Bearing fault prediction method based on improved VMD-CNN-BiLSTM

Abstract

The present invention discloses a bearing fault prediction method based on improved VMD-CNN-BiLSTM, introduces SMP chaotic mapping to initialize the distribution of eagle groups, introduces search space for global optimization in the first stage of the northern goshawk algorithm, and introduces Levy flight strategy improvement algorithm in the second stage of the northern goshawk algorithm; uses the improved northern goshawk algorithm to confirm the secondary penalty term in the VMD model; uses the effective weighted index MBVH to screen IMFs, and reconstructs the signal, and divides the reconstructed signal into a training set and a test set; uses a convolutional neural network to extract features, and performs residual connection on the output sequence of CNN and the output sequence of BiLSTM to construct an improved CNN-BiLSTM prediction model; inputs the test set, recombines and superimposes the model prediction values, obtains the output sequence, and completes the bearing fault prediction. Compared with the prior art, the present invention improves the time and accuracy of bearing fault prediction by using the UNGO algorithm to optimize the VMD denoising model, MBVH to screen the signal, and optimizes the CNN-BiLSTM prediction model.

Description

Technical Field

[0001] The invention belongs to the field of fault prediction, and in particular relates to a bearing fault prediction method based on an improved VMD-CNN-BiLSTM. Background Art

[0002] The reducer is an important component in various mechanical equipment. The reducer can convert high-speed output into low-speed high-torque input to achieve power transmission and regulation. The bearings of the reducer are key components that support the movement of the rotor and bear the load. They are prone to wear, fatigue and failure during long-term operation. The reducer bearings are inside a sealed machine, and it is impossible to predict the bearings in real time. Fault prediction technology based on vibration signals can quickly and accurately predict the operating status of bearings. In the field of fault prediction, fault prediction based on vibration signals will be affected by the operating environment, noise, dust and other harsh conditions of the equipment itself. Better filtering and denoising of the collected signals can better diagnose the faults of the bearing operating status.

[0003] Early fault prediction uses computational intelligence technologies such as artificial neural networks, but the fault prediction ability is poor. The current prediction methods are mainly divided into physical methods and data-driven methods. Due to the influence of external factors such as complex environment, physical methods are not suitable for prediction. Meta-heuristic methods perform in-depth fault prediction on feature variables. Compared with machine learning models such as minimum support vector machine (SVM), random forest (RF), and decision tree (DT), they have higher requirements on data quality.

[0004] In recent years, intelligent optimization algorithms have been continuously proposed, such as the whale algorithm, genetic algorithm, particle swarm algorithm, etc., which have significantly improved the accuracy and computational efficiency of different algorithm structures. Combining deep models with intelligent optimization algorithms, adjusting the parameters of deep models through search and optimization algorithms, thereby improving the performance of the model, reducing the dimensions of model input through dimensionality reduction, feature selection and other methods, and improving the processing efficiency and generalization ability of the model.

[0005] Common signal preprocessing methods include: wavelet threshold denoising, empirical mode decomposition (EMD), singular value decomposition (SVD), filtering denoising, and variational mode decomposition (Varational Mode Decomposition). The VMD algorithm has significant advantages in vibration signal fault prediction. It is an adaptive and completely non-recursive signal processing method that can determine the number of modal decompositions of a given sequence according to actual conditions. This makes VMD advantageous in processing complex and nonlinear systems. It can adaptively capture and analyze the intrinsic characteristics of the system. However, the VMD algorithm cannot separate non-stationary signals in one step. It has problems such as incomplete denoising, high algorithm complexity, and easy distortion of target signals at the current stage. Summary of the invention

[0006] Purpose of the invention: In view of the problems pointed out in the background technology, the present invention discloses a bearing fault prediction method based on improved VMD-CNN-BiLSTM, which combines the intelligent optimization algorithm to improve VMD and CNN-BiLSTM to improve the accuracy of the deep learning model for bearing faults.

[0007] Technical solution: The present invention proposes a bearing fault prediction method based on improved VMD-CNN-BiLSTM, comprising the following steps:

[0008] Step 1: Obtain the vibration signal of the bearing parts in the reducer in the manufacturing industry;

[0009] Step 2: Using the improved northern goshawk algorithm UNGO to optimize the VMD model, using the optimized VMD model to denoise the vibration signal, and obtaining several IMFs after decomposition; the improved northern goshawk algorithm UNGO introduces SMP chaos mapping to optimize the initialization population distribution, local optimization in the neighboring search space, and Levy flight strategy to prevent falling into the local optimum;

[0010] Step 3: Use the effective weighted standard MBVH to screen the IMFs of the denoised vibration signal and reconstruct the signal;

[0011] Step 4: Divide the reconstructed signal into a training set and a test set, build an improved CNN-BiLSTM prediction model, and complete fault prediction; the improved CNN-BiLSTM prediction model uses the local features extracted by CNN as the input of BiLSTM, and CNN uses the improved Northern Goshawk algorithm UNGO in step 2 to iteratively optimize the convolution layer with the optimal step size for convolution, and residually connects the output sequence of CNN and the output sequence of BiLSTM.

[0012] Furthermore, in step 2, the acquired vibration signal is iterated with a quadratic penalty factor α based on the improved Northern Goshawk algorithm UNGO, and the information entropy increment combined with the kurtosis index is selected as the objective function to obtain the number of intrinsic mode functions K, form the optimal VMD parameters [K, α], and perform VMD decomposition into several IMFs.

[0013] Furthermore, the improved northern goshawk algorithm UNGO, the improvement steps are as follows:

[0014] Step 2.1: Initialize the population, introduce SMP chaotic mapping to optimize the initial population distribution, perform global optimization on the adjacent search space, and let the position X of the i-th goshawk be i , identify the prey hunting position in the jth dimension, then the position of the i-th goshawk is updated as:

[0015]

[0016] In the formula, represents the new position of the i-th goshawk in the j-th dimension, l is a random number in [0, 1], I is a random integer, and F pi represents the objective function value of the hunting position, F i represents the current objective function value, P i，j represents the prey hunting position of the i-th goshawk in the j-th dimension, X i,j represents the position of the i-th goshawk in the j-th dimension;

[0017] The new objective function value of the i-th goshawk Updated to:

[0018]

[0019] The position of the i-th goshawk is updated as:

[0020]

[0021] The position of the i-th goshawk after the first stage update is X i :

[0022]

[0023] In the formula, Represents the updated objective function value;

[0024]

[0025] Where, X i,j new，p1 represents the new position of the i-th goshawk in the j-th dimension, l is a random number in [0,1], I is a random integer, and F Pi represents the objective function value of the hunting position, Fi represents the current objective function value, P i,j represents the prey hunting position of the i-th goshawk in the j-th dimension, X i,j represents the position of the i-th goshawk in the j-th dimension, and I represents.

[0026] The new objective function value of the i-th goshawk Updated to:

[0027]

[0028] The position of the i-th goshawk is updated as:

[0029]

[0030] The position of the i-th goshawk after the first stage update is X i :

[0031]

[0032] Step 2.2: In the second stage of the algorithm, the neighboring search space is introduced for local optimization. The position of the i-th goshawk becomes:

[0033]

[0034] Step 2.3: In the second stage of the algorithm, the Levy flight strategy is introduced to prevent falling into the local optimum. The position of the i-th northern goshawk in the second stage of the northern goshawk algorithm is Updated to:

[0035]

[0036] In the formula, levy is obtained by the Levy formula, and M is the relevant neighboring search range;

[0037] The objective function value and final position of the i-th goshawk in the second stage become:

[0038]

[0039]

[0040] Furthermore, in step 3, the effective weighted standard MBVH is used to screen IMFs, and the MBVH formula is as follows:

[0041]

[0042] Where: S(i) is the i-th modal component x i (t); N is the length of the modal component; K is the number of modal components; C(i) is the i-th modal component xi (t) is the correlation coefficient between the original signal; E(·) is the expectation operator, and the weighted kurtosis value in each modal component is calculated. The formula is as follows:

[0043]

[0044] When the MBVH value is greater than 0, the IMF is retained; otherwise, the component is considered to be an invalid component or noise and is deleted.

[0045] Furthermore, the improved CNN-BiLSTM prediction model in step 4 is specifically as follows:

[0046] (1) Input a vibration signal, and represent the input signal source as a time series, where each time step corresponds to a vibration data point;

[0047] (2) Using CNN convolutional neural network to extract local features based on the input vibration signal;

[0048] (3) The CNN pooled matrix is ​​used as the input of the bidirectional long short-term memory network, and the bidirectional long short-term memory network is used to learn the global context information to obtain the output sequence of BiLSTM;

[0049] (4) The output sequence of CNN and the output sequence of BiLSTM are organically fused through residual connection.

[0050] Furthermore, the CNN convolutional neural network performs 3×3 convolutional layers with an improved northern goshawk algorithm to iteratively optimize the optimal step size, slides the convolution kernel on the input vibration signal, slides one time step each time, convolves the convolution kernel with the input vibration signal data point to generate a related feature map, selects Relu as the activation function, performs batch normalization, activation layer, and pooling layer processing respectively, and performs residual connection between the convolution layer and the pooling layer to obtain an output sequence of local features.

[0051] Beneficial effects:

[0052] (1) The improved CNN-BiLSTM prediction model of the present invention uses residual connections between each convolutional layer and pooling layer, so that CNN can better exert its feature extraction ability, alleviate the gradient vanishing problem in deep networks, and enable the model to extract features at a deeper level. The residual connection can effectively add the output sequences of CNN and BiLSTM, organically integrate the local features extracted by CNN and the global context information learned by BiLSTM, and enrich the model's ability to express the features of the input sequence.

[0053] (2) The present invention uses the improved northern goshawk algorithm to optimize the VMD model, iteratively optimizes the penalty factor ɑ in the VMD algorithm, and the selection of the secondary penalty factor ɑ directly affects the acquisition of the final result, and decomposes the IMF under different states.

[0054] (3) The present invention improves the northern goshawk algorithm and introduces SMP chaotic mapping to initialize the hawk group, which can solve the problem that the number of populations initialized by the algorithm is uneven, causing the individuals in the population to fall into the local optimal state. In addition, in the second stage of the algorithm (pursuit and escape stage), the neighboring search space is introduced to perform local optimization, which speeds up the search for the optimal solution in the current search area and improves the overall search efficiency. The Levy flight strategy is introduced into the second stage of the northern goshawk algorithm to prevent the problem of falling into the local optimal state.

[0055] (4) The present invention screens IMFs based on the effective weighted standard MBVH. When the obtained IMBVH is greater than 0, the IMF is retained. Otherwise, the component is considered to be an invalid component and is deleted as noise. Deleting the invalid component enhances the validity of the signal source data, data readability and fault prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is the overall flow chart based on the improved VMD-CNN-Bilstmm;

[0057] Figure 2 Flowchart of the improved northern goshawk algorithm;

[0058] Figure 3 Run the flowchart for the improved CNN-Bilstmm. DETAILED DESCRIPTION

[0059] See attached Figure 1 The present invention discloses a bearing fault prediction method based on improved VMD-CNN-BiLSTM. The improved Northern Goshawk algorithm (UNGO) is used to iteratively select the secondary penalty factor ɑ in VMD, and the new index EWSC based on the effective weighted correlation sparsity index is used to screen several IMFs decomposed in VMD, and the reconstructed signal is divided into a training set and a test set. An improved CNN-BiLSTM model is constructed, the test set is input, the model prediction values ​​are reorganized and superimposed, and an output sequence is obtained to complete the subsequent fault classification. The specific implementation method includes the following steps:

[0060] Step 1: Obtain the vibration signals of noisy bearings and other components in the reducer in the manufacturing industry and perform UNGO-VMD decomposition. Set the signal receiving model as:

[0061] f(t)=l(t)+e(t)

[0062] Where f(t) is the vibration signal and e(t) is the noise signal.

[0063] Step 2: Use UNOG-VMD to decompose the acquired vibration signal, use the new indicator MBVH to screen IMFs, reconstruct the screened IMFs and divide them into training set and test set.

[0064] Step 2.1 Iterate the obtained f(t) using the improved Northern Goshawk algorithm (UNGO) with a quadratic penalty factor ɑ, select the information entropy increment combined with the kurtosis index as the objective function, obtain the number of intrinsic mode functions K, form the optimal VMD parameters [K, ɑ], and perform VMD decomposition into several IMFs.

[0065] Step 2.2 further decomposes the UNGO-VMD model to obtain several IMFs. The process is as follows:

[0066] Step 2.3 The signal is an amplitude-frequency modulated signal with the characteristic mode function, and its formula is as follows:

[0067]

[0068] In the formula, A k (t) is the instantaneous amplitude, φ k (t) is u k The phase of (t).

[0069] Step 2.4 uses Hilbert transform to solve the K value and then transforms it into solving the constrained variational problem. The formula is as follows:

[0070]

[0071] Where U K is the kth modal component signal after VMD decomposition, ω k is the center frequency of the kth modal component signal, t represents time, δ(t) is the Dirac function, j is the imaginary number sign, f is the input vibration signal, and k is an integer greater than or equal to 1.

[0072] Step 2.5: Transform the constrained variational problem into an unconstrained problem. Taking advantage of the quadratic penalty term and the Lagrangian multiplier method, the augmented Lagrangian function is introduced. The variational formula is as follows:

[0073]

[0074] In the formula, ɑ is the quadratic penalty factor, λ is the Lagrange multiplier, Derivative the expression in brackets with respect to t.

[0075] Step 2.6 The improved northern goshawk algorithm (UNGO) performs iterative optimization to obtain the secondary penalty factor ɑ and initialize the parameters in the above formula. Further explanation of the above improved northern goshawk algorithm (UNGO): At the initialization of the eagle group, the eagle group is marked as U, and the eagle group consists of N eagles. The goshawk population is represented by a matrix as follows:

[0076]

[0077] U is the population matrix of the northern goshawk; N is the population size of the northern goshawk; M is the dimension of the problem to be solved. The target vector F of the northern goshawk population algorithm.

[0078]

[0079] In the formula, F i =F(U i ) is the objective function value of the i-th northern goshawk position.

[0080] The phenomenon of uneven number of populations in the algorithm initialization may cause the individuals in the population to fall into the local optimal problem. SMP chaotic mapping is introduced to initialize the eagle group. The iterative formula is as follows:

[0081]

[0082] Where X(t) represents the state value at the current moment, η∈(0,1), μ∈(0,1), and r is a random number between 0 and 1.

[0083] In the first stage of the algorithm, the search space is globally optimized, and the prey hunting position P of the i-th goshawk is i Set to:

[0084] P i =X k

[0085] Wherein, K is a random integer in the range of [1, i)∪(i, N], i=1, 2, 3...N, K=1, 2, ...i...N.

[0086] The position of the prey hunting position identified by the i-th goshawk in the j-dimension is:

[0087]

[0088] In the formula, represents the new position of the i-th goshawk in the j-th dimension, l is a random number in [0,1], I is a random integer, and F pi represents the objective function value of the hunting position, Fi represents the current objective function value, and P i,j represents the prey hunting position of the i-th goshawk in the j-th dimension, P i,jrepresents the position of the i-th goshawk in the j-th dimension.

[0089] The new objective function value of the i-th goshawk Updated to:

[0090]

[0091] The position of the i-th goshawk is updated as:

[0092]

[0093] The position of the i-th goshawk after the first stage update is X i :

[0094]

[0095] In the formula, represents the new position of the i-th goshawk in the j-th dimension, represents the updated objective function value, F i Represents the objective function value before updating.

[0096] The second stage of the algorithm (the pursuit and escape stage) introduces the neighboring search space to strengthen the local optimization. becomes:

[0097]

[0098] in, represents the updated position of the goshawk, X i,j Indicates that M is the relevant adjacent search range, and its acquisition formula is as follows:

[0099] M=0.02(1-n / N)

[0100] Where n is the current number of iterations and N is the maximum number of iterations.

[0101] The Levy flight strategy is introduced into the second stage of the Northern Goshawk algorithm to prevent falling into the local optimal problem. The Levy flight strategy formula is as follows:

[0102]

[0103]

[0104] Where β is a random number in [0,2], u and v follow a normal distribution, u~N(0,σ2), v~N(0,1), and rand(0,1) is a random number in [0,1].

[0105] The position X of the i-th northern goshawk in the second stage of the northern goshawk algorithm inew,p2 Updated to:

[0106]

[0107] Where levy is obtained by the above Levy strategy.

[0108] The objective function value and final optimization solution for the i-th goshawk in the second stage become:

[0109]

[0110]

[0111] Step 2.7 introduces the penalty factor ɑ and updates the component frequency U through Parseval / Plancherel Fourier isometric transformation K and center frequency, which is expressed as follows:

[0112]

[0113]

[0114] Where U k is the frequency of k modal component signals, X(ω) is the original input signal, and n is the current number of iterations. The updated Lagrange formula is as follows:

[0115]

[0116] Where ι is the update parameter of the Lagrange multiplier, ι is taken as 10 -3 .

[0117] Step 2.8 sets the judgment threshold ε. When the equation is satisfied, the iteration ends, otherwise the iteration continues. The formula is as follows:

[0118]

[0119] Where ω is the convergence criterion tolerance value, and ε is taken as 10 －6 .

[0120] Get several IMFs after UNGO-VMD denoising.

[0121] Step 3: IMFs are screened based on the effective weighted standard MBVH.

[0122] Step 3.1 Calculate the weighted kurtosis value in each modal component. The formula is as follows:

[0123]

[0124] Where: S(i) is the i-th modal component xi (t); N is the length of the modal component; K is the number of modal components; C(i) is the i-th modal component x i (t) is the correlation coefficient between the original signal; E(·) is the expectation operator.

[0125] Step 3.2 calculates the weighted kurtosis value in each modal component, and the formula is as follows:

[0126]

[0127] When the obtained IMBVH is greater than 0, the IMF is retained, otherwise the component is considered to be an invalid component and deleted as noise. Deleting invalid components enhances the validity of signal source data, data readability and fault prediction accuracy.

[0128] Step 4 divides the denoised IMFs reconstructed signal into a training set and a test set, builds an improved CNN-BiLSTM model, reorganizes and superimposes the model prediction values ​​to obtain the output sequence, completes the fault prediction, and evaluates the model.

[0129] Step 4.1 The input layer is the input feature matrix. CNN consists of a convolution layer and a pooling layer. After the feature matrix is ​​normalized and input into the convolution layer, a new feature matrix is ​​obtained. The convolution layer with a step size of 3×3 and the optimal step size of the iteration is used. The calculation formula of the convolution layer is as follows:

[0130]

[0131] In the formula, Hi represents the feature matrix of the i-th layer of the convolutional network, f COV represents the activation function, Wi represents the weight value of the convolution kernel used in the i-th layer, and b i Represents the bias term.

[0132] The activation function uses the Relu function, and its formula is as follows:

[0133] ReLu=max(0,x)

[0134] Where X represents the input feature matrix.

[0135] Step 4.2 The calculation formula of the maximum pooling layer is as follows:

[0136]

[0137]

[0138] In the formula, h' represents the length of the input feature matrix, W' represents the width of the input feature matrix, h represents the length of the output feature matrix after pooling, w represents the width of the output feature matrix after convolution, the convolution kernel size is k, and the step size is s.

[0139] Step 4.3 further explains the operation process of the improved CNN-BiLSTM. The improved CNN-BiLSTM bearing fault prediction model introduces residual connections, the improved Northern Goshawk algorithm iteratively optimizes the convolution layer of the optimal step size, and performs batch normalization, activation layer, and pooling layer processing respectively. The specific steps are as follows:

[0140] (1) Input a vibration signal and represent the input signal source as a time series, where each time step corresponds to a vibration data point.

[0141] (2) Select a 3x3 convolution kernel with an iterative optimal step size. Slide the convolution kernel on the input vibration signal, one time step at a time. At each position, convolve the convolution kernel with the data point in the receptive field to generate a feature map, and select ReLU as the activation function.

[0142] (3) Perform residual connection between the input and output of the convolutional layer.

[0143] (4) Perform a maximum pooling operation to reduce the dimension of the feature map.

[0144] The above output is used as input to the next convolutional layer and steps (2-4) are repeated.

[0145] The CNN output sequence and the Bilstmm output sequence are further residually connected to organically integrate the learned global context information, thereby enriching the model's ability to express the features of the input sequence.

[0146] Step 4.4 uses the pooled matrix as the input of the bidirectional long short-term memory network, processing it from both the forward and reverse directions. Detailed explanation of the relevant computational issues in the bidirectional long short-term memory network.

[0147] Step 4.5 Through two LSTM chains in opposite directions (one forward and one reverse), BiLSTM can more fully understand the contextual information of the sequence data. The forward LSTM accepts the input of the current time step and the hidden state of the previous time step, and initializes the cell state. Calculate the relevant forget gate. The formula for calculating the forget gate is as follows:

[0148] f t =σ(W f ·[h t-1 , x t ]+b f )

[0149] In the formula, ht-1 is the hidden state at the previous moment, x t is the input word at the current moment, b f Represents the bias term. Calculate the memory gate and select the information to be memorized. The calculation formula is as follows:

[0150]

[0151] In the formula, h t-1 is the hidden state at the previous moment, x t is the current input. t is the memory gate value, W i , Wc represents the current weight matrix, tanh is the hyperbolic tangent function, h t-1 is the hidden state of the previous moment, σ represents the activation function, b i 、b c Represents the bias term.

[0152]

[0153] In the formula, Temporary cell state, C t-1 is the cell state at the previous moment, C t is the cell state at the current moment, i t is the memory gate value.

[0154] Step 4.6 calculates the output gate and the current hidden state. The formula is as follows:

[0155]

[0156] In the formula, h t-1 is the hidden state of the previous moment, O t is the output gate value, h t is the hidden state, W o Represents the current weight matrix.

[0157] Step 4.7 calculates the relevant hidden layer state sequence {h0, h1, h2..., h n-1 The reverse LSTM starts from the end of the sequence, accepts the input of the current time step and the hidden state of the next time step, and performs the process of forget gate, input gate, candidate value, cell state update, output gate and hidden state output, which is the same as the forward LSTM but in the opposite direction.

[0158] Step 4.8 reorganizes and superimposes the predicted values ​​obtained in CNN-BiLSTM to obtain the output sequence and complete the bearing fault prediction.

[0159] Step 5: Use the mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE) to evaluate the prediction model. The formula is as follows:

[0160]

[0161] Where n is the number of predicted data points, y i and pi are the actual observed value and the predicted observed value at time i respectively.

[0162] The lower the MAPE value, the higher the prediction accuracy of the model.

[0163] The patent of this invention can be combined with relevant enterprise systems to perform relevant reducer bearing fault detection.

[0164] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A bearing fault prediction method based on improved VMD-CNN-BiLSTM, characterized in that: The following steps are involved: Step 1: Obtain the vibration signal of the bearing parts in the reducer in the manufacturing industry; Step 2: Using the improved northern goshawk algorithm UNGO to optimize the VMD model, using the optimized VMD model to denoise the vibration signal, and obtaining several IMFs after decomposition; the improved northern goshawk algorithm UNGO introduces SPM chaos mapping to optimize the initialization population distribution, local optimization in the neighboring search space, and Levy flight strategy to prevent falling into the local optimum; The improved northern goshawk algorithm UNGO, the improvement steps are as follows: Step 2.1: Initialize the population, introduce SPM chaotic mapping to optimize the initial population distribution, perform global optimization on the adjacent search space, and let the position X of the i-th goshawk be i , identify the prey hunting position in the jth dimension, then the position of the i-th goshawk is updated as: In the formula, represents the new position of the i-th goshawk in the j-th dimension, l is a random number in [0,1], I is a random integer, and F pi represents the objective function value of the hunting position, F i represents the current objective function value, P i,j represents the prey hunting position of the i-th goshawk in the j-th dimension, X i,j represents the position of the i-th goshawk in the j-th dimension; The new objective function value of the i-th goshawk Updated to: The position of the i-th goshawk is updated as: The position of the i-th goshawk after the first stage update is X i : In the formula, Represents the updated objective function value; Step 2.2: In the second stage of the algorithm, the neighboring search space is introduced for local optimization. The position of the i-th goshawk becomes: Step 2.3: In the second stage of the algorithm, the Levy flight strategy is introduced to prevent falling into the local optimum. The position of the i-th northern goshawk in the second stage of the northern goshawk algorithm is Updated to: In the formula, levy is obtained by the Levy formula, and M is the relevant neighboring search range; The objective function value and final position of the i-th goshawk in the second stage become: Step 3: Use the effective weighted standard MBVH to screen the IMFs of the denoised vibration signal and reconstruct the signal; The effective weighted standard MBVH is used to screen IMFs. The MBVH formula is as follows: Where: S(i) is the i-th modal component x i (t); N is the length of the modal component; K is the number of modal components; C(i) is the i-th modal component x i (t) is the correlation coefficient between the original signal; E(·) is the expectation operator, and the weighted kurtosis value in each modal component is calculated. The formula is as follows: When the MBVH value is greater than 0, the IMF is retained, otherwise the component is considered to be an invalid component and noise and is deleted; Step 4: Divide the reconstructed signal into a training set and a test set, build an improved CNN-BiLSTM prediction model, and complete fault prediction; the improved CNN-BiLSTM prediction model uses the local features extracted by CNN as the input of BiLSTM, and CNN uses the improved Northern Goshawk algorithm UNGO in step 2 to iteratively optimize the convolution layer with the optimal step size for convolution, and residually connects the output sequence of CNN and the output sequence of BiLSTM.

2. The bearing fault prediction method based on improved VMD-CNN-BiLSTM according to claim 1 is characterized in that: In the step 2, the acquired vibration signal is iterated using the quadratic penalty factor ɑ based on the improved Northern Goshawk algorithm UNGO, and the information entropy increment combined with the kurtosis index is selected as the objective function to obtain the number of intrinsic mode functions K, form the optimal VMD parameters [K, ɑ], and perform VMD decomposition into several IMFs.

3. The bearing fault prediction method based on improved VMD-CNN-BiLSTM according to claim 1, characterized in that: The improved CNN-BiLSTM prediction model in step 4 is specifically as follows: (1) Input a vibration signal, and represent the input signal source as a time series, where each time step corresponds to a vibration data point; (2) Using CNN convolutional neural network to extract local features based on the input vibration signal; (3) The CNN pooled matrix is ​​used as the input of the bidirectional long short-term memory network, and the bidirectional long short-term memory network is used to learn the global context information to obtain the output sequence of BiLSTM; (4) The output sequence of CNN and the output sequence of BiLSTM are organically fused through residual connection.

4. The bearing fault prediction method based on improved VMD-CNN-BiLSTM according to claim 3 is characterized in that: The CNN convolutional neural network performs 3×3 and improved northern goshawk algorithm to iteratively optimize the convolution layer with the optimal step size, slides the convolution kernel on the input vibration signal, slides one time step each time, convolves the convolution kernel with the input vibration signal data point to generate a related feature map, selects Relu as the activation function, performs batch normalization, activation layer, and pooling layer processing respectively, and performs residual connection between the convolution layer and the pooling layer to obtain an output sequence of local features.

Images