Bearing fault early warning method and system based on IWAOA improved SVMD denoising and LSTM-CNN classification

Abstract

The application discloses a bearing fault early warning method and system based on IWAOA improved SVMD noise reduction and LSTM-CNN classification, and the optimal decomposition layer number and penalty coefficient of successive variation modal decomposition SVMD are adaptively obtained by using IWAOA, the bearing vibration signal is decomposed by SVMD, the signal is reconstructed and denoised by screening effective modal components through a correlation coefficient, the acceleration peak value and kurtosis index of the reconstructed signal are extracted to form a feature vector, after dimension reduction by PCA, the feature vector is input into an LSTM_CNN classification model combined with a long short-term memory network and a convolutional neural network for training and classification, the time sequence dependence is captured, and the accurate identification of the normal and fault states of the bearing is realized. The application has better denoising effect, faster calculation speed, and the bearing fault early warning accuracy reaches 100%, and is suitable for fault early warning of different types of bearings and under different working conditions.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery bearing fault diagnosis technology, and in particular to a bearing fault early warning method and system based on the improved whale optimization algorithm IWAOA for successive variational mode decomposition SVMD noise reduction, PCA dimensionality reduction by fusing acceleration peak and kurtosis indices, and classification by combining an LSTM-CNN deep learning model. Background Technology

[0002] Bearings are core components in rotating machinery, widely used in critical equipment such as wind turbines, high-speed trains, aero engines, and electric motors in steel metallurgical rolling mills. The operating condition of bearings directly affects the safety and reliability of the entire equipment. However, due to the long-term effects of alternating loads, high-speed rotation, and harsh working environments, bearings are highly susceptible to wear, cracks, pitting, and other failures. Failure to detect and warn of early bearing failures in a timely manner can lead to equipment downtime or even severe damage, causing significant economic losses and safety hazards. Therefore, real-time and accurate bearing fault warnings have significant engineering value and economic importance.

[0003] Vibration signal analysis is currently the most commonly used and effective method in bearing fault diagnosis. When a bearing fails, its vibration signal exhibits periodic impact components, and these impact features contain rich fault information. However, in actual industrial settings, due to interference from factors such as complex equipment structures, high environmental noise, and multiple transmission paths, the weak impact signals generated by bearing faults are often submerged in strong background noise, making it difficult to extract and identify fault features. Therefore, effectively eliminating noise interference and extracting fault impact features is a key technical challenge for achieving accurate early warning of bearing faults.

[0004] Successive Variational Mode Decomposition (SVMD) is an adaptive signal decomposition method that has emerged in recent years. It iteratively extracts modal components from a signal. Compared to traditional Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD), SVMD offers better noise robustness and mode separation. However, the decomposition performance of the SVMD algorithm heavily relies on two key parameters: the number of decomposition levels K and the penalty coefficient alpha. In practical applications, these two parameters often need to be set manually based on experience, lacking adaptive optimization capabilities. SVMD decomposition algorithms often experience performance degradation due to inaccurate decomposition mode counts, thus limiting its application in bearing fault diagnosis.

[0005] In terms of feature extraction, peak acceleration is an effective indicator of signal impact intensity, and kurtosis is highly sensitive to the impact component in the signal, making it a classic feature in bearing fault diagnosis. However, a single feature is often insufficient to fully characterize the essential properties of the fault.

[0006] In terms of classification models, existing Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) are both difficult to effectively integrate temporal and spatial features.

[0007] In summary, existing bearing fault early warning technologies still suffer from the following technical problems: SVMD decomposition parameters rely on manual setting and lack adaptive optimization capabilities; single features are insufficient to comprehensively characterize fault features; and traditional classification methods struggle to effectively integrate temporal and spatial features. Therefore, there is an urgent need for a bearing fault early warning method that can adaptively optimize SVMD parameters, integrate multi-dimensional features, and combine deep learning classification. Summary of the Invention

[0008] The technical problem to be solved by this invention is: addressing the issues of existing bearing fault early warning technologies, such as SVMD decomposition parameters relying on manual experience settings, poor noise reduction effect, incomplete feature extraction, and low classification accuracy, by providing a bearing fault early warning method and system based on IWAOA-improved SVMD noise reduction and LSTM-CNN classification, thereby achieving accurate early warning of bearing faults.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A bearing fault early warning method based on IWAOA improved SVMD denoising and LSTM-CNN classification includes the following steps: S1: Obtain the raw acceleration vibration signal dataset of the bearing, which includes normal data and fault data; S2: The improved whale optimization algorithm IWAOA is used to adaptively optimize the decomposition parameters of SVMD. The fitness function is to minimize the envelope entropy, and the optimal number of decomposition layers and the optimal penalty coefficient of successive variational mode decomposition SVMD are obtained. S3: Perform SVMD decomposition on the dataset according to the optimal decomposition level and the optimal penalty coefficient to obtain several basic modal components (IMFs), calculate the correlation coefficient between each IMF and the original signal, and add up the IMFs with correlation coefficients greater than a preset threshold to obtain the reconstructed signal. S4: Calculate the peak acceleration and kurtosis index of the reconstructed signal to form an eigenvalue vector; S5: Reduce the eigenvalue vector to one dimension using principal component analysis (PCA) to obtain the dimensionality-reduced eigenvalues; S6: Label the reduced feature values ​​as normal and faulty respectively, and input them into the LSTM-CNN classification model that combines a long short-term memory network and a convolutional neural network for training and classification, and output the probability value of the bearing being in a normal or faulty state. S7: Classify the bearing's normal and fault characteristics based on probability values, and trigger a fault warning when the probability of a fault condition exceeds a preset probability warning threshold.

[0010] This invention proposes an adaptive optimization method for SVMD parameters based on an improved IWAOA: using the minimization of envelope entropy as the fitness function, the improved Whale Optimization Algorithm (IWAOA) adaptively searches for the optimal decomposition level K and penalty coefficient alpha of SVMD within the search space. IWAOA initializes the population using a Circle chaotic mapping and introduces adaptive inertia weights to balance global search and local exploitation capabilities. Compared to the standard WOA, it exhibits faster convergence speed and stronger optimization ability, effectively overcoming the limitations of manually setting SVMD parameters.

[0011] The present invention proposes an adaptive noise reduction scheme based on correlation coefficient thresholds: after performing SVMD decomposition based on the optimized parameters, the Pearson correlation coefficient between each IMF component and the original signal is calculated, and the effective modal components are screened by setting a threshold for signal reconstruction, thereby achieving the purpose of adaptively extracting fault impact features from noise signals.

[0012] The present invention uses PCA dimensionality reduction to fuse features: it fuses the peak acceleration and kurtosis index to construct a two-dimensional feature vector, and then projects the two-dimensional features onto a one-dimensional principal component through principal component analysis (PCA). This not only preserves the key information of the original features, but also reduces the feature dimensionality, thereby reducing the computational burden of subsequent classification models.

[0013] The present invention employs an LSTM-CNN fusion classification scheme, which organically combines the spatial feature extraction capability of CNN with the temporal modeling capability of LSTM to achieve efficient identification and accurate classification of bearing fault features.

[0014] In the above technical solution, the specific optimization process of the improved whale optimization algorithm IWAOA in step S2 is as follows: the whale population is initialized using Circle chaotic mapping to enhance population diversity; an adaptive inertial weight is introduced into the whale position update formula, and the adaptive inertial weight increases with the number of iterations; the optimal decomposition layer K and penalty coefficient alpha of SVMD are found within the set search range using the minimum of envelope entropy as the fitness function.

[0015] In the above technical solution, the preset threshold in step S2 is 0.3.

[0016] That is, the IMF components with a correlation coefficient greater than 0.3 with the original signal are selected and added together for reconstruction, while the noise-dominant components with a correlation coefficient less than or equal to 0.3 are discarded.

[0017] In the above technical solution, the specific construction method of the feature vector in step S3 is as follows: calculate the acceleration peak value and kurtosis index of all reconstructed signals of normal bearings and faulty bearings respectively, form a normal feature vector by combining the acceleration peak value and kurtosis index of normal bearing data, and form a fault feature vector by combining the acceleration peak value and kurtosis index of faulty bearing data.

[0018] In the above technical solution, the network structure of the LSTM-CNN classification model in step S5 includes: a CNN layer for extracting local spatial features of the input feature data, the CNN layer containing at least one-dimensional convolutional layers and pooling layers; an LSTM layer for capturing temporal dependencies in the feature data, the LSTM layer containing at least one LSTM unit; a fully connected layer and a Softmax classification layer for outputting classification results; the CNN layer is placed before the LSTM layer, the dimensionality-reduced feature values ​​are input into the CNN layer in time series form, and after feature extraction, they are sent to the LSTM layer for temporal modeling, and finally the fully connected layer and the Softmax layer complete the binary classification of normal and fault states.

[0019] In the above technical solution, in step S6, the LSTM-CNN classification model extracts local spatial features of the feature data through the CNN layer, captures the temporal dependencies in the feature data through the LSTM layer, and outputs the probability value of whether the bearing belongs to the normal state or the fault state through the fully connected layer and the Softmax classification layer.

[0020] In the above technical solution, the specific process of PCA dimensionality reduction in step S4 is as follows: the two-dimensional feature vector composed of acceleration peak and kurtosis index is decentered, the covariance matrix is ​​calculated, the eigenvalues ​​and eigenvectors of the covariance matrix are obtained, the eigenvector corresponding to the largest eigenvalue is selected as the projection direction, and the two-dimensional features are projected onto the one-dimensional principal components to obtain the dimensionality-reduced one-dimensional feature values.

[0021] In the above technical solution, the raw acceleration vibration signal dataset obtained in step S1 is obtained by sampling the acceleration sensor installed on the bearing housing. The sampling frequency is not less than 10 times the bearing fault characteristic frequency. The dataset contains at least 10 sets of normal bearing data and 10 sets of faulty bearing data, with each set of data corresponding to a vibration signal with a duration of 1 second.

[0022] A bearing fault early warning system based on IWAOA-improved SVMD noise reduction and LSTM-CNN classification includes the following modules: Signal acquisition module: used to acquire the raw acceleration vibration signal dataset of the bearing, which includes normal data and fault data; Parameter optimization module: Used to analyze signals using the improved whale optimization algorithm IWAOA, with the minimum of envelope entropy as the fitness function, to obtain the optimal number of decomposition layers and the optimal penalty coefficient of successive variational mode decomposition (SVMD); Signal decomposition module: used to perform SVMD decomposition on the signal according to the optimal decomposition level and the optimal penalty coefficient, to obtain several basic modal components (IMFs); Signal reconstruction module: used to calculate the correlation coefficient between each IMF and the original signal, and to add up the IMFs with correlation coefficients greater than a preset threshold to obtain the reconstructed signal; Feature extraction module: used to calculate the peak acceleration and kurtosis index of the reconstructed signal, and form a feature value vector; Dimensionality reduction module: used to reduce the eigenvalue vector to one dimension through principal component analysis (PCA) to obtain the dimensionality-reduced eigenvalues; The classification and early warning module is used to label the normal and fault feature values ​​after dimensionality reduction, input them into the LSTM-CNN classification model for training and classification, and output the bearing fault early warning result. The LSTM-CNN classification model comprises a sequentially connected input layer, a CNN feature extraction layer, an LSTM temporal modeling layer, a fully connected layer, and a Softmax output layer. The CNN feature extraction layer includes at least one-dimensional convolutional layers and pooling layers to extract local spatial features from the dimensionality-reduced feature data. The LSTM temporal modeling layer includes at least one LSTM unit to capture temporal dependencies in the feature data. The Softmax output layer outputs the probability value of the bearing belonging to a normal or faulty state, and triggers a fault warning when the probability of a faulty state exceeds a preset warning threshold.

[0023] In the above technical solution, the training process of the LSTM-CNN classification model includes: arranging the dimensionality-reduced feature values ​​in time series and inputting them into the CNN layer, extracting local spatial feature patterns through one-dimensional convolutional kernels; inputting the feature sequences extracted by the CNN into the LSTM layer, capturing long-term temporal dependencies through a gating mechanism; flattening the output features of the LSTM and sending them into a fully connected layer for nonlinear transformation; finally, the Softmax layer outputs the probability distribution of normal and fault states, and the category corresponding to the maximum probability is taken as the classification result.

[0024] A computer-readable storage medium storing a program that, when executed by a processor, implements the steps of any of the methods described above.

[0025] Compared with the prior art, the beneficial effects of this invention are: (1) Better noise reduction and faster calculation speed: By adaptively optimizing the decomposition parameters of SVMD through IWAOA, the blindness of manual trial and error is avoided, making the decomposition results more accurate and reasonable; the IMF screening strategy based on the correlation coefficient threshold can effectively remove noise interference and retain the fault impact characteristics. Compared with the unimproved SVMD, the method of the present invention pre-sets the number of modes and the penalty coefficient, resulting in lower computational complexity, and the noise-reduced signal can clearly extract the impact characteristics of bearing faults, and the fault information is completely preserved.

[0026] (2) Higher bearing warning accuracy: By fusing acceleration peak and kurtosis indices and performing PCA dimensionality reduction, the optimal feature representation was constructed; by adopting a classification method combining LSTM and CNN, the spatial feature extraction capability of CNN and the temporal modeling capability of LSTM were fully utilized, achieving efficient differentiation between normal bearing signals and fault signals, with a warning accuracy of 100%.

[0027] (3) Strong adaptability: The entire method process, from parameter optimization, signal decomposition, feature extraction to classification and early warning, has achieved adaptive processing without manual intervention. It is suitable for fault early warning of different types of bearings and different working conditions. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is an overall flowchart of the bearing fault early warning method of the present invention.

[0029] Figure 2 This is a time-domain waveform diagram (original signal) of the original vibration signal of the bearing fault in an embodiment of the present invention.

[0030] Figure 3 The above is a time-domain waveform diagram of the signal after SVMD noise reduction improved by IWAOA in an embodiment of the present invention (noise-reduced signal).

[0031] Figure 4 This is the envelope spectrum (original envelope signal) of the original signal in an embodiment of the present invention.

[0032] Figure 5 This is the envelope spectrum of the noise-reduced signal (noise-reduced envelope signal) in an embodiment of the present invention.

[0033] Figure 6 This is a confusion matrix diagram of the LSTM-CNN classification model in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] Example 1 like Figure 1 As shown, the bearing fault early warning method based on IWAOA improved SVMD denoising, fusion of peak acceleration and kurtosis PCA dimensionality reduction and LSTM-CNN classification specifically includes the following steps: Step S1: Obtain the raw acceleration vibration signal dataset of the bearing, divided into normal bearing data and fault data. Vibration signals of the bearing during operation are collected using acceleration sensors mounted on the bearing housing. The sampling frequency is determined based on the bearing's rotational speed and the fault characteristic frequency, generally not less than 10 times the fault characteristic frequency. The collected signals are divided into normal bearing datasets and fault bearing datasets, which are used for subsequent model training and testing, respectively.

[0036] Step S2: Analyze the signal X for each second in the dataset using IWAOA to obtain the optimal decomposition level K and penalty coefficient alpha for SVMD decomposition. IWAOA uses the minimum envelope entropy as the fitness function, initializes the whale population position using Circle chaotic mapping, and introduces an adaptive inertia weight that decreases nonlinearly with the number of iterations to balance the algorithm's exploration and development capabilities. Within the set search range, IWAOA iteratively optimizes and finally outputs the optimal parameter combination (K, alpha) that minimizes the envelope entropy.

[0037] Step S3: Perform SVMD decomposition on the signal according to the optimal parameters to obtain several IMFs, calculate the correlation coefficient between each IMF and the original signal, and screen the effective IMF components through the correlation coefficients to reconstruct the signal and achieve adaptive noise reduction.

[0038] The reconstructed signal is obtained by adding together IMF values ​​greater than 0.3.

[0039] Step S4: Calculate the peak acceleration and kurtosis index of the reconstructed signal to form an eigenvalue vector. The peak acceleration reflects the maximum intensity of the impact component in the signal, and the kurtosis index measures the sharpness of the peak in the signal probability distribution. Calculate the peak acceleration and kurtosis index for all reconstructed signals of both normal and faulty bearings, and construct normal and faulty eigenvalue vectors respectively.

[0040] Step S5: Reduce the eigenvalue vector to one dimension using PCA. PCA projects the original two-dimensional features (acceleration peaks and kurtosis) onto the one-dimensional principal component direction with the largest variance through a linear transformation, achieving feature dimensionality reduction while preserving the main information. Specifically, the two-dimensional eigenvalue vector is centered, the covariance matrix is ​​calculated, eigenvalues ​​and eigenvectors are obtained, and the eigenvector corresponding to the largest eigenvalue is selected as the projection direction, projecting the two-dimensional features onto the one-dimensional principal component.

[0041] Step S6: Label the reduced normal and fault feature values ​​(normal label is 1, fault label is 2) and input them into the LSTM-CNN classification model for training and classification. The LSTM-CNN classification model extracts local spatial features from the input feature data through CNN layers, captures temporal dependencies in the feature data through LSTM layers, and outputs the probability value of whether the bearing belongs to a normal or faulty state through fully connected layers and a Softmax classification layer. This LSTM-CNN classification step is the core invention of this invention, achieving efficient identification of bearing fault features through the organic combination of the spatial feature extraction capability of CNN and the temporal modeling capability of LSTM.

[0042] Step S7: Classify the bearing's normal and fault features according to the output probability value of the LSTM-CNN classification model. When the probability of the fault state exceeds the preset warning threshold, trigger a fault warning to complete the warning of bearing faults.

[0043] Example 2 The following example uses the fault warning of the load-side bearing of the female screw of a discarded wind turbine as a specific case, combined with... Figures 2 to 6 The technical solution of the present invention will be described in detail below.

[0044] Input 10 datasets each for normal and faulty conditions of the load-side bearing of the abandoned fan screw, for a total of 20 datasets. The sampling frequency is set to an appropriate value to cover the main fault characteristic frequencies of the bearing. Each dataset corresponds to a 1-second acceleration vibration signal. The time-domain waveforms of some of the original signals are shown below. Figure 2 As shown in the figure, the original signal is subject to strong noise interference, and the impact characteristics are not obvious enough, making it difficult to directly identify fault information.

[0045] The original signal was analyzed computationally, and the optimal decomposition level K and penalty coefficient alpha for SVMD decomposition were obtained using IWAOA. IWAOA was used to set a reasonable population size and maximum number of iterations, employing envelope entropy as the fitness function to optimize within the search range. Circle chaotic mapping initialization resulted in a more uniform population distribution, and adaptive inertia weights enabled the algorithm to focus on global exploration in the early stages and fine-grained local search in the later stages. Through iterative optimization, the optimal SVMD parameters for this dataset were obtained.

[0046] After performing SVMD decomposition based on the optimal parameters, the correlation coefficient between each IMF and the original signal is calculated. IMF components with a correlation coefficient greater than 0.3 are selected and added together to reconstruct the signal, thus obtaining the denoised signal.

[0047] like Figure 3 As shown in the noise-reduced signal, the impact characteristics of the bearing fault signal can be clearly extracted from the noise-reduced signal, the periodic impact components are clearly visible, and the influence of noise is significantly reduced.

[0048] Figure 4 and Figure 5 The envelope spectra of the original signal and the denoised signal are shown respectively. From Figure 4 (Original envelope signal) It can be seen that the peak value of the fault characteristic frequency (around X=14Hz) in the envelope spectrum of the original signal is disturbed by noise, with an amplitude of approximately 0.759; while Figure 5 In the (noise-reduced envelope signal), the peak value of the fault characteristic frequency (around X=13Hz) in the noise-reduced envelope spectrum is more prominent, with an amplitude of about 0.836, and the fault characteristic frequencies are basically consistent, indicating that the noise-reduced signal not only removes noise, but also completely preserves the bearing fault information.

[0049] The peak acceleration and kurtosis indices of the denoised signals from the normal and faulty bearing datasets are extracted, and the results are shown in Tables 1 and 2.

[0050] Table 1. Peak Acceleration Characteristic Values

[0051] Table 2. Characteristic values ​​of kurtosis index

[0052] As can be seen from Tables 1 and 2, the peak acceleration of the faulty bearing (mean value of about 5.4) is much higher than that of the normal bearing (mean value of about 0.027), and the kurtosis index of the faulty bearing (mean value of about 3.42) is also significantly higher than that of the normal bearing (mean value of about 3.0), indicating that these two characteristic indicators can effectively distinguish between normal and faulty states.

[0053] The calculated peak normal acceleration and kurtosis index are combined to form a feature vector `value1`, and the peak fault acceleration and kurtosis index are combined to form a feature vector `value2`. Feature vector `value1` is then reduced to one dimension (`value_normal`) using PCA, and feature vector `value2` is further reduced to one dimension (`value_fault`) using PCA. PCA dimensionality reduction removes redundant information between features, making subsequent classification more efficient.

[0054] Label `value_normal` as label 1 and `value_fault` as label 2, and then input them into the LSTM-CNN for training and classification.

[0055] like Figure 6 The CNN-LSTM confusion matrix shown demonstrates a 100% accuracy rate in classifying bearings as normal or faulty. In the confusion matrix, all 10 samples of the true class 1 (normal) are correctly predicted as class 1, and all 10 samples of the true class 2 (faulty) are correctly predicted as class 2, with no misclassifications. The validation results demonstrate that the method of this invention can effectively identify bearing fault characteristics, thereby enabling accurate early warning of bearing faults and preventing significant economic losses to equipment.

[0056] In summary, the bearing fault early warning method proposed in this invention, based on IWAOA-improved SVMD denoising, fusion of peak acceleration and kurtosis PCA dimensionality reduction, and LSTM-CNN classification, achieves high-precision early warning of bearing faults through the organic combination of intelligent optimization of decomposition parameters, adaptive denoising, fusion of feature extraction, and deep learning classification. It has strong engineering practicality and application value.

[0057] Those skilled in the art can improve or modify the invention based on the above description and in conjunction with existing technology, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A bearing fault early warning method based on IWAOA-improved SVMD denoising and LSTM-CNN classification, characterized in that, Includes the following steps: S1: Obtain the raw acceleration vibration signal dataset of the bearing, which includes normal data and fault data; S2: The improved whale optimization algorithm IWAOA is used to adaptively optimize the decomposition parameters of SVMD, and obtain the optimal number of decomposition layers and the optimal penalty coefficient of successive variational mode decomposition SVMD. S3: Perform SVMD decomposition on the dataset according to the optimal decomposition level and the optimal penalty coefficient to obtain several basic modal components (IMFs), calculate the correlation coefficient between each IMF and the original signal, and add up the IMFs with correlation coefficients greater than a preset threshold to obtain the reconstructed signal. S4: Calculate the peak acceleration and kurtosis index of the reconstructed signal to form an eigenvalue vector; S5: Reduce the eigenvalue vector to one dimension using principal component analysis (PCA) to obtain the dimensionality-reduced eigenvalues; S6: Label the reduced feature values ​​as normal and faulty respectively, and input them into the LSTM-CNN classification model that combines a long short-term memory network and a convolutional neural network for training and classification, and output the probability value of the bearing being in a normal or faulty state. S7: Classify the bearing's normal and fault characteristics based on probability values, and trigger a fault warning when the probability of a fault condition exceeds a preset probability warning threshold.

2. The bearing fault early warning method according to claim 1, characterized in that, The specific optimization process of the improved whale optimization algorithm IWAOA in step S2 is as follows: the whale population is initialized using the Circle chaotic mapping to enhance the diversity of the population; an adaptive inertial weight is introduced into the whale position update formula, and the adaptive inertial weight increases with the number of iterations; the fitness function is minimized by the envelope entropy, and the optimal decomposition layer K and penalty coefficient alpha of SVMD are found within the set search range.

3. The bearing fault early warning method according to claim 1, characterized in that, The preset threshold in step S2 is 0.

3.

4. The bearing fault early warning method according to claim 1, characterized in that, The specific construction method of the feature vector in step S4 is as follows: calculate the acceleration peak value and kurtosis index for all reconstructed signals of normal bearings and faulty bearings respectively, form a normal feature vector by combining the acceleration peak value and kurtosis index of normal bearing data, and form a fault feature vector by combining the acceleration peak value and kurtosis index of faulty bearing data.

5. The bearing fault early warning method according to claim 1, characterized in that, In step S6, the LSTM-CNN classification model extracts local spatial features of the feature data through the CNN layer, captures the temporal dependencies in the feature data through the LSTM layer, and outputs the probability value of the bearing belonging to the normal state or the fault state through the fully connected layer and the Softmax classification layer.

6. The bearing fault early warning method according to claim 1, characterized in that, The specific process of PCA dimensionality reduction in step S4 is as follows: the two-dimensional feature vector composed of acceleration peak and kurtosis index is decentered, the covariance matrix is ​​calculated, the eigenvalues ​​and eigenvectors of the covariance matrix are obtained, the eigenvector corresponding to the largest eigenvalue is selected as the projection direction, and the two-dimensional features are projected onto the one-dimensional principal components to obtain the dimensionality-reduced one-dimensional feature values.

7. The bearing fault early warning method according to claim 1, characterized in that, In step S1, the sampling frequency of the original acceleration vibration signal is not less than 10 times the bearing fault characteristic frequency. The original acceleration vibration signal dataset contains at least 10 sets of normal bearing data and 10 sets of faulty bearing data, with each set of data corresponding to a vibration signal with a duration of 1 second.

8. A bearing fault early warning system based on IWAOA-improved SVMD noise reduction and LSTM-CNN classification, characterized in that, Includes the following modules: Signal acquisition module: used to acquire the raw acceleration vibration signal dataset of the bearing, which includes normal data and fault data; Parameter optimization module: Used to analyze signals using the improved whale optimization algorithm IWAOA, with the minimum of envelope entropy as the fitness function, to obtain the optimal number of decomposition layers and the optimal penalty coefficient of successive variational mode decomposition (SVMD); Signal decomposition module: used to perform SVMD decomposition on the signal according to the optimal decomposition level and the optimal penalty coefficient, to obtain several basic modal components (IMFs); Signal reconstruction module: used to calculate the correlation coefficient between each IMF and the original signal, and to add up the IMFs with correlation coefficients greater than a preset threshold to obtain the reconstructed signal; Feature extraction module: used to calculate the peak acceleration and kurtosis index of the reconstructed signal, and form a feature value vector; Dimensionality reduction module: used to reduce the eigenvalue vector to one dimension through principal component analysis (PCA) to obtain the dimensionality-reduced eigenvalues; The classification and early warning module is used to label the normal and fault feature values ​​after dimensionality reduction, input them into the LSTM-CNN classification model for training and classification, and output the bearing fault early warning result. The LSTM-CNN classification model comprises a sequentially connected input layer, a CNN feature extraction layer, an LSTM temporal modeling layer, a fully connected layer, and a Softmax output layer. The CNN feature extraction layer includes at least one-dimensional convolutional layers and pooling layers to extract local spatial features from the dimensionality-reduced feature data. The LSTM temporal modeling layer includes at least one LSTM unit to capture temporal dependencies in the feature data. The Softmax output layer outputs the probability value of the bearing belonging to a normal or faulty state, and triggers a fault warning when the probability of a faulty state exceeds a preset warning threshold.

9. The bearing fault early warning system based on IWAOA-improved SVMD noise reduction and LSTM-CNN classification according to claim 8, characterized in that, The training process of the LSTM-CNN classification model includes: arranging the dimensionality-reduced feature values ​​in time series and inputting them into the CNN layer, extracting local spatial feature patterns through one-dimensional convolutional kernels; inputting the feature sequences extracted by the CNN into the LSTM layer, capturing long-term temporal dependencies through a gating mechanism; flattening the output features of the LSTM and feeding them into a fully connected layer for nonlinear transformation; finally, the Softmax layer outputs the probability distributions of normal and fault states, and the category corresponding to the maximum probability is taken as the classification result.

10. A computer-readable storage medium storing a program thereon, characterized in that: When the program is executed by a processor, it implements the steps of the method described in any one of claims 1-7.

Images