Training method, diagnosis method and diagnosis device of bearing fault diagnosis model

Abstract

The present invention proposes a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model, wherein the training method comprises: collecting fault vibration signals of different types of bearing faults to obtain a sample set; performing empirical Fourier decomposition on the fault vibration signals in the sample set to obtain multiple modal signal components; screening the multiple modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence; processing the reconstructed signal sequence by an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set; and training the bearing fault diagnosis model by using the training feature set to obtain a trained bearing fault diagnosis model. The present invention can effectively identify bearing faults of different types and damage degrees, with an average recognition rate of 99.74%.

Description

Technical Field

[0001] The invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model. Background Art

[0002] Mechanical equipment is developing in the direction of large-scale and precision, and its intelligence is also increasing, which also puts forward higher requirements for the reliability of equipment operation. As the core slewing bearing component in rotating machinery, a slight defect on its surface may cause the operation failure of the entire device system, resulting in huge casualties and property losses. The traditional method for rolling bearing detection requires staff to regularly remove the bearings from mechanical equipment for safety inspection, which consumes a lot of manpower and material resources and affects the industrial production process. The intelligent maintenance solution uses advanced sensors and detection technologies to obtain the status information of equipment operation. When a fault occurs, it can also predict the remaining life of the equipment by establishing physical and statistical models, and carry out maintenance work on the equipment before the equipment fails completely. Therefore, carrying out bearing fault diagnosis and status detection is of great significance to improving the safety of equipment operation.

[0003] Bearings are key precision components of rotating machinery. Various complex working conditions can cause bearings to suffer from fatigue peeling, cracks, wear, indentation and other damage, aggravating bearing vibration and causing safety hazards. In order to reduce the occurrence of accidents, it is necessary to monitor the state and diagnose faults of bearings. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model to realize bearing condition monitoring and fault diagnosis, thereby avoiding fatigue peeling, cracks, wear, indentation and other damages on the bearing that aggravate bearing vibration, thereby avoiding safety hazards.

[0005] The present invention proposes a training method for a bearing fault diagnosis model, comprising:

[0006] Collect fault vibration signals of different types of bearing faults to obtain a sample set;

[0007] Performing empirical Fourier decomposition on the fault vibration signal in the sample set to obtain multiple modal signal components;

[0008] Screening the plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence;

[0009] The reconstructed signal sequence is processed using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set;

[0010] The bearing fault diagnosis model is trained using the training feature set to obtain a trained bearing fault diagnosis model.

[0011] In one embodiment of the present invention, the different types of faults include at least normal, inner ring single point fault, outer ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, inner ring multi-point fault, outer ring multi-point fault and ball multi-point fault.

[0012] In one embodiment of the present invention, the step of performing empirical Fourier decomposition on the sample set to obtain multiple modal signal components includes:

[0013] Normalizing the Fourier spectrum of the fault vibration signal in the sample set to the interval [0, π], and predefining the number of frequency bands N required to be divided;

[0014] The initial value and the maximum value in the Fourier spectrum are set as M control points, and the final number of control points T is obtained according to the number of frequency bands N and the number of control points M;

[0015] The interval is divided according to the final number of control points T, and the position of each control point is α n , where 1≤n≤T, α 1 =0,α T+1 =π;

[0016] according to Determine the spectrum segmentation boundary ω of Fourier decomposition n , define the first T boundaries as min[α T-1 , α T ], where Λ n is the set of the first T boundaries;

[0017] Determine the continuous interval Ω=[ω i ,ω i+1 ], and calculate the analytical Fourier natural band function of each interval, where i = 1, 2, ..., T;

[0018] The real part of the analytical Fourier natural frequency band function of each interval is inversely Fourier transformed to obtain multiple modal signal components.

[0019] In one embodiment of the present invention, obtaining the final number of control points T according to the number of frequency bands N and the number of control points M includes:

[0020] When M ≥ N, only the first N segments of M control points in descending order are taken, that is, the final number of control points T is N;

[0021] When M<N, the frequency bands that can be divided in the vibration signal are less than the predefined number of frequency bands N, and the final number of control points T is M.

[0022] In one embodiment of the present invention, screening the plurality of modal signal components by kurtosis and correlation coefficient to obtain a reconstructed signal comprises:

[0023] Calculate the kurtosis value Kurt of each modal signal component n , and retain the kurtosis value Kurt n All components greater than 3 to form the set U;

[0024] Calculate the correlation coefficient ρ between each modal signal component and the fault vibration signal y n , and the correlation coefficient ρ n The standard deviation of is taken as the minimum threshold to screen each of the modal signal components as the main rotation components of the fault vibration signal to form a set V;

[0025] Take the union of set U and set V and add the signals therein to obtain the reconstructed signal sequence.

[0026] In one embodiment of the present invention, the kurtosis value Kurt n By formula:

[0027] Calculated, where efd n,k is the k-point of the n-th modal signal component; n = 1, 2, ..., T; k = 1, 2, ..., L; u and σ are efd n The mean and variance of ; E[n] represents the mathematical expectation of n.

[0028] In one embodiment of the present invention, the correlation coefficient ρ n By formula:

[0029] Calculated, among which, represents the average value of the nth modal component, z k represents the kth sampling point of the fault vibration signal z, Represents the mean of all sampling points of the fault vibration signal z.

[0030] In one embodiment of the present invention, the reconstructed signal sequence is processed using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set, including:

[0031] Given a reconstructed signal sequence x = {x i , i = 1, 2, ..., L}, and calculate the average operator P 0 and the difference operator P 1 : Among them, x i represents the reconstructed signal of the i-th interval;

[0032] Define operators at different levels s (f = 0 or 1), where

[0033]

[0034] Given a specific vector sequence [r 1 ,r 2 ,…,r s ], and determine the f value of each layer according to the node e of the sth layer. The relationship between the node e and the vector sequence is: In the formula, {r d ,d=1,2,…,s}∈{0,1} represents the average operator or difference operator of the dth layer;

[0035] Get the hierarchical components of the reconstructed signal sequence, expressed as:

[0036] Calculate the improved hierarchical slope entropy of each hierarchical component, expressed as: MHSE(x,m,s,e,γ,δ)=SE(X s,e ,m,γ,δ), where m is the embedding dimension, γ and δ are the thresholds of slope entropy.

[0037] In one embodiment of the present invention, the step of training the bearing fault diagnosis model using the training feature set to obtain a trained bearing fault diagnosis model includes:

[0038] The training feature set is represented as {u i ,i=1,2,…,p}, and define the weight between the input layer and the hidden layer as The weight between the hidden and output layers is β, the hidden layer threshold is b, and p = 2 s *(m-1):

[0039] b=[b 1 ,b 2 ,…,b Q ] T , where R, Q, and F are the number of neurons in the input layer, hidden layer, and output layer of the extreme learning machine respectively;

[0040] Assuming an activation function g(x), the output T of the extreme learning machine * Expressed as: T * =Hβ, where

[0041] Solve the minimum value of its approximate square error to obtain the weight of the output layer: And find its optimal solution as β * =H + T * , where H+ is the generalized inverse of the hidden layer output matrix.

[0042] The present invention also provides a bearing fault diagnosis method, comprising:

[0043] Collect bearing fault vibration signals;

[0044] Performing empirical Fourier decomposition on the fault vibration signal to obtain multiple modal signal components;

[0045] Screening the plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence;

[0046] The reconstructed signal sequence is processed using an improved hierarchical slope entropy algorithm to obtain fault feature information;

[0047] The fault feature information is input into a bearing fault diagnosis model trained by the bearing fault diagnosis model training method described in any one of claims 1 to 9 to obtain a diagnosis result of the bearing fault.

[0048] The present invention also provides a bearing fault diagnosis device, comprising:

[0049] A fault signal acquisition module is used to collect fault vibration signals of different types of bearing faults to obtain a sample set;

[0050] A modal signal acquisition module, connected to the fault signal acquisition module, for performing empirical Fourier decomposition on the fault vibration signal in the sample set to obtain multiple modal signal components;

[0051] A signal reconstruction module, connected to the modal signal acquisition module, for screening a plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence;

[0052] A training feature set acquisition module, connected to the signal reconstruction module, for processing the reconstructed signal sequence using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set;

[0053] A model building module, connected to the training feature set acquisition module, is used to train the bearing fault diagnosis model using the training feature set to obtain a trained bearing fault diagnosis model;

[0054] A fault diagnosis module is connected to the model building module and is used to obtain a diagnosis result.

[0055] The present invention proposes a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model, which utilizes empirical Fourier decomposition to preprocess rolling bearing fault signals, thereby filtering out the noise segment of the Fourier spectrum and highlighting the main vibration component and the impact component of the time series.

[0056] Compared with the coarse-grained sequence method, the improved hierarchical slope entropy algorithm proposed in the present invention can extract more comprehensive and rich time series scale information; at the same time, using the slope to measure the time series difference can perceive the trend and degree of amplitude change, which is more effective than the method of sorting by amplitude size using relative permutation entropy (PE).

[0057] The present invention provides a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model, which can effectively identify bearing faults of different types and damage degrees, with an average recognition rate of 99.74%. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0059] Figure 1 This is a flowchart of the bearing fault diagnosis method proposed in the present invention.

[0060] Figure 2 This is a flow chart of the bearing fault diagnosis method proposed in the present invention.

[0061] Figure 3 is the relationship between time series difference and sign.

[0062] Figure 4 is the symbol sequence pattern when the embedding dimension m=3.

[0063] Figure 5 This is a schematic diagram of the hierarchical decomposition of the bearing fault diagnosis method proposed in the present invention.

[0064] Figure 6 This is a schematic diagram of the bearing fault diagnosis device proposed in the present invention.

[0065] Figure 7 This is a waveform diagram of the first five components obtained through modal decomposition of an outer ring multi-point fault in one embodiment of the present invention.

[0066] Figure 8 This is the t-SNE visualization dimensionality reduction result of MHSE features of 9 groups of fault samples in one embodiment of the present invention.

[0067] Fig. 9 FIG. 5 is a test set classification confusion matrix in one embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.

[0069] It should be noted that the illustrations provided in this embodiment are only used to illustrate the basic concept of the present invention in a schematic manner. Therefore, the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0070] The present invention proposes a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model to realize state monitoring and fault diagnosis of the bearing, thereby avoiding fatigue peeling, cracks, wear, indentation and other damages on the bearing that aggravate the bearing vibration, thereby avoiding potential safety hazards. Specifically, Figure 1 and Figure 2 As shown, the diagnostic method comprises:

[0071] S1. Collecting fault vibration signals of different types of bearing faults to obtain a sample set;

[0072] S2. performing empirical Fourier decomposition on the fault vibration signal in the sample set to obtain multiple modal signal components;

[0073] S3, screening the plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence;

[0074] S4, using an improved hierarchical slope entropy algorithm to process the reconstructed signal sequence to obtain fault feature information and form a training feature set;

[0075] S5. Training the bearing fault diagnosis model using the training feature set to obtain a trained bearing fault diagnosis model;

[0076] like Figure 1 and Figure 2As shown, in step S1, by collecting different types of fault vibration signals of the bearing, a sample set is obtained, and the sample set is used to train and obtain a diagnostic model. In this embodiment, the different types of faults include at least normal, inner ring single point fault, outer ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, inner ring multi-point fault, outer ring multi-point fault and ball multi-point fault.

[0077] like Figure 1 and Figure 2 As shown, in step S2, the fault vibration signal in the sample set is subjected to empirical Fourier decomposition to obtain multiple modal signal components, wherein the empirical Fourier decomposition divides the frequency bands by an improved spectrum segmentation technique, and then constructs a zero-phase filter for filtering. After each frequency band is subjected to inverse Fourier transform, the vibration signal decomposition components can be obtained. Specifically, the steps of obtaining multiple modal signal components include:

[0078] S21, normalizing the fault vibration signal in the sample set to the interval [0, π] by Fourier spectrum, and predefining the number of frequency bands N required to be divided;

[0079] S22, determining M control points according to the initial value and the maximum value in the Fourier spectrum, and obtaining the final number of control points T according to the number of frequency bands N and the number of control points M;

[0080] S23, divide the interval according to the final number of control points T, and the position of each control point is α n , where 1≤n≤T, α 1 =0,α T+1 =π;

[0081] S24, according to Determine the spectrum segmentation boundary ω of Fourier decomposition n , define the first T boundaries as min[α T-1 , α T ], where Λ n is the set of the first T boundaries;

[0082] S25, determine the continuous interval Ω=[ω i ,ω i+1 ], and calculate the analytical Fourier natural band function of each interval, where i = 1, 2, ..., T;

[0083] S26. Perform inverse Fourier transform on the real part of the analytical Fourier natural frequency band function of each interval to obtain multiple modal signal components.

[0084] like Figure 1 and Figure 2As shown, it should be noted that the relationship between the final number of control points T and the number of frequency bands N and the number of control points M is: when M≥N, only the first N segments of the M control points arranged in descending order are taken, that is, the final number of control points T is N; when M<N, the frequency band that can be divided in the vibration signal is less than the predefined number of frequency bands N, and the final number of control points T is M.

[0085] like Figure 1 As shown, the interval is divided according to the final number of control points T, and the position of each control point is α n , where 1≤n≤T, α 1 =0,α T+1 =π; It should be noted that T intervals require a total of T+1 boundaries, and the first T boundaries min[α T-1 , α T ], the set of the first T boundaries is Λ n , the T+1th boundary is α T and α T+1 The midpoint of the Fourier decomposition spectrum can be determined by n , In step S25, determine the continuous interval Ω=[ω i ,ω i+1 ], and calculate the analytical Fourier natural band function of each interval, where i = 1, 2, ..., T. Specifically, assuming that x[n] is a discrete continuous sequence of length L, its discrete Fourier form can be written as:

[0086]

[0087] In the formula, is the discrete Fourier transform of the sequence X[n]. Let L be an even number, and we know that X[0] and are all real numbers. Rewrite it as:

[0088]

[0089] In the formula, Re{z 1 [n]} represents z 1 The real part of [n], z 1 [n] and z 2 [n] are complex conjugates of each other. Analytical signal z 1 [n] can be written as follows:

[0090]

[0091] Compute the analytical Fourier natural band functions for each interval:

[0092] In the formula, i=1, 2,…,T.

[0093] In step S26, after the real part of the analytical Fourier natural frequency band function of each interval is inverse Fourier transformed, a plurality of modal signal components (efd components) corresponding to each interval are obtained.

[0094] like Figure 1 and Figure 2 As shown, in step S3, the plurality of modal signal components are screened by kurtosis and correlation coefficient respectively, and the screening results are reconstructed to obtain a reconstructed signal sequence, specifically, including:

[0095] S31, calculate the kurtosis value Kurt of each modal signal component n , and retain the kurtosis value Kurt n All components greater than 3 to form a set U, where the kurtosis value Kurt n By formula:

[0096] Calculated, where efd n,k is the k-point of the n-th modal signal component; n = 1, 2, ..., T; k = 1, 2, ..., L; u and σ are efd n The mean and variance of ; E[n] represents the mathematical expectation of n.

[0097] S32, calculating the correlation coefficient ρ between each modal signal component and the fault vibration signal y n , and the correlation coefficient ρ n The standard deviation of is taken as the minimum threshold to screen each of the modal signal components as the main rotation components of the fault vibration signal to form a set V, where the correlation coefficient ρ n By formula:

[0098] Calculated, among which, represents the average value of the nth modal component, z k represents the kth sampling point of the fault vibration signal z, Represents the mean of all sampling points of the fault vibration signal z.

[0099] S33, taking the union of set U and set V and adding the signals therein to obtain a reconstructed signal sequence.

[0100] In view of the drawback that permutation entropy lacks completeness in the use of amplitude information, a slope entropy (SE) based on sign and amplitude information is proposed. SE uses a method similar to linear quantization to divide the slopes between sequence elements into different modes by defining threshold intervals, so as to measure the change trend of amplitude. Specifically, it includes:

[0101] First, assume that a time series y = {y i , i=1,2,…,L}, and reconstruct its phase space to obtain the subsequence Where t = 1, 2, ..., L-m + 1, m is the embedding dimension. Define the vertical increment threshold γ and the horizontal increment threshold δ. γ is a large quantity used to measure the significant difference between vector sequences to distinguish different fluctuation amplitudes; δ is a very small value to classify the situation of approximately equal amplitudes. Then, the time series is differentiated and the subsequences are defined as different symbols through the threshold. Figure 3 Shows the relationship between time series differencing and sign.

[0102] When t+τ -y τ When >γ, the sign is defined as +2.

[0103] When t+τ -y τ ≥δ and y t+τ -y τ <γ, the symbol is defined as +1. When γ=1, the slope change is equivalent to being in the range of [0,45°).

[0104] When |y t+τ -y τ When |≤γ, it is in the area close to 0 difference and the sign is set to 0.

[0105] When t+τ -y τ ≤-δ and y t+τ -y τ >-γ, the symbol is defined as -1. When γ=1, the slope change is equivalent to being in the range of [-45°,0).

[0106] When t+τ -y τ When <-γ, the symbol is defined as -2.

[0107] The time series is defined as above and the symbol sequence S m ={s i ,i=1,2…,L-1}, the symbolic subspace can be expressed as: Where l = 1, 2, ..., L-m + 2. The subspaces with dimensions lower than ε = m-1 are also counted, and the dimensions that can be obtained are d = 1, 2 ..., ∈, and the sequence pattern composed of them is recorded as:

[0108]

[0109] in are subspaces of different dimensions. Figure 4The symbol sequence pattern is shown when the embedding dimension m = 3. The total number of permutation patterns Z that appear in each sequence pattern is counted d , each arrangement appears h times d,i , we can get the corresponding probability p d,i =h d,i / Z d , where i = 1, 2, ..., Z d . Define the Shannon entropy for each sequence mode:

[0110]

[0111] The slope entropy can be expressed as:

[0112] SE = {se d ,d=1,2,…,ε}.

[0113] In this embodiment, if Figures 1 to 5 As shown, in step S4, it should be noted that the hierarchical entropy (HE) obtains implicit high and low frequency information through the average and difference method, but its hierarchical process needs to ensure that the data length L = 2 n (n is a positive integer). At the same time, the hierarchical entropy will gradually lose its statistical significance as the level increases, reducing the reliability of the decomposition. In this embodiment, the improved hierarchical slope entropy algorithm is used to process the reconstructed signal sequence to obtain fault feature information and form a training feature set, which overcomes the above two shortcomings and improves the ability to extract sequence scale information. Specifically, the improved hierarchical slope entropy algorithm is used to process the reconstructed signal sequence to obtain fault feature information and form a training feature set, including:

[0114] S41, given a reconstructed signal sequence x={x i , i = 1, 2, ..., L}, and calculate the average operator P 0 and the difference operator P 1 : Among them, x i represents the reconstructed signal of the i-th interval;

[0115] S42. Define operators at different levels s (f = 0 or 1), where

[0116]

[0117] S43. Given a specific vector sequence [r 1 ,r 2 ,…,r s ], and determine the f value of each layer according to the node e of the sth layer. The relationship between the node e and the vector sequence is: In the formula, {r d,d=1,2,…,s}∈{0,1} represents the average operator or differential operator of the dth layer; node e can refer to Figure 3 , e starts from 0. When s=1, there are two nodes 0 and 1; when s=2, there are four nodes 0, 1, 2, and 3.

[0118] It should be noted that in order to obtain the hierarchical information of the time series, the operator defined above needs to be used repeatedly. A specific vector sequence is given by Determine, for example, the fourth level component X of the second level 2,3 2 means s=2, i.e. the second layer; 3 means e=3, i.e. the fourth component from 0 to 3, i.e. the fourth component in the second layer, as shown in the formula:

[0119] We can get: x is actually the original sequence y, r d That is f, f can only be 0 or 1. So find X 2,3 The process is: 2 1 *r 1 +2 0 *r 2 =3, (r 1 ,r 2 ∈{0,1}), when r 1 =1, r 2 =1, this equation holds true.

[0120] According to the node e of each layer, the vector sequence corresponding to this node is determined. Each node in different layers has a corresponding vector sequence [r 1 ,r 2 ,…,r s ], for example X 2,3 The vector sequence of

[11] , that is, r 1 =1,r 2 =1;X 3,3 The vector sequence

[110] , that is, r 1 =1,r 2 =1,r 3 =0, that is, the e of each layer determines all f values ​​from the first layer to the current layer.

[0121] S44, obtaining the hierarchical components of the reconstructed signal sequence, expressed as:

[0122] S45. Calculate the improved hierarchical slope entropy of each hierarchical component, expressed as: MHSE(x,m,s,e,γ,δ)=SE(X s,e,m,γ,δ), where m is the embedding dimension, γ and δ are the thresholds of slope entropy. Some parameters of MHSE need to be determined manually, namely, embedding dimension m, decomposition level s, thresholds γ and δ. In order to ensure that MHSE contains at least two symbol patterns, the embedding dimension is m>2. In this paper, m=3 is selected. Since the number of decomposition levels is too large, the calculation is more time-consuming, and s=3 is usually set. The thresholds γ and δ determine the classification interval of the sequence slope change. In this paper, γ=1 and δ=0.001 are taken. The improved hierarchical slope entropy is the training feature set.

[0123] like Figure 1 As shown, in step S5, the bearing fault diagnosis model is trained using the training feature set to obtain a trained bearing fault diagnosis model, for example, an extreme learning machine is used to train and construct a diagnosis model, that is, the training feature set is input into the extreme learning machine to obtain a diagnosis model. The step of obtaining a trained bearing fault diagnosis model includes:

[0124] The training feature set is represented as {u i ,i=1,2,…,p}, and define the weight between the input layer and the hidden layer as The weight between the hidden and output layers is β, the hidden layer threshold is b, and p = 2 s *(m-1):

[0125] b=[b 1 ,b 2 ,…,b Q ] T , where R, Q, and F are the number of neurons in the input layer, hidden layer, and output layer of the extreme learning machine, respectively. The value of p is related to the level s and the embedding dimension m of the level entropy, that is, p = 2 s *(m-1);

[0126] Assuming an activation function g(x), the output T of the extreme learning machine * Expressed as: T * =Hβ, where It should be noted that the input layer is multiplied by the weight Add b and fit the activation function g(x) to get the hidden layer H, where A randomly takes values ​​between [-1,1], and b randomly takes values ​​between [0,1]. The activation function g(x) can be Sigmoid function, Tanh function, sin function, etc. For example, the Sigmoid function is expressed as: The activation function can be used to approximate any nonlinear function, so that neural networks can be applied to many nonlinear models.

[0127] Solve the minimum value of its approximate square error to obtain the weight of the output layer: And find its optimal solution as β * =H + T * , where H + is the generalized inverse of the hidden layer output matrix.

[0128] like Figure 1 As shown, in this embodiment, the present invention proposes a bearing fault diagnosis method, which uses the diagnosis model constructed in the above embodiment to diagnose the bearing fault, including:

[0129] S61, collecting fault vibration signals of the bearing;

[0130] S62, performing empirical Fourier decomposition on the fault vibration signal to obtain multiple modal signal components;

[0131] S63, screening the plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence;

[0132] S64, processing the reconstructed signal sequence using an improved hierarchical slope entropy algorithm to obtain fault feature information;

[0133] S65. Input the fault feature information into the bearing fault diagnosis model trained by the bearing fault diagnosis model training method to obtain a diagnosis result of the bearing fault.

[0134] like Figure 2 As shown, in some embodiments, the fault vibration signals collected in the above embodiments and the output diagnostic results can also be constructed as a test set or another test set can be constructed, and the test set can be used as an input of the diagnostic model to continuously optimize the diagnostic model.

[0135] like Figure 6As shown, the present invention also proposes a bearing fault diagnosis device, which applies the bearing fault diagnosis method described in the above embodiment. Specifically, the diagnosis model includes a fault signal acquisition module 10, a modal signal acquisition module 20, a signal reconstruction module 30, a training feature set acquisition module 40, a model construction module 50 and a fault diagnosis module 60. Among them, the fault signal acquisition module 10 is used to collect fault vibration signals of different types of bearing faults to obtain a sample set; the modal signal acquisition module 20 is connected to the fault signal acquisition module 10, and is used to perform empirical Fourier decomposition on the fault vibration signals in the sample set to obtain multiple modal signal components; the signal reconstruction module 30 is connected to the modal signal acquisition module 20, and is used to screen the multiple modal signal components by kurtosis and correlation coefficient respectively, and reconstruct the screening results to obtain a reconstructed signal sequence; the training feature set acquisition module 40 is connected to the signal reconstruction module 30, and is used to process the reconstructed signal sequence using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set; the model construction module 50 is connected to the training feature set acquisition module 40, and is used to train the bearing fault diagnosis model using the training feature set to obtain a trained bearing fault diagnosis model; the fault diagnosis module 60 is connected to the model construction module 50, and is used to obtain the diagnosis result.

[0136] like Figure 1 As shown, in this embodiment, the experimental bearings are NSK aircraft engine bearings of model NU1010EM (inner ring removable type) and N10-10EM (outer ring removable type), and the bearing component damage is processed by laser marking machine and wire cutting machine. The bearing damage types include healthy, inner ring, outer ring, roller single point and multi-point faults, as well as outer ring / roller, inner ring / roller composite faults, a total of 9 states, with numbers 1 to 9 as the category labels of different faults, and Table 1 shows their corresponding relationships.

[0137] Table 1 Correspondence table of bearing fault types, quantities and locations

[0138]

[0139] The experimental data were collected using LMS Test.lab software with a sampling frequency of 20480 Hz. The bearing operating conditions were an axial load of 3 kN and a rotation speed of 3000 rpm. All fault signals were segmented into 1000 sample points per segment. 100 samples were taken for each fault, 20 of which were randomly selected as the sample set, and the remaining 80 samples were used as the test set. There were 180 samples in the sample set and 720 samples in the test set. The test set was used to test the trained diagnostic model.

[0140] First, empirical Fourier decomposition is performed on the fault sample signal. In order to separate the spectrum bands of the noise as much as possible, the number of frequency bands N=20 is predefined. Figure 7 The first five component waveforms obtained by EFD decomposition of the outer ring multi-point fault. The decomposed and reconstructed signal contains the main frequency and impact frequency of the fault vibration signal. After the rolling bearing fault signal is decomposed and reconstructed, the MHSE algorithm is used to extract features, and a 16-dimensional feature vector can be obtained for each sample. Figure 8 The t-SNE (t-distributed Stochastic Neighbor Embedding) visualization dimension reduction results of the MHSE features of 9 groups of fault samples. The numbers 1 to 9 correspond to different fault type labels. Figure 8 It can be seen that the intra-class features of the nine faults are compact, except that some features of label 5 are close to label 7, and the distinction between the other faults is obvious. Overall, the MHSE algorithm can effectively characterize different fault types and damage degrees.

[0141] At the same time, in this embodiment, in order to verify the feature extraction performance of the improved hierarchical slope entropy algorithm, it is compared with the existing five algorithms, namely, improved hierarchical permutation entropy (MHPE), refined composite multiscale dispersion entropy (Refined Composite Multiscale D-ispersion Entropy, RCMDE), refined composite multiscale sample entropy (Refined Composite Multis-cale Sample Entropy, RCMSE), refined composite multiscale fuzzy entropy (Refined Composite Mult-iscale Fuzzy Entorpy, RCMFE), composite multiscale weighted permutation entropy (Composite Multiscale Weighted Permutation Entropy, CWMPE), and the specific parameter settings of the algorithms are shown in Table 2.

[0142] Table 2 Specific parameters of the five algorithms

[0143]

[0144] After extracting the reconstructed signal features, the silhouette coefficient and Davidson-Boting index (DBI) are used to jointly evaluate the overall clustering characteristics of different methods. The silhouette coefficient measures the distance between data by defining cohesion and dispersion. The value range is distributed in [0,1]. The closer the value is to 1, the smaller the dispersion within the class and the larger the dispersion between classes. DBI is the mean of the maximum similarity between classes. Generally, the smaller the value, the greater the difference between classes and the better the classification effect. The evaluation results of the indicators and the calculation time of a single sample are shown in Table 3.

[0145] Table 3 Evaluation results of two indicators and calculation time of a single sample

[0146]

[0147] As can be seen from Table 3, the two indicators of MHSE are the best among the six algorithms, showing good linear clustering characteristics. In terms of sample calculation time of the same sampling point, MHSE is second only to RCMDE. Although RCMDE has the fastest operation speed, its linear clustering characteristics are much lower than MHSE.

[0148] In addition, in this embodiment, in order to verify the effect of the rolling bearing fault diagnosis method in this paper, the fault characteristics of 9 health states are input into the extreme learning machine for classification experiments. The activation function of the extreme learning machine adopts the Sigmoid function, and the number of neurons is set as: input layer R = 16, hidden layer Q = 30, and output layer F = 1. Fig. 9 is the confusion matrix of an experiment. The horizontal axes 1 to 9 represent the 9 predicted bearing states, the vertical axes 1 to 9 represent the 9 actual bearing faults, and the diagonal line represents the prediction accuracy of each fault. Fig. 9 It can be seen that only one of the 720 fault test samples was misclassified as other types. From the type point of view, the inner ring multi-point fault of label 7 was diagnosed as an inner ring single-point fault. The classification accuracy of this experiment reached 99.86%, verifying the effectiveness of the proposed method.

[0149] In order to further highlight the performance of this method, MHSE is compared with the five algorithms in the feature extraction experiment. Table 4 lists the average accuracy and corresponding standard deviation of each method after 10 experiments.

[0150] Table 4 Average accuracy and corresponding standard deviation of each method after 10 experiments

[0151]

[0152] As shown in Table 4, the average accuracy of MHSE reaches 99.74% and the standard deviation is 0.103. Compared with MHPE, RCMDE, RCMSE, RCMFE and CWMPE, MHSE has the smallest standard deviation and the average accuracy is increased by 3.35%, 3.89%, 6.7%, 3.59% and 3.24% respectively. From the experimental results, the proposed method can effectively distinguish single and compound rolling bearing faults, and has good distinction between faults with different damage degrees. The recognition accuracy and stability are better than the other five algorithms.

[0153] In summary, the present invention proposes a training method, a diagnosis method and a diagnosis device for a bearing fault diagnosis model, which uses empirical Fourier decomposition to preprocess the rolling bearing fault signal, thereby filtering out the noise segment of the Fourier spectrum and highlighting the main vibration component and the impact component of the time series.

[0154] Compared with the coarse-grained sequence method, the improved hierarchical slope entropy algorithm proposed in the present invention can extract more comprehensive and rich time series scale information; at the same time, using the slope to measure the time series difference can perceive the trend and degree of amplitude change, which is more effective than the method of sorting by amplitude size using relative permutation entropy (PE).

[0155] The present invention proposes a bearing fault diagnosis method which can effectively identify bearing faults of different types and damage degrees, with an average recognition rate of 99.74%.

[0156] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope involved in the present application is not limited to the technical solution formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the inventive concept, such as a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

[0157] Except for the technical features described in the specification, the remaining technical features are known technologies to those skilled in the art. In order to highlight the innovative features of the present invention, the remaining technical features will not be described here in detail.

Claims

1. A training method for a bearing fault diagnosis model, characterized in that: include: Collect fault vibration signals of different types of bearing faults to obtain a sample set; Performing empirical Fourier decomposition on the fault vibration signal in the sample set to obtain multiple modal signal components; Screening the plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence; The reconstructed signal sequence is processed using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set; Using the training feature set to train the bearing fault diagnosis model to obtain a trained bearing fault diagnosis model; The process of processing the reconstructed signal sequence using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set includes: Given the reconstruction signal sequence , and calculate the average operator and difference operator : , ,in, represents the reconstructed signal of the i-th interval; Define operators at different levels s ,in, ； Given a specific vector sequence , and determine the f value of each layer according to the node e of the sth layer. The relationship between the node e and the vector sequence is: , where represents the average operator or difference operator of the dth layer; Get the hierarchical components of the reconstructed signal sequence, expressed as: ; Calculate the improved hierarchical slope entropy of each hierarchical component, expressed as: , where m is the embedding dimension, γ and δ are the thresholds of slope entropy.

2. The training method of the bearing fault diagnosis model according to claim 1, characterized in that: The different types of faults include at least normal, inner ring single point fault, outer ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, inner ring multi-point fault, outer ring multi-point fault and ball multi-point fault.

3. The training method of the bearing fault diagnosis model according to claim 1, characterized in that: The step of performing empirical Fourier decomposition on the sample set to obtain multiple modal signal components comprises: Normalizing the Fourier spectrum of the fault vibration signal in the sample set to the interval [0, π], and predefining the number of frequency bands N required to be divided; The initial value and the maximum value in the Fourier spectrum are set as M control points, and the final number of control points T is obtained according to the number of frequency bands N and the number of control points M; The interval is divided according to the final number of control points T, and the position of each control point is , where 1≤n≤T, , ; according to Determine the spectral split boundaries for Fourier decomposition , define the first T boundaries as ,in is the set of the first T boundaries; Determine the continuous interval , and calculate the analytical Fourier natural band function for each interval, where 1, 2, ⋯, T; The real part of the analytical Fourier natural frequency band function of each interval is inversely Fourier transformed to obtain multiple modal signal components.

4. The training method for a bearing fault diagnosis model according to claim 3, characterized in that: Acquiring the final number of control points T according to the number of frequency bands N and the number of control points M includes: When M ≥ N, only the first N segments of M control points in descending order are taken, that is, the final number of control points T is N; When M<N, the frequency bands that can be divided in the vibration signal are less than the predefined number of frequency bands N, and the final number of control points T is M.

5. The training method of the bearing fault diagnosis model according to claim 1, characterized in that: Screening the plurality of modal signal components by kurtosis and correlation coefficient to obtain a reconstructed signal comprises: Calculate the kurtosis value of each modal signal component , and retain the kurtosis value All components greater than 3 to form the set U; Calculate the correlation coefficient between each modal signal component and the fault vibration signal , and the correlation coefficient The standard deviation of is taken as the minimum threshold to screen each of the modal signal components as the main rotation components of the fault vibration signal to form a set V; Take the union of set U and set V and add the signals therein to obtain the reconstructed signal sequence.

6. The training method of the bearing fault diagnosis model according to claim 5, characterized in that: The kurtosis value By formula: Calculated, where is the k-point of the n-th modal signal component; n = 1, 2, ⋯, T; k = 1, 2, ⋯, L; u and σ are The mean and variance of represents the mathematical expectation of n.

7. The training method of the bearing fault diagnosis model according to claim 5, characterized in that: The correlation coefficient By formula: Calculated, among which, represents the average value of the nth modal component, represents the kth sampling point of the fault vibration signal z, Represents the mean of all sampling points of the fault vibration signal z.

8. The training method for a bearing fault diagnosis model according to claim 1, characterized in that: The step of training the bearing fault diagnosis model using the training feature set to obtain a trained bearing fault diagnosis model includes: The training feature set is expressed as , define the weight between the input layer and the hidden layer as φ, the weight between the hidden and output layers as β, and the hidden layer threshold as b, : , , , where R, Q, and F are the number of neurons in the input layer, hidden layer, and output layer of the extreme learning machine respectively; Set an activation function , then the output of the extreme learning machine It is expressed as: , where ; Solve the minimum value of its approximate square error to obtain the weight of the output layer: , and find its optimal solution , where is the generalized inverse of the hidden layer output matrix.

9. A bearing fault diagnosis method, characterized in that: include: Collect bearing fault vibration signals; Performing empirical Fourier decomposition on the fault vibration signal to obtain multiple modal signal components; Screening the plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence; The reconstructed signal sequence is processed using an improved hierarchical slope entropy algorithm to obtain fault feature information; The fault feature information is input into a bearing fault diagnosis model trained by the bearing fault diagnosis model training method described in any one of claims 1 to 8 to obtain a diagnosis result of the bearing fault.

10. A bearing fault diagnosis device, characterized in that: The method for diagnosing bearing faults as claimed in claim 9 is applied, comprising: A fault signal acquisition module is used to collect fault vibration signals of different types of bearing faults to obtain a sample set; A modal signal acquisition module, connected to the fault signal acquisition module, for performing empirical Fourier decomposition on the fault vibration signal in the sample set to obtain multiple modal signal components; A signal reconstruction module, connected to the modal signal acquisition module, for screening a plurality of modal signal components by kurtosis and correlation coefficient respectively, and reconstructing the screening results to obtain a reconstructed signal sequence; A training feature set acquisition module, connected to the signal reconstruction module, for processing the reconstructed signal sequence using an improved hierarchical slope entropy algorithm to obtain fault feature information and form a training feature set; A model building module, connected to the training feature set acquisition module, is used to train the bearing fault diagnosis model using the training feature set to obtain a trained bearing fault diagnosis model; A fault diagnosis module is connected to the model building module and is used to obtain a diagnosis result.

Images