extreme learning machine makes the model have higher generalization and learning capabilities. The hyperparameters of the model can be adaptively selected without manual selection, thus avoiding the uncertainty caused by manual operation.

[0117] (3) On this basis, in order to solve the problem of hyperparameter selection in deep network models, the improved Eagle algorithm is introduced to find the optimal hyperparameters of the deep hybrid kernel extreme learning machine, and a self-optimizing deep hybrid kernel extreme learning machine is constructed as the final fault diagnosis model. The improved Eagle algorithm solves the weak local search ability of the original Eagle algorithm and introduces a new balanced search method. The global search and local search are well balanced, so that the model still has good local search ability in the later iteration, preventing it from falling into the local optimum.

[0118] (4) The model has made significant progress in feature extraction and fault classification capabilities, realizing the diagnosis of bearing faults and effectively overcoming the problem of poor performance of traditional bearing fault diagnosis methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0120] Figure 1 This is a general flowchart of the self-optimizing deep hybrid kernel extreme learning machine method in the present invention.

[0121] Figure 2 This is a flowchart of the improved Sky Eagle algorithm in the present invention.

[0122] Figure 3 It is a schematic diagram of the fault diagnosis results of the traditional HKELM method in the present invention.

[0123] Figure 4 Schematic diagram of the fault diagnosis results of the traditional AO-HKELM method in the present invention.

[0124] Figure 5 Schematic diagram of fault diagnosis results of the self-optimizing deep hybrid kernel extreme learning machine method in the present invention. DETAILED DESCRIPTION

[0125] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0126] Example 1

[0127] See also Figures 1 to 5, the technical solution provided in this embodiment is a bearing fault diagnosis method of a self-optimizing deep hybrid kernel extreme learning machine. This embodiment uses the Case Western Reserve University (CWRU) bearing dataset, referred to as the CWRU dataset, which is a rolling bearing fault dataset publicly provided by the Bearing Data Center of the Electrical Laboratory of Case Western Reserve University. The experiment uses a 2-horsepower motor, a torque sensor, a power tester, etc. The rolling bearing models at the drive end and the fan end are 6205-2RS SKF and SKF6203 deep groove ball bearings, respectively. The parameter information is shown in Table 1. The experiment uses electrospark machining to introduce single-point faults to the experimental bearings, and the fault diameters are 7, 14, 21, 28 and 40 mils (mil, 1 mil = 0.001 inch = 0.0254 mm). The vibration data is collected using an accelerometer and connected to the housing with a magnetic base. The accelerometer is placed at the 12 o'clock position at the drive end and the fan end of the motor housing. In the experiment, the vibration signals collected by the recorder are all post-processed in the Matlab environment. All data files are in Matlab (*.mat) format. Digital data were collected at 12,000 samples per second, and data for the drive-end bearing fault were collected at 48,000 samples per second. The outer raceway fault is a stationary fault, so the location of the fault relative to the bearing load area directly affects the vibration response of the motor / bearing system.

[0128] Table 1 Rolling bearing parameter information

[0129]

[0130] To quantify this effect, experiments were conducted on fans and drive-end bearings with external channel faults at 3 points (directly in the load zone), 6 points (orthogonal to the load zone), and 12 points. The entire experiment used a 16-channel DAT recorder to collect vibration signals at motor speeds of 1797 rpm (0 load), 1750 rpm (1 load), 1730 rpm (2 loads), and 1722 rpm (3 loads), with a sampling frequency of 12 kHz. There are 10 states at each speed, namely the fault state at three positions of the inner ring of the bearing, the fault state at three positions of the outer ring of the bearing, the fault state at three positions of the rolling element, and the normal operation state, as shown in Table 2.

[0131] This example uses 48K drive-end data, 9 types of fault data and 1 normal data as the entire data set, and divides it into training set and test set in a ratio of 7:3. 100 groups of samples are taken for each fault, and each group contains 1024 vibration signal data.

[0132] In the bearing diagnosis simulation of this embodiment, the traditional Eagle algorithm-deep hybrid kernel extreme learning machine (AO-DHKELM), hybrid kernel extreme learning machine (HKELM) and the self-optimizing deep hybrid kernel extreme learning machine of this embodiment are used for comparison, and 50 iterations are used as the number of iterations. The initial population number of the Eagle algorithm and the improved Eagle algorithm is 30. The three hidden layers of the improved Eagle algorithm contain 19, 50 and 45 nodes respectively. The weight of the Gaussian kernel function is 0.3832, and the value is 84.6718, which is usually a parameter controlling the kernel width; the weight of the polynomial kernel function is 0.6168, and the value is 211.9314, which represents the scaling parameter, and the power exponent of the polynomial kernel is 2; the penalty coefficient of the top kernel extreme learning machine is 440.8263.

[0133] Table 2 Schematic diagram of bearing failure

[0134] Serial number Fault location Fault Size (mils) Label 1 Rolling element 7 a1 2 Rolling element 14 a2 3 Rolling element 21 a3 4 Inner Circle 7 a4 5 Inner Circle 14 a5 6 Inner Circle 21 a6 7 Outer ring 7 a7 8 Outer ring 14 a8 9 Outer ring 21 a9 10 normal a10

[0135] The accuracy index is used as the evaluation index of the test set, and the calculation formula is as follows:

[0136]

[0137] In the formula, acc is the classification accuracy, the closer its value is to 1, the better, TP (true positive) represents the number of positive samples judged as positive samples, TN (true negative) represents the number of negative samples judged as negative samples, and ALL represents the number of all samples.

[0138] Table 3 shows the fault diagnosis effects of the traditional Sky Eagle algorithm-deep hybrid kernel extreme learning machine (AO-DHKELM), hybrid kernel extreme learning machine (HKELM) and the self-optimizing deep hybrid kernel extreme learning machine of this embodiment on the CWRU data set.

[0139] Table 3 Diagnostic effect

[0140]

[0141] As can be seen from Table 3 above, the HKELM algorithm can accurately identify the normal state diagnosis just like the latter two methods; however, it is only 82.39% for rolling element faults, which still has room for improvement; the classification accuracy for inner ring faults reached 84.66%, and for outer ring faults reached 86.80%, but 90% of the faults were not correctly identified; although the traditional AO-DHKELM algorithm introduced the Sky Eagle algorithm to select the optimal value compared to the traditional HKELM algorithm, it has a weaker local search capability for complex bearing faults, and the search capability becomes increasingly weaker in the later iterations, and it is very easy to fall into the local optimum. Although it can achieve a classification accuracy of more than 96% for rolling element faults and inner and outer ring faults, it cannot be fully identified like inner ring faults and normal states, and there is still room for improvement.

[0142] Compared with the above two traditional algorithms, the self-optimizing deep hybrid kernel extreme learning machine of this embodiment introduces a new balanced search method, which can not only ensure that the global search and local search are balanced, but also enable the model to still have strong local search capabilities in the later stage of iteration. From the results, the effect is significant. Both the inner ring fault and the outer ring fault can be fully identified, and the fault classification recognition rate of the rolling element fault has also reached 98.98%, which is a significant improvement. Therefore, the self-optimizing deep hybrid kernel extreme learning machine algorithm provided in this embodiment can effectively improve the diagnosis effect of bearing faults.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on a self-optimizing deep hybrid kernel extreme learning machine, characterized in that: The following steps are involved: Step S1, performing empirical mode decomposition on the collected bearing vibration signal to obtain the intrinsic mode components that conform to the intrinsic mode principle, inputting the decomposed signal into the Hilbert transform model, performing envelope mediation analysis to extract fault information, and dividing the data set after the fault information is extracted into a training set and a test set; Step S2, setting the number of hidden layers and activation functions, creating a multi-hidden layer extreme learning machine by stacking extreme learning machine-based autoencoders, and establishing a hybrid kernel extreme learning machine model by constructing a hybrid kernel extreme learning machine as an ML-ELM classifier combined with a kernel method; Step S3, using the improved Sky Eagle algorithm to obtain various parameters, including the number of hidden layer nodes, kernel function type and weight, etc., to modify the hybrid kernel extreme learning machine model; Step S4, putting the obtained parameters into the hybrid kernel extreme learning machine model, loading the training set data for training to obtain a deep kernel extreme learning machine, calling the model to perform classification prediction on the test set, judging whether the bearing has a fault during operation and determining the fault type.

2. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 1 is characterized in that: In the step S1, the collected bearing vibration signal is subjected to empirical mode decomposition, and when the inherent mode components that conform to the inherent mode principle are input into the Hilbert transform model, the Hilbert transform is used to remove the unnecessary high-frequency vibration components of the signal, and the excitation components are retained to replace the original signal for envelope mediation analysis to extract fault information, and then the data set is divided into a test sample set and a training sample set in proportion.

3. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 1 is characterized in that: In step S2, the number of hidden layers and activation functions are set, a multi-hidden layer extreme learning machine is created by stacking extreme learning machine-based autoencoders, a hybrid kernel extreme learning machine is constructed as a multi-hidden layer extreme learning machine classifier, and a hybrid kernel extreme learning machine model is established in combination with the kernel method.

4. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 3 is characterized in that: In step S2, two kernel functions are used as basis functions, and a new hybrid kernel function is established through a multi-kernel learning method as the kernel function of the kernel extreme learning machine, thereby establishing a new hybrid kernel learning machine model.

5. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 1 is characterized in that: In step S3, an oscillating parameter that changes with the number of iterations is proposed and a new balanced search method is used to form an improved Sky Eagle algorithm to obtain various parameters, including the number of hidden layer nodes, kernel function type and weight parameters to correct the hybrid kernel extreme learning machine model.

6. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 5 is characterized in that: In step S3, in order to better balance the global and local stages, an oscillating parameter that changes with the number of iterations is proposed. After this improvement, the algorithm also performs local search in the middle of the iteration, making the algorithm search more thoroughly in the search space; while retaining the local search method in the later stage of the iteration, a global search is added, so that the algorithm has the ability to search widely in the later stage of the iteration, helping the algorithm to avoid local optimality.

7. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 5 is characterized in that: In step S3, in the improved Sky Eagle algorithm, a nonlinear decreasing parameter is used as a substitute for the linear decreasing parameter in the traditional Sky Eagle algorithm. Based on the new balance method, the algorithm performs a local search in the middle of the iteration to search the solution space; the optimal solution and historical solution obtained in the later stage of the iteration affect the search results of this step.

8. The bearing fault diagnosis method of the self-optimizing deep hybrid kernel extreme learning machine according to claim 1 is characterized in that: In step S4, the obtained parameters are put into the hybrid kernel extreme learning machine model and the training set data is loaded for training to obtain a deep kernel extreme learning machine. The model is called to perform classification prediction on the test set to determine whether the bearing has a fault during operation and determine the type of fault.

Images