Fault diagnosis method and system for rotating machinery
Through the fault diagnosis method combined with sparrow search optimization algorithm and one-dimensional convolutional neural network, the problem of rolling bearing fault diagnosis under heavy noise conditions is solved, and efficient fault recognition and stable noise suppression effect are achieved.
Patent Information
- Application Number
- CN202311314862.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-10
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-10-10
AI Technical Summary
The prior art is difficult to effectively diagnose rolling bearing failures under heavy noise conditions. Traditional signal processing methods and artificial intelligence methods are not well recognized in the face of massive sample data and noisy signals.
The sparrow search optimization algorithm is used to optimize the modal number and filter length of the feature modal decomposition algorithm, combined with one-dimensional convolutional neural network and gated loop unit, and used the self-attention mechanism to build a fault diagnosis model, eliminate noise and capture signal timing relationships .
The effectiveness and stability of rolling bearing fault diagnosis under heavy noise conditions is achieved, and the problem of uncertainty in parameter selection and insufficient timing relationship of feature extraction in traditional methods is overcome, and it has good noise reduction effect and generalization ability.
Smart Images

Figure CN117312956B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of fault diagnosis of rotating machinery under heavy noise conditions, and in particular to a fault diagnosis method and system for rotating machinery. Background Art
[0002] Rolling bearings are one of the key components of large-scale automated mechanical equipment. Since rolling bearings work in a harsh environment of heavy load, high pressure, and high-speed rotation for a long time, they are prone to various defects such as wear, corrosion, and fracture. This may cause the mechanical equipment to fail to operate, or even cause serious safety accidents. According to statistical data analysis, the failure rate of mechanical equipment caused by bearing failure is about 41%. Therefore, in order to ensure the safe operation of automated mechanical systems and reduce manual maintenance costs, it is of great significance to achieve more accurate diagnosis of rolling bearings.
[0003] In practical applications, since rolling bearings work in complex environments, the collected bearing vibration signals are highly non-stationary and nonlinear, and are prone to contain interference signals. Therefore, it is very necessary to preprocess the collected vibration signals. The main purpose of preprocessing is to separate the interference signals from the actual bearing vibration signals, so that the extracted vibration signals contain no noise signals as much as possible.
[0004] Common signal preprocessing methods include empirical mode decomposition (EMD), local mean decomposition (LMD), wavelet threshold denoising (WTD) and variational mode decomposition (VMD). Although EMD and LMD are adaptive signal processing methods, their applications are limited due to the existence of mode mixing and endpoint effects. Although WTD can achieve noise reduction, the noise reduction effect depends on the reasonable choice of threshold. VMD has strong signal decomposition capabilities and has been widely used in fault diagnosis.
[0005] However, in the current VMD analysis process, the mode components containing fault information are always idealized as narrowband components. In other words, the VMD-based method needs to rely on certain prior fault knowledge, which has certain obstacles to its practical application.
[0006] Intelligent fault diagnosis of rolling bearings is widely considered to be a pattern recognition problem. Common pattern recognition algorithms include support vector machines (SVM), K-nearest neighbor algorithms (KNN), decision trees (DT), support matrix machines (SMM) and artificial neural networks. These algorithms can accurately identify the health status of bearings, but they require a lot of time to manually extract useful features. In addition, when the above methods are faced with massive sample data and noisy vibration signals, they cannot achieve satisfactory recognition accuracy and efficiency. Therefore, it is difficult to effectively diagnose bearing faults under heavy noise conditions by relying solely on traditional signal processing methods or artificial intelligence methods. Summary of the invention
[0007] The present application provides a fault diagnosis method and system for a rotating machinery to solve the technical problem that it is difficult to effectively diagnose bearing faults under heavy noise conditions by relying on conventional signal processing methods or artificial intelligence methods.
[0008] The present application provides a fault diagnosis method for a rotating machine, comprising:
[0009] Collect the original vibration signals of rotating machinery with different fault types;
[0010] Based on the sparrow search optimization algorithm, the mode number and filter length of the characteristic mode decomposition algorithm are optimized, and the global minimum envelope entropy value and the corresponding optimal population parameter combination are obtained;
[0011] Decompose the original vibration signal using the characteristic mode decomposition algorithm after parameter optimization and obtain more than two modal components;
[0012] Calculate the kurtosis values of more than two modal components, and select the modal component with the largest kurtosis value as the optimal mode;
[0013] The optimal mode is input into a one-dimensional convolutional neural network, and the one-dimensional convolutional neural network is used to eliminate some noise and suppress overfitting;
[0014] Using gated recurrent units to capture the temporal relationship of signal sequences;
[0015] The self-attention mechanism is used to calculate the attention score corresponding to the output of the gated recurrent unit and a fault diagnosis model is constructed;
[0016] A vibration signal is obtained and input into the fault diagnosis model, and a corresponding fault type is output.
[0017] Optionally, the step of optimizing the mode number and filter length of the characteristic mode decomposition algorithm based on the sparrow search optimization algorithm and obtaining the global minimum envelope entropy value and the corresponding optimal population parameter combination includes the following steps:
[0018] Initialize population parameters;
[0019] Calculate the envelope entropy value and sort the steps: Decompose the original vibration signal using the characteristic mode decomposition algorithm at different sparrow positions, calculate the envelope entropy value corresponding to the current position, compare the sizes of each envelope entropy value to determine the current minimum envelope entropy value, and update the individual extreme value and the global extreme value;
[0020] The envelope entropy value of the original vibration signal x(j) (j=0, 1, ..., n) is calculated by the following formula:
[0021]
[0022] Where n represents the total number of sampling points, x(j) represents the jth sampling point of the signal, and E p is the envelope entropy of signal x(j), a(j) is the envelope signal obtained after Hilbert transform of x(j), and p j is the normalized form of a(j);
[0023] The formula for updating the discoverer's position is:
[0024]
[0025] Among them, t is the current iteration number, R2∈[0,1] is the warning value, R ST ∈[0.5, 1], Q is a random number that follows a normal distribution;
[0026] Joiner position update formula:
[0027]
[0028] Among them, w i,j t The worst position searched for the current population; b i,j t is the optimal position of the sparrows in the current population; Q is a random number that obeys the normal distribution; x i,j t is the current joiner position; D is the dimension of the search space;
[0029] The update formula of the sentinel position is:
[0030]
[0031] Among them, β is a random number that obeys the normal distribution; k∈[-1,1] is a uniform random number; f i is the fitness of the current sparrow individual; f g is the fitness of the global optimal position; b i,j t is the optimal position of the current population of sparrows; x i,j t is the current alerter's position;
[0032] Determine whether danger occurs. If so, update the discoverer position and the follower position; if not, update the envelope entropy value;
[0033] Determine whether the number of cycles reaches the preset number of cycles. If not, return to the step of calculating the envelope entropy value and sorting to cycle. If so, output the global minimum envelope entropy value and the corresponding optimal population parameter combination.
[0034] Optionally, the step of decomposing the original vibration signal using the characteristic mode decomposition algorithm after parameter optimization to obtain more than two modal components specifically includes the following steps:
[0035] Inputting the original vibration signal x into the characteristic mode decomposition algorithm, and initializing the parameters of the characteristic mode decomposition algorithm, the parameters including the number of decomposed modes n and the filter length L;
[0036] Initializing the FIR filter bank by using K Hanning windows and initializing the number of iterations i=1, wherein the eigenmode decomposition algorithm updates the filter coefficients by iteration to establish an adaptive FIR filter bank;
[0037] Get the filtered modal component u k i ;
[0038] Using the original vibration signal x and the decomposed modal components u k i and the estimated failure period T k i Update the filter coefficients, where the estimated fault period T k i is composed of the modal component u k i The autocorrelation spectrum reaches the local maximum R after passing through the zero point k i The corresponding moment, and let i = i + 1;
[0039] Determine whether the number of iterations is greater than or equal to the preset number of iterations. If not, return to the step of obtaining the filtered modal components. If so, calculate the correlation coefficient between each two modal components and construct the correlation coefficient matrix CC. (K×K) ,
[0040]
[0041] Where N represents the signal length; CC pq represents the modal component u p and u q The correlation matrix of and u p and u q The average value of
[0042] Select the two modal components with the largest correlation coefficients and calculate the correlation kurtosis of the two modal components;
[0043] Select the modal component with large correlation kurtosis as the decomposed modal component, and set K = K-1, where K is the number of modes;
[0044] Determine whether the current mode number K is greater than or equal to the set mode number n. If not, return to the step of obtaining the filtered modal components. If so, obtain n modal components as the final decomposition result of the characteristic mode decomposition algorithm.
[0045] Optionally, in the step of calculating the kurtosis values of two or more modal components and selecting the mode with the maximum kurtosis value as the optimal mode,
[0046] Calculate the kurtosis value of each modal component after decomposition by the eigenmode decomposition algorithm, and select the mode with the largest kurtosis value as the optimal mode of the eigenmode decomposition algorithm;
[0047] Calculate the kurtosis value of the signal sequence {x(n)}:
[0048]
[0049] Among them, x i is the ith value of the signal sequence, is the average value of the signal sequence, and n is the signal length.
[0050] Optionally, in the step of inputting the optimal mode into a one-dimensional convolutional neural network and using the convolution layer of the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting,
[0051] The one-dimensional convolutional neural network includes a convolution layer, a pooling layer, a fully connected layer and a classification layer.
[0052] Optionally, in the step of inputting the optimal mode into a one-dimensional convolutional neural network and using the convolution layer of the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting,
[0053] The optimal mode is input into a one-dimensional convolutional neural network, and the optimal mode is processed by filter convolution;
[0054] The activation function is used to generate a feature matrix, and each convolution unit corresponds to a part of the receptive field, where the receptive field refers to the area covered by the convolution unit in the input data;
[0055] The convolution operation process is shown in the following formula:
[0056]
[0057] Among them, ReLU(·) and * represent the activation function and convolution operation; x l (j) is defined as the jth local region in convolutional layer l; and are the weight and bias of the i-th convolution kernel respectively. Represents the output features after convolution operation and Relu activation.
[0058] Optionally, in the step of inputting the optimal mode into a one-dimensional convolutional neural network and using the convolution layer of the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting,
[0059] The formula for maximum pooling is as follows:
[0060]
[0061] in, represents the value of the t-th neuron in the ith feature at layer l, (t∈[(j-1)W+1,jW]). W is the width of the pooling region; It represents the output value of the i-th feature of the l+1 layer;
[0062] The fully connected layer formula is as follows:
[0063]
[0064] in, and are weights and biases; represents the value of the tth neuron in the ith feature of the lth layer, m represents the total number of features in the lth layer, and n represents the total number of neurons in the ith feature of the lth layer.
[0065] Optionally, in the step of using a gated recurrent unit to capture the timing relationship of a signal sequence,
[0066] u t =f(W u ·[a t ,h t-1 ])
[0067] r t =f(W r ·[a t ,h t-1 ])
[0068]
[0069]
[0070] Among them, a t is the model input at that moment, h t-1 is the hidden state at the previous moment, r t and u t They are respectively reset gate output result and update gate output result at that moment, is a candidate hidden state, used to describe the current input state and update gate state, h t is a hidden state, if u tIf the value of is close to 0, the hidden state will pay more attention to the information before time t, otherwise it will be close to the candidate hidden state. f is the activation function, W u , W r , W are the weight coefficients of the reset gate, update gate and candidate output set respectively, and ⊙ represents the product between matrices.
[0071] Optionally, in the step of using the self-attention mechanism to calculate the attention score corresponding to the output of the gated recurrent unit and constructing a fault diagnosis model,
[0072] For an input sequence x = {x1, x2, ..., x n}, x i ∈R k , firstly, through adaptive mapping, we obtain three vectors
[0073] q i =W q x i , k i =W k x i , v i =W v x i
[0074] is a parameter matrix;
[0075] Calculate the vector q i and vector k i The similarity between them is used to obtain an attention score α i,j , the calculation formula is:
[0076]
[0077] Among them, d k Represents k j Dimensions;
[0078] Use the softmax function to i,j进 Row normalization to get weight coefficient The calculation formula is:
[0079]
[0080] Among them, α i,j is the attention score, n represents the number of input vectors;
[0081] The vector v i Multiply and sum with the corresponding weight coefficient to get the attention value, and the calculation formula is:
[0082]
[0083] Among them, z i is x i The attention of x can simultaneously calculate the attention value corresponding to each value in x to form a new feature sequence z = {z1, z2, ..., z n}.
[0084] Correspondingly, the present application also provides a fault diagnosis system for a rotating machinery, comprising a memory and a processor, the memory being used to store executable program codes; the processor being connected to the memory, and running a computer program corresponding to the executable program codes by reading the executable program codes, so as to execute the steps in any of the above-mentioned fault diagnosis methods for rotating machinery.
[0085] The present application provides a method and system for fault diagnosis of rotating machinery. The method for fault diagnosis of rotating machinery can effectively diagnose faults of rotating machinery under heavy noise conditions, and has good stability, strong noise suppression ability, and certain generalization ability. This advantage mainly overcomes the uncertainty of artificially selecting the length and modal number of FMD filters, uses the sparrow search optimization algorithm to search for the optimal parameters of FMD, realizes adaptive FMD filtering, and has good noise reduction effect; and combines the one-dimensional convolutional network with the GRU and self-attention mechanism to maximize the temporal relationship in the signal, and retain important information as much as possible in training, overcoming the problem of insufficient feature extraction and temporal relationship utilization of previous deep learning methods. Under FMD denoising and neural network feature extraction, the algorithm can more effectively diagnose faults of rotating machinery under heavy noise conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0086] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0087] Figure 1 It is a flow chart of a fault diagnosis method for a rotating machine provided by the present application;
[0088] Figure 2 is a flow chart of step S200 in the fault diagnosis method of rotating machinery provided by the present application;
[0089] Figure 3 is a flow chart of step S300 in the fault diagnosis method of rotating machinery provided by the present application;
[0090] Figure 4It is a comparison diagram of parameter optimization of characteristic mode decomposition algorithm in the fault diagnosis method of rotating machinery provided by the present application;
[0091] Figure 5 It is the envelope spectrum after decomposition by the characteristic mode decomposition algorithm in the fault diagnosis method of rotating machinery provided by the present application;
[0092] Figure 6 It is the confusion matrix under -10db noise condition in the fault diagnosis method of rotating machinery provided in this application. DETAILED DESCRIPTION
[0093] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. In addition, it should be understood that the specific implementation methods described herein are only used to illustrate and explain the present application and are not used to limit the present application. In the present application, unless otherwise stated, the directional words used, such as "up", "down", "left", and "right", generally refer to the up, down, left, and right of the device in actual use or working state, specifically the drawing direction in the accompanying drawings.
[0094] The present application provides a method and system for fault diagnosis of a rotating machine, which are described in detail below. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments of the present application. In the following embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.
[0095] See also Figure 1-Figure 6 , the present application provides a method for fault diagnosis of rotating machinery, firstly, the parameters of the characteristic mode decomposition algorithm (FMD) are optimized by using the sparrow search optimization algorithm (SSA). The signal is decomposed by using the optimal parameters for noise reduction; finally, a neural network combining convolution, gated recurrent unit (GRU) and self-attention mechanism is used to classify the high-noise signal. This method utilizes the characteristic mode decomposition algorithm (FMD) to take into account the pulse characteristics and periodicity of the fault signal, has good robustness to noise and interference, can adaptively extract feature information from the vibration signal, and can accurately and efficiently diagnose bearing faults under heavy noise.
[0096] See also Figure 1-Figure 6 , the fault diagnosis method of rotating machinery specifically includes the following steps:
[0097] S100, collecting original vibration signals of rotating machinery of different fault types;
[0098] Set the rotation speed and sampling frequency of the bearing vibration signal acquisition test bench, and sample the original vibration signals of different bearing fault types;
[0099] S200, optimizing the mode number and filter length of the characteristic mode decomposition algorithm based on the sparrow search optimization algorithm, and obtaining the global minimum envelope entropy value and the corresponding optimal population parameter combination;
[0100] The sparrow search optimization algorithm is inspired by the foraging and anti-predation behaviors of sparrows. The algorithm has the advantages of strong optimization ability and fast convergence speed. When using the sparrow search optimization algorithm to optimize the number of modes and filter length L of the characteristic mode decomposition algorithm, the key point is the selection and calling of the fitness function. This method selects the minimum envelope entropy value as the fitness function of the sparrow search optimization algorithm. The size of the envelope entropy value reflects the sparsity of the original signal. The stronger the sparsity, the more obvious the periodic shock characteristics related to the fault, and the smaller the envelope entropy value.
[0101] Step S200 specifically includes the following steps:
[0102] S210, initializing population parameters;
[0103] In the D-dimensional search space, the sparrow population is X, and the number of sparrows in the population is n, that is, X = [x1, x2, ..., xi], i = 1, 2, ... n. Then the search position of the i-th sparrow in the population can be expressed as XD = [xi, 1, xi, 2, ..., xi, d] in the D-dimensional space. The envelope entropy value is selected as the fitness function. The sparrow with high fitness is called the discoverer, and other sparrows with low fitness are called followers. Set the parameters of the sparrow search optimization algorithm, including the upper and lower limits of the optimization, the population size and the maximum number of iterations, and the ratio of discoverers and followers, to initialize the parameters to update the position and speed of the sparrows.
[0104] S220, calculating and sorting envelope entropy values: decomposing the original vibration signal at different sparrow positions using the characteristic mode decomposition algorithm, calculating the envelope entropy value corresponding to the current position, comparing the sizes of the envelope entropy values to determine the current minimum envelope entropy value, and updating the individual extreme value and the global extreme value;
[0105] The envelope entropy value of the original vibration signal x(j) (j=0, 1, ..., n) is calculated by the following formula:
[0106]
[0107] Where n represents the total number of sampling points, x(j) represents the jth sampling point of the signal, and E p is the envelope entropy of signal x(j), a(j) is the envelope signal obtained after Hilbert transform of x(j), and pj is the normalized form of a(j);
[0108] S230, the discoverer position update formula is:
[0109]
[0110] Among them, t is the current iteration number, R2∈[0,1] is the warning value, R ST ∈[0.5, 1], Q is a random number that follows a normal distribution;
[0111] S240, joiner position update formula:
[0112]
[0113] Among them, w i,j t The worst position searched for the current population; b i,j t is the optimal position of the sparrows in the current population; Q is a random number that obeys the normal distribution; x i,j t is the current joiner position; D is the dimension of the search space;
[0114] S250, the update formula of the position of the sentinel:
[0115]
[0116] Among them, β is a random number that obeys the normal distribution; k∈[-1,1] is a uniform random number; f i is the fitness of the current sparrow individual; f g is the fitness of the global optimal position; b i,j t is the optimal position of the current population of sparrows; x i,j t is the current alerter's position;
[0117] S260, determining whether danger occurs, if so, updating the discoverer position and the follower position; if not, updating the envelope entropy value;
[0118] S270, determining whether the number of cycles reaches the preset number of cycles, if not, returning to the step of calculating the envelope entropy value and sorting to cycle, if so, outputting the global minimum envelope entropy value and the corresponding optimal population parameter combination;
[0119] In order to prevent the problem that the decomposition effect of the eigenmode decomposition algorithm is poor due to improper artificial parameter selection, resulting in unclear denoising effect, the present application adopts the sparrow search optimization algorithm to adaptively search for the optimal parameters of the eigenmode decomposition algorithm, making full use of the filtering ability of the eigenmode decomposition algorithm in vibration signals and its robustness to noise.
[0120] S300, decomposing the original vibration signal using the characteristic mode decomposition algorithm after parameter optimization, and obtaining more than two modal components;
[0121] Eigenmode decomposition establishes an adaptive FIR filter bank by iteratively updating the filter coefficients, making the filtered signal infinitely close to the target function of the correlation kurtosis of the deconvolution, and using the autocorrelation function to estimate the fault period of the measured signal, thereby decomposing the non-stationary multi-component signal into several modal components. Therefore, the eigenmode decomposition algorithm not only considers the periodicity and impact of the signal at the same time, but also has a certain robustness to interference signals and noise components, and gets rid of the dependence on the prior knowledge of the signal fault period.
[0122] Step S300 specifically includes the following steps:
[0123] S310, inputting the original vibration signal x into the characteristic mode decomposition algorithm, and initializing the parameters of the characteristic mode decomposition algorithm, the parameters including the number of decomposed modes n and the filter length L;
[0124] S320, initializing the FIR filter bank by using K Hanning windows and initializing the number of iterations i=1, wherein the eigenmode decomposition algorithm updates the filter coefficients by iteration to establish an adaptive FIR filter bank;
[0125] S330, obtaining the filtered modal component u k i ;
[0126] S340, using the original vibration signal x, decomposed modal component u k i and the estimated failure period T k i Update the filter coefficients, where the estimated fault period T k i is composed of the modal component u k i The autocorrelation spectrum reaches the local maximum R after passing through the zero point k i The corresponding moment, and let i = i + 1;
[0127] S350, determine whether the number of iterations is greater than or equal to the preset number of iterations, if not, return to the step of obtaining the filtered modal components, if yes, calculate the correlation coefficient between every two modal components, and construct the correlation coefficient matrix CC (K×K) ,
[0128]
[0129] Where N represents the signal length; CCpq represents the modal component u p and u q The correlation matrix of and u p and u q The average value of
[0130] S360, selecting two modal components with maximum correlation coefficients, and calculating the correlation kurtosis of the two modal components;
[0131] S370, selecting a modal component with a large correlation kurtosis as a decomposed modal component, and setting K=K-1, where K is the number of modes;
[0132] S380, determining whether the current modal number K is greater than or equal to the set modal number n, if not, returning to the step of obtaining the filtered modal components, if yes, obtaining n modal components as the final decomposition result of the characteristic mode decomposition algorithm;
[0133] S400, calculating the kurtosis values of two or more modal components, and selecting the modal component with the maximum kurtosis value as the optimal mode;
[0134] Calculate the kurtosis value of each modal component after decomposition by the eigenmode decomposition algorithm, and select the mode with the largest kurtosis value as the optimal mode of the eigenmode decomposition algorithm;
[0135] Calculate the kurtosis value of the signal sequence {x(n)}:
[0136]
[0137] Among them, x i is the ith value of the signal sequence, is the average value of the signal sequence, and n is the signal length.
[0138] S500, inputting the optimal mode into a one-dimensional convolutional neural network, and using the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting;
[0139] Convolutional neural network is a forward propagation deep network evolved from perceptron. It has good data adaptability and feature extraction capabilities, and has made great breakthroughs in image processing and fault diagnosis. The above one-dimensional convolutional neural network is usually composed of several convolutional layers, pooling layers, fully connected layers and classification layers.
[0140] S510, inputting the optimal mode into a one-dimensional convolutional neural network, and the optimal mode is processed by filter convolution;
[0141] S520, generating a feature matrix using an activation function, wherein each convolution unit corresponds to a portion of a receptive field, wherein the receptive field refers to an area covered by the convolution unit in the input data;
[0142] The convolution operation process is shown in the following formula:
[0143]
[0144] Among them, ReLU(·) and * represent the activation function and convolution operation; x l (j) is defined as the jth local region in convolutional layer l; and are the weight and bias of the i-th convolution kernel respectively. Represents the output features after convolution operation and Relu activation;
[0145] Each convolution unit only processes a small part of the data in the receptive field. By performing convolution operations on the receptive fields at different positions, the network can extract features at different positions, thereby achieving effective processing and feature extraction of the input data.
[0146] S530, using the pooling layer to extract important features of the feature matrix after convolution, the formula of the maximum pooling is as follows:
[0147]
[0148] in, represents the value of the t-th neuron in the ith feature at layer l, (t∈[(j-1)W+1,jW]). W is the width of the pooling region; It represents the output value of the i-th feature of the l+1 layer;
[0149] The pooling layer usually follows the convolutional layer. Its main function is to extract important features and reduce dimensions. Currently, the most commonly used ones are maximum pooling and average pooling.
[0150] S540, the function of the fully connected layer is to convert the two-dimensional features into a one-dimensional vector through convolution and pooling operations. The formula of the fully connected layer is as follows:
[0151]
[0152] in, and are weights and biases; represents the value of the tth neuron in the ith feature of the lth layer, m represents the total number of features in the lth layer, and n represents the total number of neurons in the ith feature of the lth layer;
[0153] S600, using a gated recurrent unit to capture the timing relationship of a signal sequence;
[0154] The Gated Recurrent Unit (GRU) is evolved from LSTM and is mainly composed of an update gate and a reset gate. The reset gate is used to obtain long-term relationships, while the update gate can obtain short-term relationships. It has good time-related sequence capabilities.
[0155] u t =f(W u ·[a t ,h t-1 ])
[0156] r t =f(W r ·[a t ,h t-1 ])
[0157]
[0158]
[0159] Among them, a t is the model input at that moment, h t-1 is the hidden state at the previous moment, r t and u t They are respectively reset gate output result and update gate output result at that moment, is a candidate hidden state, used to describe the current input state and update gate state, h t is a hidden state, if u t If the value of is close to 0, the hidden state will pay more attention to the information before time t, otherwise it will be close to the candidate hidden state. f is the activation function, W u , W r , W are the weight coefficients of the reset gate, update gate and candidate output set, respectively, ⊙ represents the product between matrices;
[0160] S700, using a self-attention mechanism to calculate an attention score corresponding to an output of a gated recurrent unit, and constructing a fault diagnosis model;
[0161] For an input sequence x = {x1, x2, ..., x n}, x i ∈R k , firstly, through adaptive mapping, we obtain three vectors
[0162] q i =W q x i , k i =W k x i , v i =W vx i
[0163] is a parameter matrix;
[0164] Calculate the vector q i and vector k i The similarity between them is used to obtain an attention score α i,j , the calculation formula is:
[0165]
[0166] Among them, d k Represents k j Dimensions;
[0167] Use the softmax function to i,j Normalize to get the weight coefficient The calculation formula is:
[0168]
[0169] Among them, α i,j is the attention score, n represents the number of input vectors;
[0170] The vector v i Multiply and sum with the corresponding weight coefficient to get the attention value, and the calculation formula is:
[0171]
[0172] Among them, z j is x j The attention of x can simultaneously calculate the attention value corresponding to each value in x to form a new feature sequence z = {z1, z2, ..., z n};
[0173] The fault features contained in the bearing vibration signal are easily submerged by strong noise signals and have strong non-stationary and nonlinear characteristics. In this application, a gated recurrent unit (GRU) and a self-attention mechanism are combined with a one-dimensional convolutional neural network to construct a fault diagnosis model. The main structure of the fault diagnosis model consists of three convolutional layers, a GRU layer, a self-attention layer and a classification layer.
[0174] First, a wide convolution kernel with a kernel size of 64*1 is used to deconvolve the input features, which will bring better noise robustness. The next two layers of convolution use a convolution kernel with a kernel size of 3*1 to deepen the network and suppress overfitting. Then, GRU is used to capture the timing-related information of the signal sequence. Considering that the feature vectors output by GRU are not all decisive for bearing fault diagnosis, the self-attention mechanism is used to give different weights to the output of GRU to achieve the effect of focusing on important information.
[0175] The one-dimensional convolutional neural network is improved by combining GRU and self-attention mechanism with the one-dimensional convolutional neural network. It not only utilizes the powerful feature extraction capability of the convolutional layer, but also uses GRU to capture the potential temporal relationship between signals, and uses the characteristics of the self-attention mechanism to focus on key information, retaining more useful information of the signal and improving the accuracy of fault identification.
[0176] The present application integrates a parameter-optimized eigenmode decomposition algorithm and an improved one-dimensional convolutional neural network to provide an intelligent bearing fault diagnosis method, which can automatically extract practical information about the fault and improve the accuracy of bearing fault identification.
[0177] S800, obtaining a vibration signal and inputting it into the fault diagnosis model, and outputting a corresponding fault type;
[0178] See also Figure 1-Figure 6 , the fault diagnosis method of rotating machinery can effectively diagnose the fault of rotating machinery under heavy noise conditions, and has good stability, strong noise suppression ability, and certain generalization ability. This advantage mainly overcomes the uncertainty of artificially selecting the length and modal number of FMD filters, uses the sparrow search optimization algorithm to search for the optimal parameters of FMD, realizes adaptive FMD filtering, and has good noise reduction effect; and combines the one-dimensional convolutional network with GRU and self-attention mechanism to maximize the temporal relationship in the signal, and retain important information as much as possible in training, overcoming the problem of insufficient feature extraction and temporal relationship utilization of previous deep learning methods. Under FMD denoising and neural network feature extraction, the algorithm can more effectively diagnose the fault of rotating machinery under heavy noise conditions.
[0179] The above is a detailed introduction to the fault diagnosis method and system for a rotating machinery provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
Claims
1. A method for fault diagnosis of rotating machinery, characterized in that: include: Collect the original vibration signals of rotating machinery with different fault types; Based on the sparrow search optimization algorithm, the mode number and filter length of the characteristic mode decomposition algorithm are optimized, and the global minimum envelope entropy value and the corresponding optimal population parameter combination are obtained; Decompose the original vibration signal using the characteristic mode decomposition algorithm after parameter optimization and obtain more than two modal components; Calculate the kurtosis values of more than two modal components, and select the modal component with the largest kurtosis value as the optimal mode; The optimal mode is input into a one-dimensional convolutional neural network, and the one-dimensional convolutional neural network is used to eliminate some noise and suppress overfitting; Using gated recurrent units to capture the temporal relationship of signal sequences; The self-attention mechanism is used to calculate the attention score corresponding to the output of the gated recurrent unit and a fault diagnosis model is constructed; Obtain a vibration signal and input it into the fault diagnosis model, and output the corresponding fault type, Among them, in the step of inputting the optimal mode into a one-dimensional convolutional neural network and using the convolution layer of the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting, The one-dimensional convolutional neural network includes a convolution layer, a pooling layer, a fully connected layer and a classification layer. Among them, in the step of inputting the optimal mode into a one-dimensional convolutional neural network and using the convolution layer of the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting, The optimal mode is input into a one-dimensional convolutional neural network, and the optimal mode is processed by filter convolution; The activation function is used to generate a feature matrix, and each convolution unit corresponds to a part of the receptive field, where the receptive field refers to the area covered by the convolution unit in the input data; The convolution operation process is shown in the following formula: Among them, ReLU(·) and * represent the activation function and convolution operation; x l (j) is defined as the jth local region in convolutional layer l; and are the weight and bias of the i-th convolution kernel, Represents the output features after convolution operation and Relu activation, Among them, in the step of inputting the optimal mode into a one-dimensional convolutional neural network and using the convolution layer of the one-dimensional convolutional neural network to eliminate part of the noise and suppress overfitting, The pooling layer is used to extract the important features of the feature matrix after convolution. The formula of maximum pooling is as follows: in, represents the value of the t-th neuron in the ith feature at layer l, (t∈[(j-1)W+1,jW]), where W is the width of the pooling region; It represents the output value of the i-th feature of the l+1 layer; The fully connected layer is used to convert the two-dimensional features into a one-dimensional vector. The formula of the fully connected layer is as follows: in, and are weights and biases; represents the value of the tth neuron in the ith feature of the lth layer, m represents the total number of features in the lth layer, and n represents the total number of neurons in the ith feature of the lth layer. Wherein, in the step of using a gated recurrent unit to capture the timing relationship of a signal sequence, u t =f(W u ·[a t h t-1 ]) r t =f(W r ·[a t ,h t-1 ]) Among them, a t is the model input at that moment, h t-1 is the hidden state at the previous moment, r t and u t They are respectively reset gate output result and update gate output result at that moment, is a candidate hidden state, used to describe the current input state and update gate state, h t is a hidden state, if u t If the value of is close to 0, the hidden state will pay more attention to the information before time t, otherwise it will be close to the candidate hidden state. f is the activation function, W u , W r , W are the weight coefficients of the reset gate, update gate and candidate output set, ⊙ represents the product between matrices, Among them, in the step of using the self-attention mechanism to calculate the attention score corresponding to the output of the gated recurrent unit and constructing a fault diagnosis model, For an input sequence x={x1,x2,…,x n },x i ∈R k ,First, through adaptive mapping, three vectors are obtained is a parameter matrix; Calculate the vector q i and vector k i The similarity between them is used to obtain an attention score α i,j , the calculation formula is: Among them, d k Represents k j Dimensions; Use the softmax function to i,j Normalize to get the weight coefficient The calculation formula is: Among them, α i,j is the attention score, n represents the number of input vectors; The vector v i Multiply and sum with the corresponding weight coefficient to get the attention value, and the calculation formula is: Among them, z i is x i At the same time, the attention value corresponding to each value in x is calculated to form a new feature sequence z = {z1, z2, ..., z n }.
2. The fault diagnosis method for rotating machinery according to claim 1, characterized in that: The step of optimizing the mode number and filter length of the characteristic mode decomposition algorithm based on the sparrow search optimization algorithm and obtaining the global minimum envelope entropy value and the corresponding optimal population parameter combination includes the following steps: Initialize population parameters; Calculate the envelope entropy value and sort the steps: Decompose the original vibration signal using the characteristic mode decomposition algorithm at different sparrow positions, calculate the envelope entropy value corresponding to the current position, compare the sizes of each envelope entropy value to determine the current minimum envelope entropy value, and update the individual extreme value and the global extreme value; The envelope entropy value of the original vibration signal x(j) (j=0,1,…,n) is calculated by the following formula: Where n represents the total number of sampling points, x(j) represents the jth sampling point of the signal, and E p is the envelope entropy of signal x(j), a(j) is the envelope signal obtained after Hilbert transform of x(j), and p j is the normalized form of a(j); The formula for updating the discoverer's position is: Among them, t is the current iteration number, R2∈[0,1] is the warning value, R ST ∈[0.5, 1], Q is a random number that follows a normal distribution; Joiner position update formula: Among them, w i,j t The worst position searched for the current population; b i,j t is the optimal position of the sparrows in the current population; Q is a random number that obeys the normal distribution; x i,j t is the current joiner position; D is the dimension of the search space; The update formula of the sentinel position is: Among them, β is a random number that obeys the normal distribution; κ∈[-1,1] is a uniform random number; f i is the fitness of the current sparrow individual; f g is the fitness of the global optimal position; b i,j t is the optimal position of the current population of sparrows; x i,j t is the current alerter's position; Determine whether danger occurs. If so, update the discoverer position and the follower position; if not, update the envelope entropy value; Determine whether the number of cycles reaches the preset number of cycles. If not, return to the step of calculating the envelope entropy value and sorting to cycle. If so, output the global minimum envelope entropy value and the corresponding optimal population parameter combination.
3. The fault diagnosis method for rotating machinery according to claim 1, characterized in that: The step of decomposing the original vibration signal using the characteristic mode decomposition algorithm after parameter optimization and obtaining more than two modal components specifically includes the following steps: Input the original vibration signal x into the characteristic mode decomposition algorithm, and initialize the parameters of the characteristic mode decomposition algorithm, which include the number of decomposed modes n and the filter length L; Initializing the FIR filter bank by using K Hanning windows and initializing the number of iterations i=1, wherein the eigenmode decomposition algorithm updates the filter coefficients by iteration to establish an adaptive FIR filter bank; Get the filtered modal component u k i ; Using the original vibration signal x and the decomposed modal components u k i and the estimated failure period T k i Update the filter coefficients, where the estimated fault period T k i is composed of the modal component u k i The autocorrelation spectrum reaches the local maximum R after passing through the zero point k i The corresponding moment, and let i = i + 1; Determine whether the number of iterations is greater than or equal to the preset number of iterations. If not, return to the step of obtaining the filtered modal components. If so, calculate the correlation coefficient between each two modal components and construct the correlation coefficient matrix CC. (K×K) , Where N represents the signal length; CC pq represents the modal component u p and u q The correlation matrix of and u p and u q The average value of Select the two modal components with the largest correlation coefficients and calculate the correlation kurtosis of the two modal components; Select the modal component with large correlation kurtosis as the decomposed modal component, and set K = K-1, where K is the number of modes; Determine whether the current mode number K is greater than or equal to the set mode number n. If not, return to the step of obtaining the filtered modal components. If so, obtain n modal components as the final decomposition result of the characteristic mode decomposition algorithm.
4. The fault diagnosis method for rotating machinery according to claim 1, characterized in that: In the step of calculating the kurtosis values of two or more modal components and selecting the mode with the maximum kurtosis value as the optimal mode, Calculate the kurtosis value of each modal component after decomposition by the eigenmode decomposition algorithm, and select the mode with the largest kurtosis value as the optimal mode of the eigenmode decomposition algorithm; Calculate the kurtosis value of the signal sequence {x(n)}: Among them, x i is the ith value of the signal sequence, is the average value of the signal sequence, and n is the signal length.
5. A fault diagnosis system for rotating machinery, characterized in that: include: A memory for storing executable program codes; as well as A processor is connected to the memory, and runs a computer program corresponding to the executable program code by reading the executable program code, so as to execute the steps in the fault diagnosis method of a rotating machinery as claimed in any one of claims 1 to 4.
