Rolling bearing fault diagnosis method, device, electronic device and storage medium

Abstract

The present application discloses a rolling bearing fault diagnosis method, device, electronic device and storage medium, the method comprising: obtaining multiple vibration signal samples, and setting corresponding fault classification labels, performing variational modal decomposition on the multiple vibration signal samples, determining multiple groups of one-dimensional singular value fault feature vector samples, and forming a training data set; inputting multiple groups of training data sets into an initial BP neural network model based on fractional order gradient, and performing classification training on the initial BP neural network model based on fractional order gradient using a fractional order algorithm, and obtaining a fully trained BP neural network model based on fractional order gradient. The vibration signal is processed by variational modal decomposition, which improves the adaptability of rolling bearing fault diagnosis; in addition, the initial BP neural network model based on fractional order gradient is trained using a fractional order algorithm, which effectively improves the performance of the BP neural network model based on fractional order gradient and improves the accuracy of the diagnosis result.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis method, device, electronic equipment and storage medium. Background Art

[0002] Rolling bearings are precision mechanical components that convert the sliding friction between the rotating shaft and the shaft seat into rolling friction, thereby reducing friction losses. As a key component of rotating machinery, rolling bearings play an important role in the performance, stability and life cycle of the entire machinery.

[0003] At present, although the vibration signal of rolling bearings can be extracted well, the original vibration signal of the bearing contains not only a lot of information about the health of the bearing, but also a lot of noise and interference. Therefore, it is difficult to accurately extract the fault characteristics of the bearing using general time domain or frequency domain signal processing methods, and the variational mode decomposition (VMD) in the time-frequency domain analysis method needs to set parameters before use, which greatly reduces the adaptability and practical value of variational mode decomposition. In addition, the traditional artificial neural network fault classifier has the problems of slow convergence speed and easy to fall into local extreme values. At present, it is mainly processed by gradient descent algorithm, and the accuracy is not high.

[0004] Therefore, there are problems in the prior art that the rolling bearing fault diagnosis has low adaptability and low accuracy of the diagnosis result. Summary of the invention

[0005] In view of this, it is necessary to provide a rolling bearing fault diagnosis method, device, electronic device and storage medium to solve the problems in the prior art that rolling bearing fault diagnosis has low adaptability and low accuracy of diagnosis results.

[0006] In order to solve the above problems, the present invention provides a rolling bearing fault diagnosis method, comprising:

[0007] Obtain multiple vibration signal samples of rolling bearings under different fault conditions, set corresponding fault classification labels, perform variational mode decomposition on the multiple vibration signal samples, determine multiple groups of one-dimensional singular value fault feature vector samples, and form a training data set;

[0008] Input multiple sets of training data sets into the initial BP neural network model based on fractional gradient, take one-dimensional singular value fault feature vector samples as input, take corresponding fault classification labels as output, use fractional algorithm to classify and train the initial BP neural network model based on fractional gradient, and obtain a fully trained BP neural network model based on fractional gradient;

[0009] The vibration signal of the rolling bearing is obtained, and the vibration signal is subjected to variational mode decomposition to determine the one-dimensional singular value fault feature vector. The one-dimensional singular value fault feature vector is input into a well-trained BP neural network model based on fractional gradient to perform rolling bearing fault diagnosis.

[0010] Furthermore, variational mode decomposition is performed on multiple vibration signal samples to determine multiple groups of one-dimensional singular value fault feature vector samples, including:

[0011] Determine the parameters of variational mode decomposition according to the sine-cosine algorithm, wherein the parameters are the number of intrinsic mode vibration signal components and the penalty factor;

[0012] Perform variational mode decomposition on the vibration signal to obtain multiple eigenmode vibration signal components;

[0013] Singular value decomposition is performed on multiple intrinsic mode vibration signal components to determine the one-dimensional singular value fault feature vector of the vibration signal.

[0014] Furthermore, the parameters of variational mode decomposition are determined according to the sine-cosine algorithm, including:

[0015] The number of candidate solutions, the maximum number of iterations, the lower bound of the number of intrinsic mode vibration signal components, the lower bound of the penalty factor and the fitness function are obtained; the parameters of the variational mode decomposition are determined by iterative calculation using the sine and cosine update functions.

[0016] Furthermore, singular value decomposition is performed on multiple intrinsic mode vibration signal components to determine a one-dimensional singular value fault feature vector of the vibration signal, including:

[0017] According to the intrinsic mode vibration signal component, the corresponding initial single-sided spectrum is determined by using Hilbert transform;

[0018] According to the center frequency, the initial single-sided spectrum is adjusted to the baseband to obtain an adjusted single-sided spectrum;

[0019] According to the unilateral spectrum, the frequency band width of the unilateral spectrum is estimated through regularization and Gaussian smoothing, and the variational model is determined according to the frequency band width and constraint conditions;

[0020] Multiple intrinsic mode vibration signal components are input into the variational model to determine the one-dimensional singular value fault feature vector of the vibration signal.

[0021] Furthermore, the initial BP neural network model based on fractional order gradient is classified and trained using a fractional order algorithm to obtain a fully trained BP neural network model based on fractional order gradient, including:

[0022] Input multiple sets of training data sets into the initial fractional gradient-based BP neural network model to obtain the output of each layer node;

[0023] Calculate the fractional gradient of each node according to the error function to update the weight matrix, and set the number of iterations to increase by one;

[0024] The training data set is repeatedly input into the initial BP neural network model based on fractional gradient until the number of iterations reaches a preset value, thereby obtaining a fully trained BP neural network model based on fractional gradient.

[0025] Furthermore, multiple sets of training data sets are input into the initial fractional gradient-based BP neural network model to obtain the output of each layer node, including:

[0026] Obtaining the number of neurons in the input layer, the number of neurons in the hidden layer, and the number of neurons in the output layer of the initial fractional gradient-based BP neural network model, the first connection weight matrix between the input layer and the hidden layer, and the second connection weight matrix between the hidden layer and the input layer;

[0027] According to multiple sets of training data sets, the activation functions of the hidden layer and the output layer are used to obtain the output of the nodes in each layer.

[0028] Furthermore, the weight matrix is ​​updated by calculating the fractional gradient of each node according to the error function, including:

[0029] According to the first connection weight matrix and the second connection weight matrix, using the error function, determining an update formula for the weight connection matrix;

[0030] According to the updating formula of the weight connection matrix, the fractional gradient of each node is determined to update the weight matrix.

[0031] In order to solve the above problems, the present invention provides a rolling bearing fault diagnosis device, comprising:

[0032] A training data acquisition module is used to obtain multiple groups of vibration signal samples of rolling bearings under different fault conditions and set corresponding fault classification labels to form a training data set;

[0033] An initial model building module is used to build an initial BP neural network model based on fractional order gradient, wherein the initial BP neural network model based on fractional order gradient performs variational modal decomposition on the vibration signal to determine a one-dimensional singular value fault feature vector;

[0034] A model training module is used to input multiple groups of vibration signal samples into the initial BP neural network model based on fractional order gradient, output corresponding fault classification labels, and use the fractional order algorithm to perform classification training on the initial BP neural network model based on fractional order gradient to obtain a fully trained BP neural network model based on fractional order gradient;

[0035] The fault diagnosis module is used to perform rolling bearing fault diagnosis using a fully trained BP neural network model based on fractional gradient.

[0036] The present invention also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, a rolling bearing fault diagnosis method as described in any of the above technical solutions is implemented.

[0037] The present invention also provides a computer-readable storage medium, wherein the program medium stores computer program instructions. When the computer program instructions are executed by a computer, a rolling bearing fault diagnosis method as described in any of the above technical solutions is implemented.

[0038] The beneficial effect of adopting the above embodiment is: the present invention provides a rolling bearing fault diagnosis method, including: obtaining multiple vibration signal samples of the rolling bearing under different fault conditions, and setting fault classification labels accordingly, performing variational modal decomposition on the multiple vibration signal samples, and determining multiple groups of one-dimensional singular value fault feature vector samples to form a training data set; inputting the multiple groups of training data sets into an initial BP neural network model based on fractional gradient, taking the one-dimensional singular value fault feature vector samples as input and the corresponding fault classification labels as output, and using a fractional order algorithm to perform classification training on the initial BP neural network model based on fractional gradient to obtain a fully trained BP neural network model based on fractional gradient; obtaining the vibration signal of the rolling bearing, performing variational modal decomposition on the vibration signal, determining the one-dimensional singular value fault feature vector, and inputting the one-dimensional singular value fault feature vector into the fully trained BP neural network model based on fractional gradient to perform rolling bearing fault diagnosis. The vibration signal is processed according to variational modal decomposition to obtain the corresponding one-dimensional singular value fault feature vector sample. Since the parameters in the variational modal decomposition are automatically calculated by the model, the adaptability of rolling bearing fault diagnosis is improved. In addition, in this embodiment, the initial BP neural network model based on fractional order gradient is trained using a fractional order algorithm to obtain a BP neural network model based on fractional order gradient, which effectively improves the performance of the BP neural network model based on fractional order gradient and improves the accuracy of the diagnosis result. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of a flow chart of an embodiment of a rolling bearing fault diagnosis method provided by the present invention;

[0040] Figure 2 A schematic diagram of a flow chart of an embodiment of determining a one-dimensional singular value fault feature vector sample provided by the present invention;

[0041] Figure 3 A schematic diagram of a flow chart of an embodiment of determining a one-dimensional singular value fault feature vector provided by the present invention;

[0042] Figure 4 A schematic diagram of a flow chart of an embodiment of using a fractional order algorithm to train an initial BP neural network model based on a fractional order gradient provided by the present invention;

[0043] Figure 5 A structural block diagram of an embodiment of a rolling bearing fault diagnosis device provided by the present invention;

[0044] Figure 6 This is a structural block diagram of an embodiment of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0046] Before describing the embodiments, the rolling bearings, fault types and fractional order algorithms are described:

[0047] Rolling bearings are generally composed of four parts: an inner ring, an outer ring, rolling elements and a cage. The function of the inner ring is to cooperate with the shaft and rotate with the shaft; the function of the outer ring is to cooperate with the bearing seat and play a supporting role; the rolling elements are evenly distributed between the inner and outer rings with the help of the cage, and the shape, size and number of the cage directly affect the performance and life of the rolling bearing; the cage can evenly distribute the rolling elements, guide the rolling elements to rotate and play a lubricating role.

[0048] Rolling bearing fault types are divided into rolling element, bearing inner ring and bearing outer ring according to the fault location, and divided into 0.1778mm, 0.3556mm and 0.5334mm according to the fault diameter. The fault diagnosis results of rolling bearings include 9 fault states and normal state.

[0049] As the most commonly used optimization algorithm, gradient descent plays an important role in all walks of life.

[0050] Fractional-order algorithm is an algorithm that uses fractional-order calculus to perform calculations. As an extension of integer-order calculus, fractional-order calculus improves the performance of fractional-order algorithm.

[0051] In order to accurately diagnose and identify rolling bearing faults, the present invention provides a rolling bearing fault diagnosis method, device, electronic device and storage medium, which are described in detail below.

[0052] Figure 1 A schematic diagram of a flow chart of an embodiment of a rolling bearing fault diagnosis method provided by the present invention, as shown in FIG. Figure 1 As shown, the rolling bearing fault diagnosis method includes:

[0053] Step S11: obtaining multiple vibration signal samples of the rolling bearing under different fault conditions, and setting corresponding fault classification labels, performing variational modal decomposition on the multiple vibration signal samples, and determining multiple groups of one-dimensional singular value fault feature vector samples to form a training data set;

[0054] Step S12: input multiple sets of training data sets into the initial BP neural network model based on fractional order gradient, take the one-dimensional singular value fault feature vector sample as input, take the corresponding fault classification label as output, use the fractional order algorithm to perform classification training on the initial BP neural network model based on fractional order gradient, and obtain a fully trained BP neural network model based on fractional order gradient;

[0055] Step S13: Acquire the vibration signal of the rolling bearing, perform variational mode decomposition on the vibration signal, determine the one-dimensional singular value fault feature vector, input the one-dimensional singular value fault feature vector into a well-trained BP neural network model based on fractional gradient, and perform rolling bearing fault diagnosis.

[0056] In this embodiment, first, multiple vibration signal samples of the rolling bearing under different fault conditions and their corresponding fault classification labels are obtained; secondly, variational modal decomposition is performed on the multiple vibration signal samples to determine multiple groups of one-dimensional singular value fault feature vector samples, thereby determining a training data set; then, the multiple groups of training data sets are input into an initial BP neural network model based on fractional-order gradient, and the initial BP neural network model based on fractional-order gradient is classified and trained using a fractional-order algorithm to obtain a fully trained BP neural network model based on fractional-order gradient; finally, the real-time vibration signal of the rolling bearing is obtained, and variational modal decomposition is performed on the vibration signal to determine the one-dimensional singular value fault feature vector, and based on the fully trained BP neural network model based on fractional-order gradient, the fault classification label of the rolling bearing is determined.

[0057] In this embodiment, the vibration signal is processed according to variational modal decomposition to obtain the corresponding one-dimensional singular value fault feature vector sample. Since the parameters in the variational modal decomposition are automatically calculated by the model, the adaptability of rolling bearing fault diagnosis is improved. In addition, in this embodiment, the initial BP neural network model based on fractional order gradient is trained using a fractional order algorithm to obtain a BP neural network model based on fractional order gradient, which effectively improves the performance of the BP neural network model based on fractional order gradient and improves the accuracy of the diagnosis results.

[0058] As a preferred embodiment, in step S11, in order to obtain multiple groups of one-dimensional singular value fault feature vector samples according to multiple vibration signal samples, such as Figure 2 As shown, Figure 2A flow chart of an embodiment of determining a one-dimensional singular value fault feature vector sample provided by the present invention includes:

[0059] Step S111: Determine the parameters of variational mode decomposition according to the sine-cosine algorithm, wherein the parameters are the number of intrinsic mode vibration signal components and the penalty factor.

[0060] Step S112: performing variational modal decomposition on the vibration signal to obtain a plurality of intrinsic modal vibration signal components.

[0061] Step S113: performing singular value decomposition on multiple intrinsic mode vibration signal components to determine a one-dimensional singular value fault feature vector of the vibration signal.

[0062] In this embodiment, the parameters of variational modal decomposition are first determined according to the sine-cosine algorithm to obtain the parameters of variational modal decomposition with high adaptability and high reliability, and then the vibration signal is subjected to variational modal decomposition to obtain multiple intrinsic modal vibration signal components, and finally the one-dimensional singular value fault characteristic vector of the vibration signal is determined according to the multiple intrinsic modal vibration signal components. The above method determines the parameters of variational modal decomposition through the sine-cosine algorithm, realizes the automatic adjustment of the parameters of variational modal decomposition, and improves the adaptability of variational modal decomposition.

[0063] As a preferred embodiment, in step S111, first, the number of candidate solutions, the maximum number of iterations, the lower bound of the number of intrinsic modal vibration signal components and the lower bound of the penalty factor are set; secondly, a random solution of the candidate solution is initialized, and a fitness function is set, wherein the fitness function in this embodiment is the envelope spectrum peak factor; then, the position of the solution and the target point and the optimal parameter combination are updated according to the range of the sine and cosine functions; finally, the parameters of the variational modal decomposition are determined according to the envelope spectrum peak factor.

[0064] In a specific embodiment, first, let Among them, k is the lower bound of the number of intrinsic mode vibration signal components, and α is the lower bound of the penalty factor.

[0065] Then, the solution and the position of the target point and the optimal parameter combination are updated according to the range of the sine and cosine functions. The specific expression of the sine and cosine update function is:

[0066]

[0067]

[0068] Where t is the current iteration number, T is the maximum iteration number, a is a constant, r1 is a linear decreasing function of the function value in the interval [0,2], r2 is a random number in the interval [0,2π], r3 is a random number in the interval [0,2], and r4 is a random number in the interval (0,1). is the position of the current solution in the i-th dimension at the t-th iteration, is the position of the target point in the i-th dimension at the t-th iteration.

[0069] In this embodiment, the ratio of the maximum value of the signal envelope spectrum in the range of [n*fr, fs / 2] to the effective value is called the envelope spectrum peak factor, where fr represents the bearing rotation frequency, fs represents the signal sampling frequency, and n is generally taken as 2 to avoid the influence of the rotation frequency on the envelope spectrum peak factor.

[0070] In particular, in general envelope spectra, the closer to the impact frequency, the larger the peak value. The envelope spectrum peak factor takes into account the characteristics of large energy in the resonance frequency band and strong periodicity of the filtered signal. The larger the value of the envelope spectrum peak factor, the higher the signal-to-noise ratio and the greater the energy, and the stronger the periodic impact component.

[0071] As a preferred embodiment, in order to determine the one-dimensional singular value fault feature vector according to the intrinsic mode vibration signal component of the vibration signal, such as Figure 3 As shown, Figure 3 A flow chart of an embodiment of determining a one-dimensional singular value fault feature vector provided by the present invention includes:

[0072] Step S1121: Determine the corresponding initial single-sided spectrum according to the intrinsic mode vibration signal component by using Hilbert transform.

[0073] Step S1122: adjusting the initial single-sided spectrum to the baseband according to the center frequency to obtain an adjusted single-sided spectrum.

[0074] Step S1123: According to the single-sided spectrum, the frequency band width of the single-sided spectrum is estimated through regularization and Gaussian smoothing, and the variational model is determined according to the frequency band width and constraint conditions.

[0075] Step S1124: input multiple intrinsic mode vibration signal components into the variational model to determine a one-dimensional singular value fault feature vector of the vibration signal.

[0076] In a specific embodiment, assuming that the input signal x t is an eigenmode vibration signal component containing different frequency components, then in step S1121, the Hilbert transform is used to obtain the initial single-sided spectrum of each eigenmode vibration signal component, wherein the formula for obtaining the initial single-sided spectrum is:

[0077]

[0078] Where f0 is the initial one-sided spectrum, δ(g) is the Dirichlet function, j is the imaginary factor, * is the convolution symbol, and u kis the eigenmode vibration signal component.

[0079] In step S1122, with a pre-estimated center frequency Multiply them, modulate the initial single-sided spectrum to the baseband, and obtain the modulated single-sided spectrum. The formula for obtaining the modulated single-sided spectrum is:

[0080]

[0081] Among them, f1 is a single-sided spectrum.

[0082] In step S1123, by squaring L 2 Regularization and H 1 Gaussian smoothing estimates the bandwidth of each intrinsic mode vibration signal component, introduces constraints, and finally constructs a constrained variational model:

[0083]

[0084] Among them, Q is the variational model, is the gradient operator, K is the number of eigenmode vibration signal components, {u k}={u1,...,u k} and {ω k}={ω1,...,ω k} are the sets of each eigenmode vibration signal component and the corresponding center frequency.

[0085] Furthermore, a quadratic penalty factor α and Lagrangian multiplier λ are introduced to obtain the augmented Lagrangian function:

[0086]

[0087] In the above way, the minimization problem is transformed into an alternating search for {u k} and {ω k}, and finally {u k},{ω k The update form of} and λ is:

[0088]

[0089]

[0090]

[0091] Where τ is the update factor of a Lagrange multiplier, n is the number of iterations n = {1,...,N}, and Corresponding to ui The Fourier transforms of (t), f(t), and λ(t).

[0092] In this embodiment, through the intrinsic mode vibration signal components, the corresponding one-sided spectrum is obtained; then, according to the constraint conditions, the variational model is determined; finally, according to the variational model, the one-dimensional singular value fault feature vector of the vibration signal is determined. By performing variational mode decomposition on the vibration signal, the frequency-domain dissection of the vibration signal and the effective separation of each vibration signal component are adaptively realized, and the one-dimensional singular value fault feature vector of the vibration signal is determined.

[0093] After obtaining the one-dimensional singular value fault feature vector of the vibration signal, in step S12, it is also necessary to perform classification training on the initial BP neural network model based on the fractional gradient to obtain a trained complete BP neural network model based on the fractional gradient.

[0094] Specifically, in order to improve the detection accuracy of the BP neural network model based on the fractional gradient, in this embodiment, the fractional order algorithm is used to perform classification training on the initial BP neural network model based on the fractional gradient.

[0095] Before stating the method of using the fractional order algorithm to perform classification training on the initial BP neural network model based on the fractional gradient, the fractional gradient descent method is first elaborated:

[0096] For the function h(t) ∈ C n ([0, +∞), R), its Caputo fractional order derivative is defined as follows:[[]]

[0097]

[0098] where t0 and t represent the upper and lower bounds of differentiation, t0 = 0, t ≥ 0; ε represents the order of differentiation, n ∈ N + , n - 1 < ε < n; Γ(·) is the gamma function.

[0099] Performing Taylor series expansion on equation (1), we get:[[]]

[0100]

[0101] where 0 < ε < 1.

[0102] Assuming the objective function is h(t), the iterative process of the gradient descent method is:[[]]

[0103] t k+1 = t k - ρ·h'(t) (3)

[0104] where ρ > 0 represents the iteration step size.

[0105] Replacing the first-order derivative in equation (3) with the fractional-order derivative defined by Caputo, we get the fractional-order gradient descent (FOGD) iterative process:

[0106]

[0107] Among them, 0<ε<1.

[0108] Replace the lower limit of differentiation from the fixed point t0 to t k-1 ,have:

[0109]

[0110]

[0111] Among them, 0<ε<1.

[0112] Keep the first term of the series in equation (6), and the simplified FOGD algorithm is:

[0113]

[0114] Combine the constant Γ(2-ε) with ρ while still using ρ to represent it; replace t in the derivative k-1 Use t k Substitute, and we get:

[0115] t k+1 =t k -ρ·h (1) (t k )(t k -t k-1 ) 1-ε (8)

[0116] Among them, 0<ε<2.

[0117] Give(t k -t k-1 ) 1-ε Adding the absolute value symbol, the final fractional gradient descent algorithm is:

[0118] t k+1 =t k -ρ·h (1) (t k )|t k -t k-1 | 1-ε (9)

[0119] Among them, 0<ε<2.

[0120] In the above fractional gradient descent method, the function h(t) is iterated multiple times to obtain the final fractional gradient descent algorithm.

[0121] Since the fractional-order gradient descent method uses fractional-order derivatives for iterative processing, compared with the commonly used integer-order gradient descent method, the fractional-order gradient descent method has a wider processing range and higher accuracy of the results.

[0122] In step S12, in order to obtain a fully trained BP neural network model based on fractional gradient, Figure 4 As shown, Figure 4 The flowchart of an embodiment of using a fractional order algorithm to train an initial BP neural network model based on a fractional order gradient provided by the present invention includes:

[0123] Step S121: input the training data set into the initial fractional gradient-based BP neural network model to obtain the output of each layer node.

[0124] Step S122: Calculate the fractional gradient of each node according to the error function to update the weight matrix, and set the number of iterations to be plus one.

[0125] Step S123: repeatedly inputting vibration signal samples into the initial fractional-order gradient-based BP neural network model until the number of iterations reaches a preset value, thereby obtaining a fully trained fractional-order gradient-based BP neural network model.

[0126] In this embodiment, the fractional order gradient of each node is calculated by the error function to update the weight matrix, and then the parameters in the initial BP neural network model based on the fractional order gradient are adjusted through multiple iterations, and finally a fully trained BP neural network model based on the fractional order gradient is obtained. Since the weight matrix updated by the fractional order gradient can reflect the deviation of the parameters in the initial BP neural network model based on the fractional order gradient, and the accuracy is relatively high, therefore, the BP neural network model based on the fractional order gradient with the initial training based on the fractional order gradient can be obtained with a relatively high judgment accuracy.

[0127] Furthermore, in step S121, multiple groups of vibration signal samples are input into the initial fractional gradient-based BP neural network model to obtain the output of each layer node, including:

[0128] Obtaining the number of neurons in the input layer, the number of neurons in the hidden layer, and the number of neurons in the output layer of the initial fractional gradient-based BP neural network model, the first connection weight matrix between the input layer and the hidden layer, and the second connection weight matrix between the hidden layer and the input layer;

[0129] According to multiple groups of vibration signal samples, the output of each layer node is obtained by using the activation function of the hidden layer and the output layer.

[0130] In a specific embodiment, the number of neurons in the input layer, hidden layer and output layer of the initial fractional gradient-based BP neural network model is set to d, q, and l respectively; the training sample set is x i and i is the input of the i-th sample and the corresponding ideal output; the input is described by d attributes, and the output is an l-dimensional real-valued vector; the first connection weight matrix between the input layer and the hidden layer is defined as {w nh ,n=1,2,...,d,h=1,2,...,l}, the second connection weight matrix between the hidden layer and the input layer is defined as {v hj ,h=1,2,...,q,j=1,2,...,l}; the activation functions of the hidden layer and the output layer are f,g:R→R respectively; the input of the hidden layer is and output Calculated by formulas (10) and (11), the input of the output layer is and output Calculated by formula (12) and (13) respectively:

[0131]

[0132]

[0133]

[0134]

[0135] In the above embodiment, the output of each layer node is obtained by using the activation functions of the hidden layer and the output layer according to the training sample set.

[0136] In step S122, the weight matrix is ​​updated by calculating the fractional gradient of each node according to the error function, including:

[0137] According to the first connection weight matrix and the second connection weight matrix, using the error function, determining an update formula for the weight connection matrix;

[0138] According to the updating formula of the weight connection matrix, the fractional gradient of each node is determined to update the weight matrix.

[0139] In step S123, after obtaining the output of each layer node, back propagation is required to adjust the parameters of the initial BP neural network model based on fractional gradient, including:

[0140] Calculate the fractional gradient of each node according to the error function to update the weight matrix, and set the number of iterations to increase by one;

[0141] In a specific embodiment, the mean square error function E(w,v) used to evaluate the neural network is:

[0142]

[0143] in, is the ideal output of the i-th sample, assuming Then formula (14) can be rewritten as:

[0144]

[0145] For any initial connection weight matrix and The update formula of the kth weight connection matrix based on the fractional gradient descent method is as follows:

[0146]

[0147]

[0148] Among them, η is the neural network learning rate.

[0149] According to the definition of Caputo fractional derivative, the fractional derivative of a composite function is expressed as the product of integer differential and fractional differential. Assume that h(s(t)) is a composite function, its ε-order derivative with respect to t is:

[0150]

[0151] According to formula (18), it can be deduced that:

[0152]

[0153]

[0154] in,

[0155]

[0156]

[0157]

[0158]

[0159] Finally, the update formula of the k-th weight connection matrix of the BP neural network based on fractional order optimization is as follows:

[0160]

[0161]

[0162] In the above manner, the initial BP neural network model based on fractional order gradient is trained using the fractional order algorithm. Since the fractional order algorithm can more accurately obtain the parameter values ​​in the fractional order gradient update weight matrix during the processing process, the performance of the final BP neural network model based on fractional order gradient can be enhanced, and the judgment accuracy can be improved.

[0163] In order to solve the above problems, the present invention also provides a rolling bearing fault diagnosis device, such as Figure 5 As shown, Figure 5 This is a structural block diagram of an embodiment of a rolling bearing fault diagnosis device provided by the present invention. The rolling bearing fault diagnosis device 500 includes:

[0164] The training data acquisition module 501 is used to acquire multiple groups of vibration signal samples of rolling bearings under different fault conditions, and set corresponding fault classification labels to form a training data set;

[0165] An initial model building module 502 is used to build an initial BP neural network model based on fractional order gradient, wherein the initial BP neural network model based on fractional order gradient performs variational modal decomposition on the vibration signal to determine a one-dimensional singular value fault feature vector;

[0166] The model training module 503 is used to input multiple groups of vibration signal samples into the initial BP neural network model based on fractional order gradient, output corresponding fault classification labels, and perform classification training on the initial BP neural network model based on fractional order gradient using a fractional order algorithm to obtain a fully trained BP neural network model based on fractional order gradient;

[0167] The fault diagnosis module 504 is used to perform rolling bearing fault diagnosis using a fully trained BP neural network model based on fractional order gradient.

[0168] The present invention also provides an electronic device, such as Figure 6 As shown, Figure 6 This is a structural block diagram of an embodiment of an electronic device provided by the present invention. The electronic device 600 may be a computing device such as a mobile terminal, a desktop computer, a notebook, a palm computer, and a server. The electronic device 600 includes a processor 601 and a memory 602, wherein the memory 602 stores a rolling bearing fault diagnosis program 603.

[0169] In some embodiments, the memory 602 may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory 602 may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Further, the memory 602 may also include both an internal storage unit of the computer device and an external storage device. The memory 602 is used to store application software and various types of data installed in the computer device, such as program codes for installing the computer device. The memory 602 may also be used to temporarily store data that has been output or is to be output. In one embodiment, the rolling bearing fault diagnosis program 603 may be executed by the processor 601, thereby realizing the rolling bearing fault diagnosis method of each embodiment of the present invention.

[0170] In some embodiments, the processor 601 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run program codes or process data stored in the memory 602, such as executing a rolling bearing fault diagnosis program.

[0171] This embodiment also provides a computer-readable storage medium, on which a rolling bearing fault diagnosis program is stored. When the program is executed by a processor of a computer, the rolling bearing fault diagnosis method described above is implemented.

[0172] The technical solution in this application is used to classify and detect bearing faults, which not only improves the adaptability of the BP neural network model based on fractional gradient, but also improves the accuracy of fault diagnosis results. Experimental analysis was conducted using the open benchmark data set for bearings provided by Case Western Reserve University in the United States, and the accuracy of bearing fault diagnosis by this technical solution reached 97.5%.

[0173] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0174] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method, characterized in that: include: Acquire multiple vibration signal samples of the rolling bearing under different fault conditions, set fault classification labels accordingly, perform variational modal decomposition on the multiple vibration signal samples, determine multiple groups of one-dimensional singular value fault feature vector samples, and form a training data set; Inputting the multiple sets of training data sets into the initial BP neural network model based on fractional gradient, taking the one-dimensional singular value fault feature vector sample as input and the corresponding fault classification label as output, and using the fractional algorithm to perform classification training on the initial BP neural network model based on fractional gradient to obtain a fully trained BP neural network model based on fractional gradient; Acquire the vibration signal of the rolling bearing, perform the variational modal decomposition on the vibration signal, determine a one-dimensional singular value fault feature vector, input the one-dimensional singular value fault feature vector into the fully trained BP neural network model based on fractional gradient, and perform rolling bearing fault diagnosis; The method of using a fractional order algorithm to classify and train the initial BP neural network model based on fractional order gradient to obtain a fully trained BP neural network model based on fractional order gradient includes: Inputting the plurality of training data sets into the initial BP neural network model based on fractional gradient to obtain the output of each layer node; Calculate the fractional gradient of each node according to the error function to update the weight matrix, and set the number of iterations to increase by one; Repeatedly inputting the training data set into the initial BP neural network model based on fractional gradient until the number of iterations reaches a preset value, thereby obtaining a fully trained BP neural network model based on fractional gradient; The step of inputting the plurality of training data sets into the initial BP neural network model based on fractional gradient to obtain the output of each layer node includes: Obtaining the number of neurons in the input layer, the number of neurons in the hidden layer, and the number of neurons in the output layer of the initial fractional gradient-based BP neural network model, a first connection weight matrix between the input layer and the hidden layer, and a second connection weight matrix between the hidden layer and the input layer; According to the plurality of training data sets, the output of the nodes of each layer is obtained by using the activation functions of the hidden layer and the output layer; The step of calculating the fractional gradient of each node according to the error function to update the weight matrix includes: According to the first connection weight matrix and the second connection weight matrix, using the error function, determining an update formula for the weight connection matrix; Determine the fractional gradient update weight matrix of each node according to the update formula of the weight connection matrix; In addition, the final BP neural network based on fractional order optimization The update formula of the secondary weight connection matrix is ​​as follows: 。 2. The rolling bearing fault diagnosis method according to claim 1, characterized in that: The performing variational mode decomposition on the plurality of vibration signal samples to determine a plurality of groups of one-dimensional singular value fault feature vector samples comprises: Determining the parameters of the variational modal decomposition according to the sine-cosine algorithm, wherein the parameters are the number of intrinsic modal vibration signal components and a penalty factor; Performing the variational modal decomposition on the vibration signal to obtain a plurality of eigenmode vibration signal components; Singular value decomposition is performed on the multiple intrinsic mode vibration signal components to determine a one-dimensional singular value fault feature vector of the vibration signal.

3. The rolling bearing fault diagnosis method according to claim 2, characterized in that: The step of determining the parameters of the variational mode decomposition according to the sine-cosine algorithm comprises: The number of candidate solutions, the maximum number of iterations, the lower bound of the number of the intrinsic mode vibration signal components, the lower bound of the penalty factor and the fitness function are obtained; and the parameters of the variational mode decomposition are determined by iterative calculation using the sine-cosine update function.

4. The rolling bearing fault diagnosis method according to claim 2, characterized in that: The performing singular value decomposition on the multiple intrinsic mode vibration signal components to determine a one-dimensional singular value fault feature vector of the vibration signal includes: According to the intrinsic mode vibration signal component, using Hilbert transform, determine the corresponding initial single-sided spectrum; According to the center frequency, the initial single-sided spectrum is adjusted to the baseband to obtain an adjusted single-sided spectrum; According to the unilateral spectrum, the frequency band width of the unilateral spectrum is estimated by regularization and Gaussian smoothing, and a variational model is determined according to the frequency band width and constraint conditions; The multiple intrinsic mode vibration signal components are input into the variational model to determine a one-dimensional singular value fault feature vector of the vibration signal.

5. A rolling bearing fault diagnosis device, characterized in that: include: A training data acquisition module is used to obtain multiple groups of vibration signal samples of rolling bearings under different fault conditions and set corresponding fault classification labels to form a training data set; An initial model building module is used to build an initial BP neural network model based on fractional order gradient, wherein the initial BP neural network model based on fractional order gradient performs variational modal decomposition on the vibration signal to determine a one-dimensional singular value fault feature vector; A model training module, used for inputting the plurality of groups of vibration signal samples into the initial BP neural network model based on fractional order gradient, outputting the corresponding fault classification labels, and performing classification training on the initial BP neural network model based on fractional order gradient using a fractional order algorithm to obtain a fully trained BP neural network model based on fractional order gradient; A fault diagnosis module, used for performing rolling bearing fault diagnosis using the fully trained BP neural network model based on fractional gradient; The method of using a fractional order algorithm to classify and train the initial BP neural network model based on fractional order gradient to obtain a fully trained BP neural network model based on fractional order gradient includes: Inputting the plurality of training data sets into the initial BP neural network model based on fractional gradient to obtain the output of each layer node; Calculate the fractional gradient of each node according to the error function to update the weight matrix, and set the number of iterations to increase by one; Repeatedly inputting the training data set into the initial BP neural network model based on fractional gradient until the number of iterations reaches a preset value, thereby obtaining a fully trained BP neural network model based on fractional gradient; The step of inputting the plurality of training data sets into the initial BP neural network model based on fractional gradient to obtain the output of each layer node includes: Obtaining the number of neurons in the input layer, the number of neurons in the hidden layer, and the number of neurons in the output layer of the initial fractional gradient-based BP neural network model, a first connection weight matrix between the input layer and the hidden layer, and a second connection weight matrix between the hidden layer and the input layer; According to the plurality of training data sets, the output of the nodes of each layer is obtained by using the activation functions of the hidden layer and the output layer; The step of calculating the fractional gradient of each node according to the error function to update the weight matrix includes: According to the first connection weight matrix and the second connection weight matrix, using the error function, determining an update formula for the weight connection matrix; Determine the fractional gradient update weight matrix of each node according to the update formula of the weight connection matrix; In addition, the final BP neural network based on fractional order optimization The update formula of the secondary weight connection matrix is ​​as follows: 。 6. An electronic device, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the rolling bearing fault diagnosis method according to any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium, characterized in that: The program medium stores computer program instructions, and when the computer program instructions are executed by a computer, the computer is caused to execute the rolling bearing fault diagnosis method according to any one of claims 1 to 4.

Images