A packaging machine bearing fault classification method based on convolutional neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA TOBACCO HENAN IND CO LTD
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-19
AI Technical Summary
In existing technologies, the fault diagnosis of bearings in cigarette production packaging machines suffers from problems such as strong noise interference, low signal-to-noise ratio, unbalanced datasets, and low diagnostic accuracy, making it difficult to accurately identify fault characteristics through one-dimensional time-domain and frequency-domain signal processing.
A convolutional neural network-based approach is adopted, which extracts the time-frequency domain features of fault signals through variational mode decomposition and continuous wavelet transform, and combines data augmentation techniques to train the convolutional neural network for fault classification.
This improves the accuracy and generalization performance of bearing fault diagnosis, enabling more efficient fault identification and classification.
Smart Images
Figure CN122241379A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cigarette manufacturing technology, and more specifically, to a method for classifying packaging machine bearing faults based on convolutional neural networks. Background Technology
[0002] In automated cigarette production and packaging machinery, rotating structures constitute a major part. Over prolonged use, they inevitably experience malfunctions, leading to increased production and maintenance costs. Key components of rotating machinery, such as rolling bearings and gears, frequently exhibit early defects, such as minor wear and pitting. Bearings are an indispensable critical component in rotating machinery, and the operating condition of the rotating equipment largely depends on the condition of the bearings. However, currently, cigarette factory operators typically rely on their experience to judge the bearing's operating condition and malfunction type, significantly impacting the accuracy of judgment and production efficiency.
[0003] In the prior art, (1) the working environment of the packaging machine bearing is complex, with strong noise interference, low signal-to-noise ratio, and poor identification. The fault characteristic signal of the target bearing is usually mixed with other source signals, making it difficult to identify and extract the fault characteristics from the time domain diagram and spectrum diagram of the original vibration signal.
[0004] (2) In the actual working conditions of the packaging machine, the data collected by the vibration sensor is more in the normal operation state and less in the fault state. The imbalance of the dataset leads to the poor effect of the diagnosis method based on the neural network.
[0005] (3) Bearing fault diagnosis usually uses one-dimensional time domain and frequency domain signal processing methods. One-dimensional signals contain fewer fault feature points, resulting in lower diagnostic accuracy. Summary of the Invention
[0006] The purpose of this invention is to provide a method for classifying packaging machine bearing faults based on convolutional neural networks, in order to overcome the shortcomings of existing technologies.
[0007] To achieve the above objectives, this application employs the following technical solution:
[0008] A method for classifying bearing faults in packaging machines based on convolutional neural networks, comprising the following steps:
[0009] S1. Obtain the operating data of the packaging machine bearing to obtain the vibration signal dataset of the normal operating state and different fault types of the packaging machine bearing, as a data sample for establishing the model;
[0010] S2. For the data samples established in step S1, determine the optimal decomposition K value of the improved variational mode decomposition method in the data sample preprocessing by using the principle of minimizing the average Pearson correlation coefficient.
[0011] S3. Based on the optimal decomposition K value obtained in step S2, extract the time-domain features of the fault signal, use variational mode decomposition to preprocess the acquired original signal dataset, decompose the original vibration signal into K component signals, perform correlation analysis between the K component signals and the original signal, select the highly correlated components for reconstruction, remove the interference of redundant components, and obtain an effective one-dimensional time-domain fault signal.
[0012] S4. For the effective one-dimensional time-domain fault signals obtained in step S3, the overlapping sampling method is used to increase the number of training samples, thereby expanding the training samples and fully extracting the fault information contained in each fault type.
[0013] S5. Extract the time-frequency domain features of the fault signal and use the continuous wavelet transform method to convert the expanded one-dimensional time-domain fault signal into a two-dimensional time-frequency domain fault image.
[0014] S6. For the two-dimensional time-frequency domain fault image obtained in step S5, adjust the image size, divide the image samples into training set, validation set and test set and input them into the convolutional neural network for training.
[0015] S7. Test the neural network trained in step S6 using the test set to obtain the diagnostic classification results of the intelligent diagnostic model.
[0016] Furthermore, the step S1 of acquiring the packaging machine bearing operating data to obtain a dataset of vibration signals showing the normal operating state and different fault types of the packaging machine bearing includes:
[0017] Vibration signals from the bearings of the packaging machine are acquired based on the normal operating status and different fault types of the bearings.
[0018] The vibration signal is stored in a PC to obtain the original vibration signal dataset.
[0019] Furthermore, the formula for calculating the average Pearson correlation coefficient used in step S2 to determine the optimal K value of the variational mode decomposition method is as follows:
[0020] ,
[0021] in, Here, Pearson correlation coefficient is given, and K is the optimal decomposition value of the variational mode. The Pearson correlation coefficient is calculated for each IMF component after signal decomposition, as follows:
[0022] ,
[0023] Furthermore, the specific method of variational mode decomposition for decomposing the effective signal in step S3 is as follows:
[0024] Constructing and solving a constrained variational problem is as follows:
[0025] The one-sided spectrum of the signal is obtained through Hilbert transform:
[0026] ,
[0027] Pre-estimate the center frequency of each analytical signal. Modal spectrum is modulated to the fundamental frequency band via frequency shift:
[0028] ,
[0029] The demodulated signal is processed using Gaussian smoothing, and the bandwidth of each BLIMF is estimated. The constrained variational problem constructed after processing can be expressed as:
[0030] ;
[0031] The solution process for the variational problem is as follows:
[0032] Introducing a quadratic penalty factor and a Lagrange operator, it is transformed into an unconstrained variational problem, yielding the augmented Lagrange function as:
[0033] ;
[0034] in, Represents a punishment factor; Represents the Lagrange multipliers; The middle part represents the inner product operation;
[0035] The saddle points of the augmented Lagrange expression can be iteratively updated using the multiplier alternation direction algorithm. , , Finding the saddle point of the unconstrained variational problem yields the final solution of the augmented Lagrangian function; in this algorithm, the generated decomposition components and their corresponding center frequencies... The update process in ADMM is as follows:
[0036] ;
[0037] ;
[0038] Lagrange multiplier The update is performed using the following formula, where n is the iteration parameter. It updates the parameters;
[0039] ;
[0040] The update process will repeat until the following conditions are met:
[0041] ,
[0042] For higher accuracy, when the above conditions are met, the signal can be adaptively decomposed into multiple different modal component signals.
[0043] Furthermore, the specific method of the overlap sampling in step S4 is as follows:
[0044] ,
[0045] in, Indicates the number of samples. Indicates the length of the sampled signal. Indicates the length of a single sample signal. This indicates the step size for each movement.
[0046] Furthermore, the specific method of the continuous wavelet transform in step S5 is as follows:
[0047] ,
[0048] in, These are the wavelet transform coefficients. For dependency a, wavelet basis functions, As a scale factor, The translation parameter is used. A Morlet wavelet, similar in waveform to the bearing fault vibration signal, is selected as the mother wavelet for wavelet transform of the vibration signal.
[0049] Furthermore, step S6, which involves adjusting the image size, dividing the image samples into training, validation, and test sets, and inputting them into the convolutional neural network for training, includes:
[0050] Different types of fault images, after continuous wavelet transform processing, undergo random angle rotation, flipping, scaling, and translation for data augmentation before being input into the neural network. The processed time-domain signals are then converted into a 64×64×3 RGB time-frequency image dataset. For each type of fault signal, after transformation, samples of the same fault type are labeled, and the image training set is shuffled. Supervised training is then used to input the images into the neural network for recognition and classification. Attached Figure Description
[0051] Figure 1 The following are the overall steps of the packaging machine bearing fault classification method based on convolutional neural networks of the present invention;
[0052] Figure 2This is a graph showing the variation of the average Pearson coefficient in an example of the packaging machine bearing fault classification method based on convolutional neural networks of the present invention.
[0053] Figure 3 The variational mode decomposition signal component diagram of a packaging machine bearing fault signal is shown as an example of the packaging machine bearing fault classification method based on convolutional neural network of the present invention.
[0054] Figure 4 This is a comparison image of the packaging machine bearing fault signal before and after reconstruction, which is an example of the packaging machine bearing fault classification method based on convolutional neural network of the present invention.
[0055] Figure 5 This is a schematic diagram of overlapping sampling of packaging machine bearing fault signals, which is an example of the packaging machine bearing fault classification method based on convolutional neural networks of the present invention.
[0056] Figure 6 A two-dimensional time-frequency domain image of a packaging machine bearing fault signal, which is an example of the packaging machine bearing fault classification method based on convolutional neural networks of the present invention.
[0057] Figure 7 The graph shows the accuracy of convolutional neural network fault identification, which is an example of the packaging machine bearing fault classification method based on convolutional neural network of the present invention.
[0058] Figure 8 The graph shows the convolutional neural network loss function curve for an example of the packaging machine bearing fault classification method based on convolutional neural networks of the present invention.
[0059] Figure 9 This is a confusion matrix diagram of the test set classification results for an example of the packaging machine bearing fault classification method based on convolutional neural networks of the present invention. Detailed Implementation
[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0061] like Figure 1 As shown, this application provides a method for classifying packaging machine bearing faults based on convolutional neural networks, the steps of which include S1 to S7:
[0062] S1. Obtain the operating data of the packaging machine bearing to obtain the vibration signal dataset of the normal operating state and different fault types of the packaging machine bearing, as a data sample for establishing the model;
[0063] Specifically, step S1 includes S11 to S12.
[0064] S11, Based on the vibration sensing equipment, obtain the vibration signals of the packaging machine bearing under normal operation and operation under different fault types;
[0065] S12, the vibration signal is stored in the PC to obtain the original vibration signal dataset;
[0066] Understandably, this example uses the radial position of the first cigarette pusher in a packaging machine as an example, sampling the vibration signal at a frequency of 25kHz. The original vibration signal dataset includes four different operating state types: normal operating state vibration signal, outer ring fault, inner ring fault, and rolling element fault. Except for the healthy state, each fault operating state includes three levels, with the fault severity increasing from low to high. Therefore, the fault diagnosis model includes a total of 10 different operating states. This invention uses a supervised training method when training the convolutional neural network, labeling the 10 signals as shown in the table below:
[0067]
[0068] Based on the bearing failure vibration state under different fault types, the principle to be observed when sampling the signal is to retain vibration information for at least one rotation cycle. The sampling length can be calculated using the following formula:
[0069] ,
[0070] In the formula This represents the number of sampling points contained in each signal segment; The rotational speed of the packaging machine spindle is 1072 rpm. The sampling frequency is 25kHz. After calculation, the number of sampling points L is determined to be 1500.
[0071] S2. The collected 10 kinds of original vibration signal data are preprocessed using an improved variational mode decomposition method to eliminate other source signals and noise interference, and obtain a pure target bearing vibration signal.
[0072] Specifically, step S2 includes S21 to S22.
[0073] S21, First, the optimal K value for the variational mode decomposition method is determined using the principle of minimizing the average Pearson correlation coefficient. The formula for calculating the average Pearson correlation coefficient is as follows:
[0074] ,
[0075] in, Here, Pearson correlation coefficient is given, and K is the optimal decomposition value of the variational mode. These are the individual IMF components after signal decomposition. The Pearson correlation coefficient is calculated as follows:
[0076] ,
[0077] Specifically, the average Pearson coefficient value corresponding to the K value within the iteration range is calculated. Based on the principle of minimizing the average Pearson coefficient, the optimal mode decomposition quantity K value corresponding to different fault states is determined. This maximizes the difference between adjacent frequency bands after variational mode decomposition of signals in different fault types, resulting in the best decomposition effect for bearing fault signals. Taking fault tag 3 data as an example, the iteration range of K value is set to [5-10]. The line graph showing the change of the average Pearson coefficient value with the K value within the iteration interval is shown below. Figure 2 As shown, the minimum average Pearson coefficient is 0.066559, corresponding to a K value of 8, meaning that the original signal achieves the best results when decomposed into 8 IMF components.
[0078] S22, Secondly, after determining the optimal decomposition K value for each set of data, the variational mode decomposition method is used to decompose the data. The specific method of variational mode decomposition is as follows:
[0079] Constructing and solving a constrained variational problem is as follows:
[0080] The one-sided spectrum of the signal is obtained through Hilbert transform.
[0081] ,
[0082] Pre-estimate the center frequency of each analytical signal. The modal spectrum is modulated to the baseband via frequency shifting;
[0083] ,
[0084] The demodulated signal is processed using Gaussian smoothing, and the bandwidth of each BLIMF is estimated. The constrained variational problem constructed after processing can be expressed as:
[0085] ;
[0086] The solution process for the variational problem is as follows:
[0087] Introducing a quadratic penalty factor and a Lagrange operator, it is transformed into an unconstrained variational problem, yielding the augmented Lagrange function as follows:
[0088] ;
[0089] in, Represents a punishment factor; Represents the Lagrange multipliers; The middle part is the inner product operation.
[0090] The saddle points of the augmented Lagrange expression can be iteratively updated using the multiplier alternation direction algorithm. , , Finding the saddle point of the unconstrained variational problem yields the final solution of the augmented Lagrangian function. In this algorithm, the generated decomposition components and their corresponding center frequencies... The update process in ADMM is as follows:
[0091] ;
[0092] ;
[0093] Lagrange multiplier The update is performed using the following formula, where n is the iteration parameter. It updates the parameters.
[0094] ;
[0095] The update process will repeat until the following conditions are met:
[0096] ,
[0097] For higher accuracy, when the above conditions are met, the signal can be adaptively decomposed into multiple different modal component signals.
[0098] This example uses fault tag 3 data. According to step S21, the optimal decomposition K value for fault 3 data is 8. Therefore, the fault 3 data signal is divided into 8 IMF components using variational mode decomposition. The time-domain waveform of the decomposed signal is shown below. Figure 3 As shown.
[0099] S23, after decomposing the original signal into K component signals, correlation analysis is performed between each component signal and the original signal to reconstruct the signal. The specific method is as follows:
[0100] The eight decomposed component signals were sequentially correlated with the original signal. Specifically, the correlation coefficient between each component signal and the original signal was calculated sequentially using the Pearson correlation coefficient formula. The correlation coefficient ranged from -1 to +1. When the correlation coefficient was positive, the two signals were positively correlated; when the correlation coefficient was negative, the two signals were negatively correlated; and when the correlation coefficient was 0, the two signals were not correlated. The closer the absolute value of the correlation coefficient was to 1, the higher the correlation.
[0101] After substituting the values into the Pearson correlation coefficient formula, the correlation coefficients between each component of the fault 3 data and the original signal are shown in the table below:
[0102]
[0103] IMF1, 4, 5, and 6, which have correlation coefficients greater than 0.3, were selected to reconstruct the signals. The comparison charts before and after reconstruction are shown below. Figure 4 As shown.
[0104] S3. The overlapping sampling method is used to increase the number of training samples, thereby expanding the training samples and fully extracting the fault information contained in each fault type.
[0105] In bearing fault identification of packaging machines, there is often a problem of insufficient actual fault data, resulting in low diagnostic generalization performance. Therefore, this example uses overlapping sampling to increase the number of training samples. The specific method is as follows:
[0106] Based on the packaging machine spindle speed of 1072 rpm, the sampling frequency is determined to be 25 kHz. The length of a single sample signal is determined using the sampling length calculation formula, which is as follows:
[0107] ,
[0108] After calculation, the value of the number of sampling points L is determined to be 1500. In this example, with a sampling frequency of 25KHz and a sampling time of 10s, the total length of the sampled signal is 250,000 data points. By overlapping sampling of training samples, if the length of a single sample is 1500 data points and the moving step size is set to 200, the number of training samples that can be obtained is 1244 according to the following formula.
[0109] ,
[0110] in, Indicates the number of samples. Indicates the length of the sampled signal. Indicates the length of a single sample signal. The diagram illustrating the signal overlap sampling at each movement step is shown below. Figure 5As shown, overlapping sampling can be used to expand the training samples and fully extract the fault information contained in each fault type. Furthermore, the training samples processed using this method can improve the generalization performance of the training model, thereby increasing the model's fault identification accuracy.
[0111] S4. Use the continuous wavelet transform method to convert the expanded one-dimensional time-domain fault signal into a two-dimensional time-frequency domain fault image.
[0112] This example uses fault data 3 as an example. The specific method of continuous wavelet transform is as follows:
[0113] ,
[0114] in, These are the wavelet transform coefficients. For dependency a, wavelet basis functions, As a scale factor, The translation parameter is used. A Morlet wavelet, similar in waveform to the bearing fault vibration signal, is selected as the mother wavelet for wavelet transform of the vibration signal.
[0115] The two-dimensional time-frequency domain image of the fault signal after continuous wavelet transform is as follows: Figure 6 As shown.
[0116] S5. Adjust the image size, divide the image samples into training set, validation set and test set and input them into the convolutional neural network for training;
[0117] Specifically, step S5 includes S51 to S52;
[0118] S51, In this example, for each fault type signal, data augmentation methods yield 1244 training samples. 800 samples are randomly selected from these as the sample set for fault diagnosis. The ten fault types contain a total of 8000 samples, which are randomly divided into a training set of 7000 samples, a test set of 250 samples, and a validation set of 750 samples. The preprocessed time-domain data is analyzed using continuous wavelet transform to obtain ten different types of fault diagnosis images. Before inputting into the neural network, random angle rotation, flipping, scaling, and translation are applied to the fault images for data augmentation. The processed time-domain signals are converted into a 64×64×3 RGB time-frequency image dataset. For each different type of fault signal, after conversion, samples of the same fault type are labeled, and the image training set is shuffled. Supervised training is then used to input the images into the neural network for recognition and classification.
[0119] S52, This example uses a convolutional neural network to identify and classify two-dimensional fault images. The detailed parameters of each layer of the convolutional neural network and the size of the output features are shown in the table below.
[0120]
[0121] In the convolutional neural network, the stride and padding parameters of the convolutional kernel traversal are both set to 1, and the kernel size is 3. The kernel size and stride of the max-pooling layer are set to 2, and the padding is set to 1. The output classification number of the last layer is 10. The model uses a supervised training method. During the preprocessing of the dataset, signals of different fault types are labeled. The learning rate is 0.0001. This network structure uses 5 convolutional layers and 4 max-pooling layers, with 3 fully connected layers after the max-pooling layers. To avoid gradient vanishing problems during backpropagation, the ReLU nonlinear activation function is used between the convolutional layers, which also increases the nonlinear relationship between the layers of the neural network. In the fully connected layers, the data gain Dropout parameter is set to 0.5, and half of the neurons are randomly deactivated. This method can avoid problems such as getting stuck in local optima during the diagnostic model iteration process, and can also improve the accuracy and generalization performance of the diagnostic model. Setting these hyperparameters of the network can enable the network to achieve better convergence, thereby establishing a complex nonlinear mapping between bearing fault identification and known fault types, realizing intelligent diagnosis of bearing faults.
[0122] S6. Input the test set into the trained convolutional neural network for testing to obtain the diagnostic classification results of the intelligent diagnostic model;
[0123] The network is trained using a mini-batch method, with each batch processing 32 samples (Bathch_size = 32). One training epoch requires 219 batches. Using mini-batch improves training speed and saves time. One iteration ends after all batches have been trained. The neural network is trained for 50 epochs. The accuracy and loss function curves after training are shown below. Figure 7 , Figure 8 As shown.
[0124] The training results from 50 rounds show that the diagnostic model proposed in this example achieves a stable recognition accuracy on the validation set after 15 rounds of training. At this point, the model's loss function curve has not yet converged to its minimum, and the backpropagation of the neural network continues, causing the model's loss value to continue converging to its minimum. After 34 rounds of training, the model's recognition accuracy stabilizes, reaching a maximum of 97.3% on the validation set, with an average training result of 97.10% after 30 rounds. After 50 rounds of training, the average recognition accuracy of the diagnostic model reaches 97.12%, and the loss function stabilizes at around 0.009. To avoid overfitting of the diagnostic model, the results after 50 rounds of training are used as the diagnostic model for the test set.
[0125] The parameters of the diagnostic model trained for 50 rounds are saved, and 250 samples from the test set are input into the neural network to identify and classify different fault types. The confusion matrix of the test set sample identification and prediction results is plotted as follows. Figure 9 As shown.
[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for classifying bearing faults in packaging machines based on convolutional neural networks, characterized in that, Includes the following steps: S1. Obtain the operating data of the packaging machine bearing to obtain the vibration signal dataset of the normal operating state and different fault types of the packaging machine bearing, as a data sample for establishing the model; S2. For the data samples established in step S1, determine the optimal decomposition K value of the improved variational mode decomposition method in the data sample preprocessing using the principle of minimizing the average Pearson correlation coefficient; S3. Based on the optimal decomposition K value obtained in step S2, extract the time-domain features of the fault signal, preprocess the acquired original signal dataset using variational mode decomposition, decompose the original vibration signal into K component signals, perform correlation analysis between the K component signals and the original signal, select the highly correlated components for reconstruction, remove the interference of redundant components, and obtain an effective one-dimensional time-domain fault signal; S4. For the effective one-dimensional time-domain fault signals obtained in step S3, the overlapping sampling method is used to increase the number of training samples, thereby expanding the training samples and fully extracting the fault information contained in each fault type. S5. Extract the time-frequency domain features of the fault signal and use the continuous wavelet transform method to convert the expanded one-dimensional time-domain fault signal into a two-dimensional time-frequency domain fault image. S6. For the two-dimensional time-frequency domain fault image obtained in step S5, adjust the image size, divide the image samples into training set, validation set and test set, and input them into the convolutional neural network for training; S7. Test the neural network trained in step S6 using the test set to obtain the diagnostic classification results of the intelligent diagnostic model.
2. The packaging machine bearing fault classification method based on convolutional neural networks according to claim 1, characterized in that, The step S1, which involves acquiring the operating data of the packaging machine bearing to obtain a dataset of vibration signals indicating the normal operating state and different fault types of the packaging machine bearing, includes: Vibration signals from the bearings of the packaging machine are acquired based on the normal operating status and different fault types of the bearings. The vibration signal is stored in a PC to obtain the original vibration signal dataset.
3. The packaging machine bearing fault classification method based on convolutional neural networks according to claim 1, characterized in that, The formula for calculating the average Pearson correlation coefficient used in step S2 to determine the optimal K value of the variational mode decomposition method is as follows: ， in, Here, Pearson correlation coefficient is given, and K is the optimal decomposition value of the variational mode. The Pearson correlation coefficient is calculated for each IMF component after signal decomposition, as follows: 。 4. The packaging machine bearing fault classification method based on convolutional neural networks according to claim 1, characterized in that, The specific method of variational mode decomposition for decomposing the effective signal in step S3 is as follows: Constructing and solving a constrained variational problem is as follows: The one-sided spectrum of the signal is obtained through Hilbert transform. ， Pre-estimate the center frequency of each analytical signal. The modal spectrum is modulated to the fundamental frequency band via frequency shifting. ， The demodulated signal is processed using Gaussian smoothing, and the bandwidth of each BLIMF is estimated. The constrained variational problem constructed after processing can be expressed as: ， The solution process for the variational problem is as follows: Introducing a quadratic penalty factor and a Lagrange operator, it is transformed into an unconstrained variational problem, yielding the augmented Lagrange function as follows: ， in, Represents a punishment factor; Represents the Lagrange multipliers; The middle part is the inner product operation; The saddle points of the augmented Lagrange expression are iteratively updated using the multiplier alternation direction algorithm. , , Finding the saddle point of the unconstrained variational problem yields the final solution of the augmented Lagrangian function. In this algorithm, the generated decomposition components and their corresponding center frequencies are... The update process in ADMM is as follows: ， ， Lagrange multiplier The update is performed using the following formula, where n is the iteration parameter. It updates the parameters. ， The update process will repeat until the following conditions are met: ， For accuracy, when the above conditions are met, the signal is adaptively decomposed into multiple different modal component signals.
5. The packaging machine bearing fault classification method based on convolutional neural networks according to claim 1, characterized in that, The specific method of the overlap sampling method in step S4 is as follows: ， in, Indicates the number of samples. Indicates the length of the sampled signal. Indicates the length of a single sample signal. This indicates the step size for each movement.
6. The packaging machine bearing fault classification method based on convolutional neural networks according to claim 1, characterized in that, The specific method of the continuous wavelet transform in step S5 is as follows: ， in, These are the wavelet transform coefficients. For dependency a, wavelet basis functions, As a scale factor, To use the translation parameters, Morlet, which is similar to the waveform of the bearing fault vibration signal, is selected as the mother wavelet to perform wavelet transform on the vibration signal.
7. The packaging machine bearing fault classification method based on convolutional neural networks according to claim 1, characterized in that, Step S6, which involves adjusting the image size, dividing the image samples into training, validation, and test sets, and inputting them into the convolutional neural network for training, includes: Different types of fault images processed by continuous wavelet transform are subjected to random angle rotation, flipping, scaling transformation and translation before being input into the neural network for data augmentation. The processed time-domain signal is converted into a 64×64×3 RGB time-frequency image dataset. For each type of fault signal, after conversion, each sample of the same fault type is labeled and the image training set is shuffled. The images are then input into the neural network for recognition and classification using a supervised training method.