A rolling bearing vibration signal prediction method based on an improved Cordyceps sinensis optimization algorithm
By improving the Cordyceps sinensis optimization algorithm, combining Hammersley sequence initialization and multidifference Cauchy-Gaussian mutation strategy, and optimizing the parameters of VMD and BiLSTM, the decomposition and hyperparameter setting problems in rolling bearing vibration signal prediction are solved, thereby improving prediction accuracy and robustness.
Patent Information
- Application Number
- CN202611477760.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for predicting vibration signals in rolling bearings suffer from problems such as low signal decomposition quality, reliance on human experience for hyperparameter settings, and insufficient optimization performance of optimization algorithms, resulting in poor prediction accuracy and robustness.
The optimization algorithm for Cordyceps sinensis is improved by introducing Hammersley sequence initialization, adaptive perturbation, and multi-difference Cauchy-Gaussian mutation strategy. The parameters of variational mode decomposition and bidirectional long short-term memory network are optimized to improve signal decomposition quality and prediction accuracy.
By improving the optimization algorithm and parameter settings, the decomposition quality and prediction accuracy of rolling bearing vibration signals are enhanced, and the stability and adaptability of the prediction results are improved, providing a reliable basis for the condition monitoring of rotating machinery.
Smart Images

Figure CN123115690A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vibration signal prediction technology, and in particular to a method for predicting vibration signals of rolling bearings based on an improved Cordyceps sinensis optimization algorithm. Background Technology
[0002] Rolling bearings are crucial support components in rotating machinery, widely used in industries such as petrochemicals, metallurgy, and energy. Under long-term operation and complex conditions, rolling bearings are prone to performance degradation and changes in vibration signal characteristics due to factors such as load fluctuations, mechanical shocks, and component coupling. When a bearing malfunctions or fails, it can affect the normal operation of equipment, and in severe cases, cause downtime or safety accidents. Therefore, accurately predicting the trend of rolling bearing vibration signals is of great significance for equipment condition monitoring and fault early warning.
[0003] Bearing vibration signals typically exhibit nonlinear, non-stationary, and multi-scale coupling characteristics, and are susceptible to noise interference, making their variation patterns difficult to extract directly. Existing techniques often employ empirical mode decomposition (EMD) and ensemble EMD methods for signal decomposition, but these may suffer from problems such as mode aliasing, residual noise, and parameter dependence on human experience. Variational mode decomposition (VMD) can improve mode aliasing to some extent, but its decomposition effect is significantly affected by the number of modal components and the penalty factor. Inappropriate parameter settings can easily lead to under-decomposition, over-decomposition, or mode aliasing, affecting the quality of input features for subsequent prediction models.
[0004] In time series prediction, Bidirectional Long Short-Term Memory (BiLSTM) networks can extract temporal dependency information from both forward and backward directions, exhibiting strong nonlinear representation capabilities. However, their prediction performance is affected by hyperparameters such as the number of hidden layer units, learning rate, and L2 regularization coefficient. Parameters that are too large or too small can lead to overfitting, insufficient feature extraction ability, or decreased convergence stability. In existing technologies, these hyperparameters are often set based on human experience or use uniform parameters for each signal component, making it difficult to adapt to the differences in temporal characteristics among different components.
[0005] To reduce the impact of manual parameter tuning, swarm intelligence optimization algorithms are increasingly being used for parameter optimization. The Cordyceps sinensis optimization algorithm (CFO) has good global search and local exploitation capabilities, but traditional CFO still suffers from problems such as uneven population distribution due to random initialization, decreased diversity in the later stages of iteration, stagnation in individual position updates, insufficient utilization of population difference information, and susceptibility to getting trapped in local optima, thus affecting optimization accuracy and convergence stability.
[0006] In summary, existing technologies for predicting vibration signals from rolling bearings in rotating machinery still have room for improvement in areas such as complex signal decomposition, determination of prediction model hyperparameters, and optimization algorithm performance. How to reduce the influence of manual parameter settings, improve signal decomposition quality, rationally determine BiLSTM hyperparameters, and enhance the optimization algorithm's performance, thereby improving the accuracy and robustness of bearing vibration signal prediction, is a key technical problem that needs to be solved in this field. Summary of the Invention
[0007] The purpose of this invention is to provide a method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm. The improved Cordyceps sinensis optimization algorithm is constructed by introducing Hammersley sequence initialization, adaptive perturbation, and multi-difference Cauchy-Gaussian mutation strategy. The algorithm is then used to optimize VMD parameters and BiLSTM hyperparameters, thereby reducing the influence of manual parameter tuning and improving the decomposition quality, prediction accuracy, and robustness of rolling bearing vibration signals in rotating machinery.
[0008] To achieve the above objectives, this invention provides a method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm, comprising: S1. Obtain the vibration signal collected under the operating state of the centrifugal compressor drive end bearing, and preprocess the vibration signal to obtain an initial vibration sequence; S2. Based on the Cordyceps sinensis optimization algorithm, Hammersley sequence initialization strategy, adaptive perturbation strategy and multi-difference Cauchy-Gaussian mutation strategy are introduced to improve the population initialization method and the population individual position update method, and to construct an improved Cordyceps sinensis optimization algorithm. S3. Using the penalty factor and the number of modal components of variational mode decomposition as parameters to be optimized, the improved Cordyceps sinensis optimization algorithm is used to find the optimal penalty factor and the optimal number of modal components. S4. Construct a variational mode decomposition model based on the optimal penalty factor and the optimal number of modal components, and input the initial vibration sequence into the variational mode decomposition model to obtain several modal components and residual components; S5. Using a sliding window approach, prediction samples are constructed for the modal components and residual components respectively. The L2 regularization coefficient, learning rate, and number of hidden layer units of the bidirectional long short-term memory network are used as parameters to be optimized. The improved Cordyceps sinensis optimization algorithm is used to perform optimization respectively to obtain the optimal combination of prediction hyperparameters for each component. Based on the optimal combination of prediction hyperparameters, a prediction model for each component is constructed. S6. Input the modal components and residual components into a bidirectional long short-term memory network constructed based on the optimal prediction hyperparameter combination corresponding to each component for prediction, and obtain the prediction results of each component; reconstruct and denormalize the prediction results of each component to obtain the complete prediction result of the bearing vibration signal.
[0009] Preferably, in step S2, the process of constructing the improved Cordyceps sinensis optimization algorithm includes: S2.1. The Hammersley sequence initialization strategy is used to generate an initial population in the search space; S2.2 Entering the exploration stage, the population individuals randomly update their positions using either the wave-forward operator or the spiral-ascending operator; S2.3. Entering the parasitic stage, individuals in the population update their positions by adopting reparasitic behavior, optimal parasitic behavior, or adaptive perturbation strategy according to the selection probability. S2.4 Entering the mutation stage, a multi-difference Cauchy-Gaussian mutation strategy is adopted to construct multi-difference mutation directions, and Cauchy perturbation or Gaussian perturbation is dynamically selected according to the iteration process to generate mutation candidate individuals.
[0010] Preferably, the Hammersley sequence initialization strategy in step S2.1 includes: The ratio of individual index to population size is used as the first dimension of the Hammersley sequence point. Select distinct prime numbers, expand the individual indices according to the base of each prime number, and generate the remaining coordinates of the Hammersley sequence points using the corresponding reflection function; By combining the first dimension coordinate with the other dimensions coordinates, we obtain the D-dimensional Hammersley low-difference sequence points; Mapping the coordinates of the Hammersley low-difference sequence points to the corresponding search intervals yields the initial positions of individuals in the population. The mapping formula is as follows:
[0011] in, Indicates the first The individual in the first Initial position on the dimension, Indicates the first Hammersley point at the j Coordinates on the dimension and The first Lower and upper bounds of a dimensional variable. .
[0012] Preferably, in the exploration stage of step S2.2, the update formulas for the wave advance operator and the spiral ascent operator are as follows: Wave forward operator:
[0013]
[0014] Spiral Ascent Operator:
[0015]
[0016] in, Indicates the exploration stage The updated location of each individual This indicates the location of the individual with the optimal fitness value during the exploration phase. and These represent the exploration stages. and the The current location of each individual. The wave disturbance factor. As a spiral regulating factor, This represents the current iteration number. The maximum number of iterations, Let be a random vector with elements in the interval [0,1]. A random number in the range [0,1]. It is a random integer that takes the value 1 or 2.
[0017] Preferably, in the parasitic stage of step S2.3, the update formulas for the reparasitic behavior, optimal parasitic behavior, and adaptive perturbation strategy are as follows: Reparasitic behavior update formula:
[0018] Optimal Parasitic Behavior Update Formula:
[0019]
[0020] in, Indicates the parasitic stage The updated location of each individual This indicates the position of the individual with the optimal fitness value during the parasitic stage. and These represent the parasitic stage, number 1 and 2. and the The current location of each individual. For a random scalar that follows a standard normal distribution, The optimal parasitic perturbation factor, δ It is a random perturbation vector that follows a normal distribution; The update formula for the adaptive perturbation strategy is:
[0021]
[0022] in, , Let represent the positions of two random individuals in the parasitic stage population, and let ? be a random perturbation vector following a normal distribution. This indicates taking the absolute value of each component of the vector.
[0023] Preferably, in step S2.4, the update formula for the multi-difference Cauchy-Gaussian mutation strategy is:
[0024] in:
[0025]
[0026] in, Indicates the parasitic stage The updated location of each individual This indicates the position of the individual with the optimal fitness value during the parasitic stage. These represent the parasitic stage, number 1 and 2. The current location of each individual. , and Let [the vector be] a random weight vector in [0,1]. The Cauchy perturbation factor, is the Gaussian perturbation factor.
[0027] Preferably, the optimization using the improved Cordyceps sinensis optimization algorithm in step S3 includes: The penalty factor and the number of modal components of variational mode decomposition are used as parameters to be optimized. Using the minimum envelope entropy as the fitness function, the formula for calculating the minimum envelope entropy is as follows:
[0028]
[0029] in, Represents the minimum envelope entropy. Indicates the number of modal components. Indicates the number of signal sampling points. Indicates the first The envelope signal obtained by the Hilbert transform of the nth modal component is in the nth... The amplitude at each sampling point This represents the corresponding normalized envelope amplitude.
[0030] An improved Cordyceps sinensis optimization algorithm was used to iteratively find the optimal combination of parameters with the best fitness as the optimal combination of variational mode decomposition parameters, and a variational mode decomposition model was constructed based on the optimal combination of variational mode decomposition parameters.
[0031] Preferably, step S4, which involves inputting the initial vibration sequence into the variational mode decomposition model to obtain several modal components and residual components, includes: A variational constrained optimization model is constructed, and a penalty factor and Lagrange multipliers are introduced to transform the variational constrained optimization model into an unconstrained form. The alternating direction multiplier method is used to iteratively update each modal component, center frequency, and Lagrange multiplier until the convergence condition is met, resulting in several modal components and residual components.
[0032] Preferably, the hyperparameters of the bidirectional long short-term memory network in step S5 are optimized using an improved Cordyceps sinensis optimization algorithm, including: The L2 regularization coefficient, learning rate, and number of hidden layer units of the bidirectional long short-term memory network are used as hyperparameters to be optimized. A sliding window approach is used to construct training and test samples for each modal component and residual component; The root mean square error is used as the fitness function; An improved Cordyceps sinensis optimization algorithm was used to iteratively optimize the hyperparameters of each signal component to obtain the optimal combination of prediction hyperparameters for each signal component. Based on the optimal combination of prediction hyperparameters for each signal component, a corresponding bidirectional long short-term memory network prediction model was constructed.
[0033] Preferably, in step S6, the step of inputting each signal component into a bidirectional long short-term memory network for prediction and reconstructing the prediction results of each signal component includes: Each modal component and residual component is input into a bidirectional long short-term memory network constructed based on its optimal prediction hyperparameter combination to obtain the prediction results of each signal component. The prediction results of each signal component are reconstructed using a linear superposition method. The reconstruction formula is as follows:
[0034] in, This represents the final reconstructed predicted sequence of bearing vibration signals. Indicates the first The prediction results for each component, Indicates the number of components involved in the reconstruction; The reconstructed signal is inversely normalized to obtain the final predicted result of the rolling bearing vibration signal.
[0035] Compared with the prior art, the present invention discloses the following technical effects: (1) This invention introduces a Hammersley sequence initialization strategy, an adaptive perturbation strategy, and a multi-difference Cauchy-Gaussian mutation strategy into the Cordyceps sinensis optimization algorithm. Compared with random initialization, the Hammersley sequence can improve the distribution uniformity of the initial population and the search space coverage; the adaptive perturbation strategy can alleviate the positional stagnation of individuals in the population; and the multi-difference Cauchy-Gaussian mutation strategy can supplement the population difference information and enhance the algorithm's ability to escape local optima. The synergistic improvement of multiple strategies enables the improved Cordyceps sinensis optimization algorithm to achieve better optimization accuracy, convergence stability, and global search performance.
[0036] (2) This invention utilizes an improved Cordyceps sinensis optimization algorithm to optimize the penalty factor and the number of modal components in variational mode decomposition. Compared with the method of relying on manual experience to set parameters, this method can improve the adaptability of variational mode decomposition to different vibration signals, reduce the mode aliasing, over-decomposition or under-decomposition phenomena caused by improper parameter selection, thereby improving the separation effect of different scale features and the signal decomposition quality in the vibration signal of rotating machinery rolling bearings.
[0037] (3) This invention utilizes an improved Cordyceps sinensis optimization algorithm to optimize the L2 regularization coefficient, learning rate, and number of hidden layer units of the bidirectional long short-term memory network corresponding to each modal component and residual component. Compared with using uniform network parameters for each signal component or relying on manual parameter tuning, this method can configure corresponding prediction hyperparameters according to the temporal characteristics of different components, reduce the impact of unreasonable parameter settings on prediction results, and thus improve the fitting ability and prediction accuracy of the prediction model for each signal component.
[0038] (4) This invention decomposes the vibration signal of a rolling bearing in rotating machinery, inputs each modal component and residual component into a corresponding bidirectional long short-term memory network for prediction, and reconstructs and denormalizes the prediction results of each signal component. This method can reduce the influence of nonlinear, non-stationary, and multi-scale coupling characteristics in the original vibration signal on direct prediction, thereby improving the accuracy and stability of the overall prediction results of the bearing vibration signal, and providing a reliable basis for monitoring the operating status and analyzing the vibration trend of rolling bearings in rotating machinery.
[0039] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This is a flowchart of the rolling bearing vibration signal prediction method based on the improved Cordyceps sinensis optimization algorithm according to an embodiment of the present invention. Figure 2 This is a flowchart of the ICFO-VMD-BiLSTM prediction model according to an embodiment of the present invention; Figure 3 is a comparison of the convergence curves of the ICFO algorithm and the comparative optimization algorithm in the CEC2017 test function according to the embodiments of the present invention. Among them, (a) is the convergence result on the CEC2017-F1 test function, (b) is the convergence result on the CEC2017-F9 test function, (c) is the convergence result on the CEC2017-F13 test function, and (d) is the convergence result on the CEC2017-F28 test function. Figure 4 This is a variational mode decomposition result diagram of the vibration signal of the drive end bearing of the centrifugal compressor according to an embodiment of the present invention; Figure 5 This is a comparison chart of the prediction errors of the ICFO-VMD-BiLSTM prediction model and the CFO-VMD-BiLSTM, SMA-VMD-BiLSTM, and GWO-VMD-BiLSTM prediction models in bearing vibration signal prediction according to an embodiment of the present invention. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0043] Example 1 This embodiment provides a method for predicting the vibration signal of rolling bearings based on an improved Cordyceps sinensis optimization algorithm, such as... Figure 1 As shown, it includes the following steps: S1. Vibration signal acquisition and preprocessing The vibration signal collected during the operation of the centrifugal compressor drive end bearing is acquired, and the vibration signal data is normalized to map the vibration signal amplitude to a preset value range to obtain an initial vibration sequence.
[0044] This embodiment uses the maximum-minimum value normalization method, and its normalization formula is as follows:
[0045] in, The first part represents the original vibration signal. One observation value, This represents the normalized signal value. and These represent the minimum and maximum values in the original signal, respectively.
[0046] The normalization process enables vibration signal data to have a uniform numerical scale, which is beneficial to improving the numerical stability and convergence performance of subsequent variational mode decomposition and bidirectional long short-term memory network training processes.
[0047] S2. Constructing an improved optimization algorithm for Cordyceps sinensis. Based on the Cordyceps sinensis optimization algorithm (CFO), an improved Cordyceps sinensis optimization algorithm (ICFO) is constructed by introducing the Hammersley sequence initialization strategy, the adaptive perturbation strategy, and the multi-difference Cauchy-Gaussian mutation strategy. For example... Figure 2 As shown, ICFO includes the initialization phase, exploration phase, parasitic phase, and mutation phase.
[0048] S2.1 Hammersley Sequence Initialization Phase Traditional CFO methods use random initialization to generate the population, which can easily lead to uneven distribution and insufficient coverage of individuals in the solution space. To improve the problem of uneven initial population distribution, this embodiment introduces a Hammersley sequence initialization strategy to initialize the population within the search space.
[0049] For any non-negative integer Prime number base ,Will Represented as Number system expansion format:
[0050] This definition is based on prime numbers. Reflection function:
[0051] Let the population size be The dimension of the parameters to be optimized is Then the first indivual The Hammersley point can be represented as:
[0052] in, Indicates the first There are several Hammersley points. Finally, the Hammersley points are mapped to the actual search space:
[0053] in, Indicates the first The individual in the first Initial position on the dimension, and The first Lower and upper bounds of a dimensional variable. .
[0054] This strategy can effectively improve the spatial coverage of the initial population and enhance the global search capability of the algorithm in the early stages.
[0055] S2.2, Exploration Stage The exploration phase simulates the behavior of Cordyceps sinensis searching for parasitic targets in the soil, and a wave-forward operator and a spiral-ascending operator are designed. Individuals in the population randomly update their positions using either the wave-forward or spiral-ascending operator, allowing them to update their positions in the search space with different movement directions and search step sizes. The wave-forward operator expands the population's search range, enhancing the global exploration capability of unknown areas; the spiral-ascending operator refines the search around the areas where high-quality individuals are located, improving the algorithm's ability to find potential high-quality solutions.
[0056] Wave advance operator update formula:
[0057]
[0058] Spiral Ascent Operator Update Formula:
[0059]
[0060] in, Indicates the exploration stage The updated location of each individual This indicates the location of the individual with the optimal fitness value during the exploration phase. and These represent the exploration stages. and the The current location of each individual. The wave disturbance factor. As a spiral regulating factor, This represents the current iteration number. The maximum number of iterations, Let be a random vector with elements in the interval [0,1]. A random number in the range [0,1]. It is a random integer that takes the value 1 or 2.
[0061] S2.3, Parasitic stage The parasitic stage simulates the parasitism and reparasitism process of Cordyceps sinensis on its host. Two individual position update methods were designed: reparasitism behavior and optimal parasitism behavior. An adaptive perturbation strategy was also integrated into this stage. Individuals in the population update their positions based on the selection probability that changes with the iteration process, using either reparasitism, optimal parasitism, or adaptive perturbation.
[0062] Reparasitic behavior update formula:
[0063] Optimal Parasitic Behavior Update Formula:
[0064]
[0065] Adaptive perturbation policy update formula:
[0066]
[0067] in, Indicates the parasitic stage The updated location of each individual This indicates the position of the individual with the optimal fitness value during the parasitic stage. and These represent the parasitic stage, number 1 and 2. and the The current location of each individual. For a random scalar that follows a standard normal distribution, The optimal parasitic perturbation factor, , Let represent the positions of two random individuals in the parasitic stage population, and let ? be a random perturbation vector following a normal distribution. This indicates taking the absolute value of each component of the vector.
[0068] S2.4, Mutation Stage After the parasitic phase, a multi-difference Cauchy-Gaussian mutation strategy is introduced. By fusing the guiding difference between the current best individual and the current individual, as well as the positional differences between multiple groups of random individuals, a well-defined multi-difference mutation direction is constructed. Cauchy perturbation or Gaussian perturbation is dynamically selected based on the iteration progress to generate mutation candidate individuals. In the early stages of iteration, the heavy-tailed characteristic of the Cauchy distribution is utilized to generate irregular mutations with a large span. In the later stages of iteration, Gaussian perturbation is used to slightly correct the individual positions, thereby supplementing population diversity and improving the algorithm's ability to escape local optima.
[0069] The formula for updating individual locations during this stage is as follows:
[0070]
[0071]
[0072] in, Indicates the parasitic stage The updated location of each individual This indicates the position of the individual with the optimal fitness value during the parasitic stage. These represent the parasitic stage, number 1 and 2. The current location of each individual. , and Let [the vector be] a random weight vector in [0,1]. The Cauchy perturbation factor, is the Gaussian perturbation factor.
[0073] S3. VMD Parameter Optimization and Signal Decomposition Based on ICFO Penalty factor for variational mode decomposition α and number of modal components K As parameters to be optimized, the improved Cordyceps sinensis optimization algorithm is used for optimization.
[0074] Specifically, the parameters to be optimized in the variational mode decomposition and their search range are determined, including: the number of modal components. and penalty factor Simultaneously, the population size and maximum number of iterations for the ICFO algorithm are set; the ICFO algorithm is used to generate the positions of individuals in the population, and each individual corresponds to a set of candidate parameter combinations consisting of a penalty factor and the number of modal components. The candidate parameter combinations are substituted into the variational mode decomposition model to decompose the initial vibration sequence, obtaining multiple corresponding modal components. During the iteration process, ICFO updates the individual positions of the population through its core stage, realizing dynamic optimization of the variational mode decomposition parameter combinations. Each iteration performs parameter updates, signal decomposition, fitness calculation, and records the parameter combination with the current optimal fitness value. When the preset optimization termination condition is met, the parameter combination with the optimal fitness value is output as the optimal variational mode decomposition parameter combination.
[0075] This embodiment uses the minimum envelope entropy as the fitness function, extracts the envelope signal of each signal component and calculates the envelope entropy, and uses this to evaluate the decomposition effect of variational mode decomposition on the initial vibration sequence under different parameter combinations. The formula for calculating the minimum envelope entropy is:
[0076]
[0077] in, Represents the minimum envelope entropy. Indicates the number of modal components. Indicates the number of signal sampling points. Indicates the first The envelope signal obtained by the Hilbert transform of the nth modal component is in the nth... The amplitude at each sampling point This represents the corresponding normalized envelope amplitude.
[0078] By using the optimization mechanism for VMD parameters based on ICFO, adaptive optimization can be achieved in terms of the number of modes and bandwidth constraint parameters, thereby improving the mode separation capability and reducing mode aliasing.
[0079] S4. Construct and decompose the VMD model. The variational mode decomposition (VMD) model is constructed based on the optimal combination of variational mode decomposition parameters obtained in step S3. VMD decomposes the input signal into several narrowband mode components with specific center frequencies by constructing a variational constraint optimization model. The specific steps are as follows: Construct a variational constrained optimization model:
[0080] in, For the first Each modal component, subscript , The number of modal components. For the first The center frequency of each modal component For the Dirac impulse function, Represents the Hilbert transform kernel function. The imaginary unit, For convolution operations, This is the initial vibration sequence. Represents the time variable The partial derivative operation, The square of the L2 norm, To express summation, Pi is a constant. For the first Each modal component at time... The value of .
[0081] In the above-mentioned constrained problem, a penalty factor is introduced. and Lagrange multipliers This transforms a constrained problem into an unconstrained problem:
[0082] in, As a penalty factor, For Lagrange multipliers, Represent the Lagrange function, For convolution operations, This indicates the inner product operation.
[0083] The alternating direction multiplier method is used for iterative solution to obtain each modal component. and its center frequency :
[0084]
[0085]
[0086] in, and They represent the first The modal component and the center frequency at the ... The update value for the next iteration. Indicates except the first The summation of all modal components other than the first modal component. The first Lagrange multiplier represents the first... iteration The first Lagrange multiplier represents the first... iteration This represents integration. , , They represent , , Representation in the frequency domain, For the initial vibration sequence at time... The value of , For the Lagrange multipliers at time The value of , For frequency variables, This is the step size parameter; The iteration stops when the following convergence condition is met:
[0087] in, The preset convergence threshold is used; Finally, several modal components and one residual component are obtained, and the sum of all components is equal to the initial vibration sequence.
[0088] S5. Optimizing BiLSTM hyperparameters using ICFO (1) Constructing samples using a sliding window The modal components and residual components are divided into training and testing sets according to time sequence, preferably in an 8:2 ratio. Training and testing samples for each signal component are constructed using a sliding window of a preset length. For any component, continuous data within the sliding window is used as the input sample, and data at the next time step after the sliding window is used as the prediction target, forming a single-step prediction sample.
[0089] (2) ICFO optimization of BiLSTM hyperparameters The improved Cordyceps sinensis optimization algorithm (ICFO) is used to iteratively optimize the hyperparameters of the bidirectional long short-term memory network (BiLSTM) corresponding to each modal component and residual component.
[0090] The hyperparameters to be optimized and their search range for the bidirectional long short-term memory network are determined. These hyperparameters include the L2 regularization coefficient, learning rate, and number of hidden layer units. The population size and maximum number of iterations for the improved Cordyceps sinensis optimization algorithm are also set. An initial population is generated using the ICFO algorithm. Each individual in the population corresponds to a set of candidate hyperparameter combinations consisting of the L2 regularization coefficient, learning rate, and number of hidden layer units. These candidate hyperparameter combinations are then used to construct BiLSTM networks for each signal component. Subsequently, the corresponding BiLSTM networks are trained using training samples from each signal component. The root mean square error (RMSE) is calculated based on the prediction results of the test samples. The RMSE is used as the fitness value to evaluate the prediction performance of different candidate hyperparameter combinations. A smaller RMSE indicates better prediction performance for the corresponding candidate hyperparameter combination. The formula for calculating the RMSE is:
[0091] in, and They represent the first The true and predicted values of each sample Indicates the number of sample data.
[0092] During the iteration process, the ICFO algorithm iteratively updates each candidate hyperparameter combination according to the population individual position update method described in step S2, and repeatedly executes network construction, model training, prediction, and fitness calculation until the maximum number of iterations is reached or the preset optimization termination condition is met. Finally, the fitness-optimal hyperparameter combination of the BiLSTM network for each signal component is output as the globally optimal prediction hyperparameter combination, and the final prediction model for each signal component is constructed based on the globally optimal prediction hyperparameter combination.
[0093] Among them, the number of hidden layer units is used to adjust the network's feature extraction and temporal information representation capabilities, the learning rate is used to adjust the update step size of the network parameters, and the L2 regularization coefficient is used to constrain the network weight size and reduce the risk of overfitting of the model.
[0094] S6. Prediction, Reconstruction, and Denormalization The modal components and residual components are respectively input into a bidirectional long short-term memory network constructed based on the optimal prediction hyperparameter combination corresponding to each component for prediction, and the prediction results of each component are obtained. The prediction results of each component are reconstructed and denormalized to obtain the complete prediction result of the bearing vibration signal.
[0095] Specifically, the test set obtained in S5 is input into the final prediction model of each signal component, and time series prediction is performed on the several modal components and residual components to obtain the prediction results corresponding to each signal component.
[0096] The prediction results of each component are reconstructed using a linear superposition method. The reconstruction formula is as follows:
[0097] in, This represents the final reconstructed predicted sequence of bearing vibration signals. Indicates the first The prediction results for each component, This indicates the number of components involved in the reconstruction.
[0098] The reconstructed signal is denormalized to map it back to the original amplitude range, thus obtaining the final predicted result of the rolling bearing vibration signal.
[0099] Example 2 1. Convergence performance verification based on CEC2017 test function To verify the optimization performance of the improved Cordyceps sinensis optimization algorithm (ICFO) constructed in this embodiment, F1, F9, F13, and F28 were selected from the CEC2017 test function set for convergence performance analysis. F1 is a unimodal function that has undergone offset and rotation processing, used to evaluate the algorithm's convergence efficiency and local exploitation ability in complex unimodal spaces; F9 is a multimodal function with highly oscillating characteristics, reflecting the algorithm's ability to maintain population diversity and avoid local optimum stagnation; F13 is a mixed function, which can be used to examine the algorithm's ability to coordinate heterogeneous search structures; and F28 is a composite function, used to verify the algorithm's stability and comprehensive solution ability in high-dimensional nonlinear optimization problems.
[0100] The ICFO algorithm was compared with six existing mainstream optimization algorithms, including the Cordyceps Sinensis Optimization (CFO), Particle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), Harris Eagle Optimization (HHO), Slime Mold Optimization (SMA), and Sparrow Search Algorithm (SSA). Simulation parameters were set as follows: optimization problem dimension was 30, population size was 50, and maximum number of iterations Max_Iter = 1000. To ensure the fairness and reliability of the experimental results, each algorithm was run independently 30 times under the same initial parameter configuration and constraints. The optimal fitness value and iteration convergence process were recorded for each run. The average convergence curve of each run was used as the evaluation criterion for the algorithm's convergence performance.
[0101] Figure 3 shows the convergence curves of the ICFO algorithm and six mainstream optimization algorithms on the selected test functions. Figures 3(a) and (b) show that the ICFO algorithm maintains a fast convergence speed and achieves high optimization accuracy in the CEC2017-F1 and F9 function tests. It demonstrates strong global search and local exploitation capabilities during the iteration process, ultimately achieving optimization results superior to most of the compared algorithms. Figures 3(c) and (d) show that the ICFO algorithm maintains a stable convergence trend even when facing complex search environments such as mixed and composite functions, exhibiting good convergence stability, optimization accuracy, and robustness, verifying the effectiveness and adaptability of the algorithm in complex optimization problems.
[0102] 2. Comprehensive performance verification based on CEC2017 and CEC2022 test functions To verify the effectiveness and stability of the improved Cordyceps sinensis optimization algorithm (ICFO) constructed in this embodiment across different types of optimization problems, typical test functions were selected from the benchmark function sets of CEC2017 and CEC2022, respectively, to construct an experimental test platform containing 20 single-objective real-valued optimization problems. The selected benchmark function set covers four types: unimodal functions, multimodal functions, mixed functions, and combined functions, which can effectively simulate the complex search space and high-dimensional nonlinear characteristics in practical engineering.
[0103] The ICFO algorithm was compared with six typical intelligent optimization algorithms, including the Cordyceps Sinensis Optimization (CFO), Particle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), Harris Eagle Optimization (HHO), Slime Mold Optimization (SMA), and Sparrow Search Algorithm (SSA). Experimental parameters were set as follows: all test functions were 10-dimensional, the population size was 50, and the maximum number of iterations was 1000. To reduce the influence of random factors on the experimental results, each algorithm was run independently 30 times on each test function, and the average fitness value of the multiple runs was calculated as the basis for evaluating algorithm performance. Based on this, the Friedman nonparametric rank-sum test was used to rank and analyze the comprehensive performance of each algorithm on different test functions.
[0104] Table 1 shows the Friedman average rank values of the ICFO algorithm and six mainstream optimization algorithms on the benchmark function set. As shown in Table 1, the ICFO algorithm achieves the lowest average rank value (1.43), ranking first among the seven algorithms and significantly outperforming the traditional CFO algorithm (3.45), which ranks third. This indicates that the ICFO algorithm has better optimization performance in the overall performance evaluation, demonstrating its superior optimization ability, stability, and adaptability to complex optimization problems.
[0105] Table 1 Ranking of Seven Algorithms
[0106] Example 3 This embodiment uses the rolling bearing vibration signal prediction method based on the improved Cordyceps sinensis optimization algorithm described in Example 1, and verifies its application using vibration signals from the drive end bearing of a centrifugal compressor collected by a certain enterprise. The specific implementation process is as follows.
[0107] 1. Data Acquisition and Preprocessing. 1200 consecutive sampling points were selected from the acquired raw bearing vibration signals as the research samples, and the sample signals were preprocessed using the maximum-minimum normalization method to obtain the normalized initial vibration sequence.
[0108] 2. VMD Parameter Optimization Based on ICFO. An improved Cordyceps sinensis optimization algorithm (ICFO) is used to optimize key parameters of Variational Mode Decomposition (VMD), including the number of modal components. K and penalty factor α The ICFO population size was set to 30, the maximum number of iterations was set to 20, and the minimum envelope entropy was used as the optimization objective function. The optimal parameter combination for VMD was obtained through iterative search, and the specific results are shown in Table 2.
[0109] Table 2 Results of VMD parameter optimization
[0110] 3. VMD Decomposition Based on Optimal Parameters. The normalized initial vibration sequence is input into the VMD model constructed based on the optimal parameter combination for decomposition, obtaining each modal component and residual component. The decomposition results are as follows: Figure 4 As shown.
[0111] 4. BiLSTM hyperparameter optimization based on ICFO.
[0112] An improved Cordyceps sinensis optimization algorithm (ICFO) was used to optimize the key hyperparameters of a bidirectional long short-term memory (BiLSTM) network. The optimized hyperparameters included the number of hidden layer units, the L2 regularization coefficient, and the learning rate. The population size of the ICFO algorithm was set to 10, and the maximum number of iterations was set to 100. The BiLSTM network was trained using training sets of each modal component (IMF) and residual components. The fitness function was minimized using the root mean square error (RMSE). The optimal combination of hyperparameters for BiLSTM was obtained through iterative optimization, and the results are shown in Table 3.
[0113] Table 3 Results of BiLSTM hyperparameter optimization
[0114] Based on the obtained optimal hyperparameter combination, a BiLSTM prediction model for each signal component is constructed for subsequent vibration signal prediction.
[0115] 5. Prediction, reconstruction, and inverse normalization of each component The test set divided by each signal component is input into the BiLSTM prediction model of each signal component constructed based on the optimal hyperparameters for time series prediction to obtain the prediction results of each signal component; the prediction results of each modal component are reconstructed, and the reconstructed signal is inversely normalized to obtain the final prediction result of the vibration signal of the compressor drive end bearing.
[0116] 6. Comparative Experiment Setup To further verify the predictive performance of the method proposed in this invention, multiple sets of comparative models were set up for experimental analysis. The comparative models include: (1) CFO-VMD-BiLSTM model: The Cordyceps sinensis optimization algorithm (CFO) is used to optimize the key parameters of VMD, and the decomposition results are input into the CFO-optimized BiLSTM network for vibration signal prediction. (2) SMA-VMD-BiLSTM model: The Slime Mold Optimization Algorithm (SMA) is used to optimize the key parameters of VMD, and the decomposition results are input into the BiLSTM network optimized by SMA for vibration signal prediction. (3) GWO-VMD-BiLSTM model: The key parameters of VMD are optimized by the Grey Wolf Optimization Algorithm (GWO), and the decomposition results are input into the BiLSTM network optimized by GWO for vibration signal prediction.
[0117] To ensure the fairness of the comparative experiments, all models were trained and tested under the same parameter settings and experimental conditions.
[0118] 7. Predictive performance evaluation To intuitively evaluate the prediction performance, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were selected. 2 ( ) are used as predictive performance evaluation metrics. Among them, RMSE measures the overall deviation between predicted and actual values; the smaller the value, the higher the overall prediction accuracy of the model. MAE characterizes the mean absolute deviation between predicted and actual values, and can intuitively reflect the overall prediction error level of the model. R 2 It is used to measure the degree of fit between the model's predictions and the actual data. The closer the value is to 1, the better the model's fit performance.
[0119] The predictive performance evaluation results of each model are shown in Table 4.
[0120] Table 4 Predictive performance evaluation results
[0121] Table 4 shows that the ICFO-VMD-BiLSTM model performs well in terms of RMSE, MAE, and R... 2 All three evaluation metrics were superior to the other comparative models. Specifically, the ICFO-VMD-BiLSTM model had an RMSE of 0.0234, which was 35.18%, 32.56%, and 43.75% lower than the CFO-VMD-BiLSTM, SMA-VMD-BiLSTM, and GWO-VMD-BiLSTM models, respectively; its MAE was 0.0188, which was 36.27%, 32.13%, and 46.29% lower than the three comparative models, respectively; R...2 It reached 0.9619, which is higher than the other three comparison models. Figure 5 This is a comparison of the prediction errors of the ICFO-VMD-BiLSTM prediction model with those of the CFO-VMD-BiLSTM, SMA-VMD-BiLSTM, and GWO-VMD-BiLSTM prediction models in bearing vibration signal prediction.
[0122] The experimental results above show that the ICFO-VMD-BiLSTM model constructed in this invention can effectively reduce the prediction error of bearing vibration signals, improve prediction accuracy and the degree of fitting to real vibration signals.
[0123] Those skilled in the art will understand that the above method can be implemented by a computer program, which can be stored in a computer-readable storage medium, and the processor implements the above method when executing the computer program.
[0124] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Those skilled in the art can flexibly select other technical features according to actual needs. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should fall within the protection scope of the claims of the present invention. To avoid obscuring the core technical solutions of the present invention, well-known technologies and conventional methods are not described in detail herein.
Claims
1. A method for predicting the vibration signal of rolling bearings based on an improved Cordyceps sinensis optimization algorithm, characterized in that, include: S1. Obtain the vibration signal collected under the operating state of the centrifugal compressor drive end bearing, and preprocess the vibration signal to obtain an initial vibration sequence; S2. Based on the Cordyceps sinensis optimization algorithm, Hammersley sequence initialization strategy, adaptive perturbation strategy and multi-difference Cauchy-Gaussian mutation strategy are introduced to improve the population initialization method and the population individual position update method, and to construct an improved Cordyceps sinensis optimization algorithm. S3. Using the penalty factor and the number of modal components of variational mode decomposition as parameters to be optimized, the improved Cordyceps sinensis optimization algorithm is used to find the optimal penalty factor and the optimal number of modal components. S4. Construct a variational mode decomposition model based on the optimal penalty factor and the optimal number of modal components, and input the initial vibration sequence into the variational mode decomposition model to obtain several modal components and residual components; S5. Using a sliding window approach, prediction samples are constructed for the modal components and residual components respectively. The L2 regularization coefficient, learning rate, and number of hidden layer units of the bidirectional long short-term memory network are used as parameters to be optimized. The improved Cordyceps sinensis optimization algorithm is used to perform optimization respectively to obtain the optimal combination of prediction hyperparameters for each component. Based on the optimal combination of prediction hyperparameters, a prediction model for each component is constructed. S6. Input the modal components and residual components into a bidirectional long short-term memory network constructed based on the optimal prediction hyperparameter combination corresponding to each component for prediction, and obtain the prediction results of each component; reconstruct and denormalize the prediction results of each component to obtain the complete prediction result of the bearing vibration signal.
2. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 1, characterized in that, Step S2, the process of constructing the improved Cordyceps sinensis optimization algorithm includes: S2.
1. The Hammersley sequence initialization strategy is used to generate an initial population in the search space; S2.2 Entering the exploration stage, the population individuals randomly update their positions using either the wave-forward operator or the spiral-ascending operator; S2.
3. Entering the parasitic stage, individuals in the population update their positions by adopting reparasitic behavior, optimal parasitic behavior, or adaptive perturbation strategy according to the selection probability. S2.4 Entering the mutation stage, a multi-difference Cauchy-Gaussian mutation strategy is adopted to construct multi-difference mutation directions, and Cauchy perturbation or Gaussian perturbation is dynamically selected according to the iteration process to generate mutation candidate individuals.
3. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 2, characterized in that, The Hammersley sequence initialization strategy in step S2.1 includes: The ratio of individual index to population size is used as the first dimension of the Hammersley sequence point. Select distinct prime numbers, expand the individual indices according to the base of each prime number, and generate the remaining coordinates of the Hammersley sequence points using the corresponding reflection function; By combining the first dimension coordinate with the other dimensions coordinates, we obtain the D-dimensional Hammersley low-difference sequence points; Mapping the coordinates of the Hammersley low-difference sequence points to the corresponding search intervals yields the initial positions of individuals in the population. The mapping formula is as follows: in, Indicates the first The individual in the first Initial position on the dimension, Indicates the first Hammersley point at the j Coordinates on the dimension and The first Lower and upper bounds of a dimensional variable. .
4. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 2, characterized in that, In the exploration phase described in step S2.2, the update formulas for the wave advance operator and the spiral ascent operator are as follows: Wave forward operator: Spiral Ascent Operator: in, Indicates the exploration stage The updated location of each individual This indicates the location of the individual with the optimal fitness value during the exploration phase. and These represent the exploration stages. and the The current location of each individual. The wave disturbance factor. As a spiral regulating factor, This represents the current iteration number. The maximum number of iterations, Let be a random vector with elements in the interval [0,1]. A random number in the range [0,1]. It is a random integer that takes the value 1 or 2.
5. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 2, characterized in that, In the parasitic phase of step S2.3, the update formulas for the reparasitic behavior, optimal parasitic behavior, and adaptive perturbation strategy are as follows: Reparasitic behavior update formula: Optimal Parasitic Behavior Update Formula: in, Indicates the parasitic stage The updated location of each individual This indicates the position of the individual with the optimal fitness value during the parasitic stage. and These represent the parasitic stage, number 1 and 2. and the The current location of each individual. For a random scalar that follows a standard normal distribution, The optimal parasitic perturbation factor, δ It is a random perturbation vector that follows a normal distribution; The update formula for the adaptive perturbation strategy is: in, , Let represent the positions of two random individuals in the parasitic stage population, and let ? be a random perturbation vector following a normal distribution. This indicates taking the absolute value of each component of the vector.
6. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 2, characterized in that, In step S2.4, the update formula for the multi-difference Cauchy-Gaussian mutation strategy is: in: in, Indicates the parasitic stage The updated location of each individual This indicates the position of the individual with the optimal fitness value during the parasitic stage. These represent the parasitic stage, number 1 and 2. The current location of each individual. , and Let [the vector be] a random weight vector in [0,1]. The Cauchy perturbation factor, is the Gaussian perturbation factor.
7. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 1, characterized in that, Step S3, which involves using the improved Cordyceps sinensis optimization algorithm for optimization, includes: The penalty factor and the number of modal components of variational mode decomposition are used as parameters to be optimized. Using the minimum envelope entropy as the fitness function, the formula for calculating the minimum envelope entropy is as follows: in, Represents the minimum envelope entropy. Indicates the number of modal components. Indicates the number of signal sampling points. Indicates the first The envelope signal obtained by the Hilbert transform of the nth modal component is in the nth... The amplitude at each sampling point This represents the corresponding normalized envelope amplitude. An improved Cordyceps sinensis optimization algorithm is used for iterative optimization to determine the parameter combination with the best fitness as the optimal variational mode decomposition parameter combination. A variational mode decomposition model is then constructed based on this optimal combination of parameter values.
8. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 1, characterized in that, Step S4, which involves inputting the initial vibration sequence into the variational mode decomposition model to obtain several modal components and residual components, includes: A variational constrained optimization model is constructed, and a penalty factor and Lagrange multipliers are introduced to transform the variational constrained optimization model into an unconstrained form. The alternating direction multiplier method is used to iteratively update each modal component, center frequency, and Lagrange multiplier until the convergence condition is met, resulting in several modal components and residual components.
9. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 1, characterized in that, The hyperparameters of the bidirectional long short-term memory network described in step S5 are optimized using an improved Cordyceps sinensis optimization algorithm, including: The L2 regularization coefficient, learning rate, and number of hidden layer units of the bidirectional long short-term memory network are used as hyperparameters to be optimized. A sliding window approach is used to construct training and test samples for each modal component and residual component; The root mean square error is used as the fitness function; An improved Cordyceps sinensis optimization algorithm was used to iteratively optimize the hyperparameters of each signal component to obtain the optimal combination of prediction hyperparameters for each signal component. Based on the optimal combination of prediction hyperparameters for each signal component, a corresponding bidirectional long short-term memory network prediction model was constructed.
10. The method for predicting rolling bearing vibration signals based on an improved Cordyceps sinensis optimization algorithm according to claim 1, characterized in that, In step S6, the step of inputting each signal component into a bidirectional long short-term memory network for prediction and reconstructing the prediction results of each signal component includes: Each modal component and residual component is input into a bidirectional long short-term memory network constructed based on its optimal prediction hyperparameter combination to obtain the prediction results of each signal component. The prediction results of each signal component are reconstructed using a linear superposition method. The reconstruction formula is as follows: in, This represents the final reconstructed predicted sequence of bearing vibration signals. Indicates the first The prediction results for each component, Indicates the number of components involved in the reconstruction; The reconstructed signal is inversely normalized to obtain the final predicted result of the rolling bearing vibration signal.
