Rolling bearing fault diagnosis method based on small sample data

Abstract

The invention discloses a rolling bearing fault diagnosis method based on small sample data, and the method is based on a SqueezeNet convolutional neural network, carries out the fine tuning based on the utilization of a pre-training model through a transfer learning technology, and remarkably improves the learning capability and diagnosis accuracy of a small sample data set. Especially for rolling bearing faults, the method can accurately identify different fault types and severity thereof. The fault prediction model adopted by the method is small in parameter quantity, low in calculation complexity, suitable for embedded equipment or resource-limited scenes, capable of meeting the requirement for real-time fault diagnosis, capable of being applied to fault diagnosis of the rolling bearing and capable of being popularized to state monitoring and fault diagnosis tasks of other mechanical equipment, and high in practicability. And the method has relatively high universality and flexibility.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault diagnosis, and particularly relates to a rolling bearing fault diagnosis method based on small sample data, realizing intelligent identification and diagnosis of rolling bearing fault types. Background Art

[0002] Rolling bearings are key components widely used in mechanical equipment, and their performance directly affects the operation efficiency and safety of the equipment. However, due to the complex working environment, rolling bearings are easily affected by factors such as wear, fatigue, and external vibration, resulting in failures. Traditional fault diagnosis methods mainly rely on empirical feature extraction and manual analysis, suffering from problems such as low diagnosis efficiency, poor accuracy, and limited applicability.

[0003] In recent years, with the rapid development of deep learning technology, convolutional neural networks have shown significant advantages in the field of fault diagnosis, capable of automatically extracting high-order features from complex data. However, the large-scale deep network models used in existing methods have a high computational complexity and are not suitable for embedded devices or real-time application scenarios. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the present invention proposes a rolling bearing fault diagnosis method based on small sample data. By combining a lightweight network model and transfer learning technology, the diagnosis efficiency and accuracy are improved, and the deficiencies of the traditional method, such as high demand for computing resources and dependence on manual feature extraction, are solved.

[0005] The technical solution of the present invention to solve the above technical problems is: designing a rolling bearing fault diagnosis method based on small sample data, characterized in that the method specifically includes the following steps:

[0006] Step 1: Obtain the characteristic data of the rolling bearing under different fault states and different working conditions, and establish a rolling bearing fault diagnosis data set; in this data set, one piece of data includes the acceleration vibration signal x(t) of a certain duration after preprocessing, and the corresponding fault state is used as a label;

[0007] Step 2: Through the continuous wavelet transform method, convert the acceleration vibration signal x(t) in the rolling bearing fault diagnosis data set into a two-dimensional RGB image through time-frequency domain transformation to obtain a training data set;

[0008] Step 3: Establish a fault diagnosis network model

[0009] The fault diagnosis network model is an improvement on the pre-trained SqueezeNet convolutional neural network. The SqueezeNet convolutional neural network specifically includes a first convolutional layer, a first max pooling layer, a first Fire module, a second Fire module, a third Fire module, a second max pooling layer, a fourth Fire module, a fifth Fire module, a sixth Fire module, a seventh Fire module, a third max pooling layer, an eighth Fire module, a second convolutional layer, a global average pooling layer, and a classification regression layer. These fifteen parts are connected in sequence. The input data is first input into the first convolutional layer and processed through the fifteen parts in sequence. The output of the previous part is the input of the next part. Finally, the classification regression layer outputs the probability of the input data belonging to each category prediction. The category with the largest probability value is selected as the prediction category of the SqueezeNet convolutional neural network for the input data.

[0010] The fault diagnosis network model refers to replacing the second convolutional layer in the pre-trained SqueezeNet convolutional neural network with a fully connected layer with a 1×1 convolutional kernel. The output channels of this fully connected layer are set to the number of categories of the labels, and the trainable parameters of this fully connected layer are initialized to random values, thus obtaining the fault diagnosis network model.

[0011] Step 4: Train the fault diagnosis network model

[0012] Step 4.1: Divide the training data set in Step 2 into a training set and a validation set randomly and without repetition according to a certain number ratio based on the number of data items. Both the training set and the validation set contain training data under different fault states and different working conditions.

[0013] Step 4.2: Set the initial learning rate to 0.0001, and set the maximum number of training epochs and the maximum number of iterations per epoch.

[0014] Step 4.3: Initialize the trainable parameters of the fully connected layer of the fault diagnosis network model to random values, and fix the remaining trainable parameters to the pre-trained values of the pre-trained SqueezeNet convolutional neural network; Use the two-dimensional RGB image of a piece of training data in the training set in Step 4.1 as the input of the fault diagnosis network model, and the corresponding label as the reference value of the model output. Use cross-entropy loss to calculate the training loss; According to the training loss of a piece of training data, use SGDM as the optimization algorithm to update the trainable parameters of the fully connected layer backward once; Use the updated network parameters as the initial parameters when training the next piece of training data, input the next piece of training data in the training set, and repeat continuously until the last piece of data in the training set is trained, completing one round of training of the training set; After each round of training is completed, input the training data in the validation set in Step 4.1 into the network that has completed one round of training, and calculate the accuracy of the current fault diagnosis network model's prediction of the validation set:

[0015]

[0016] wherein, is an indicator function; when the prediction result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i is the i-th piece of training data in the validation set;

[0017] Use the network parameters that have completed one round of training as the initial parameters for the next round of training, and repeat continuously. When the change in the accuracy of the fault diagnosis network model's prediction of the validation set after 8 consecutive rounds of training is no more than 0.01%, or the number of training rounds reaches the preset value, complete the training of the first cycle to obtain the initial training value of the trainable parameters of the fully connected layer;

[0018] Step 4.4: Use the initial training values of the trainable parameters of the fully connected layer of the fault diagnosis network model and the pre-training values of the remaining trainable parameters as the initial values of the network model; repeat the process of dividing the training set and the validation set in Step 4.1 to obtain a new training set and a new validation set; use the two-dimensional RGB image of a piece of training data in the new training set as the input of the fault diagnosis network model, use its label as the reference value of the model output, and use cross-entropy loss to calculate the training loss; according to the training loss of a piece of training data, use SGDM as the optimization algorithm to update all the trainable parameters of the fault diagnosis network model backward once; use the updated network parameters as the initial parameters when training the next piece of training data, input the next piece of training data in the new training set, and repeat continuously until the last piece of training data in the new training set is trained, completing one round of training of the new training set; after each round of training is completed, input the training data in the new validation set into the network model that has completed one round of training, and calculate the accuracy of the current fault diagnosis network model's prediction of the new validation set:

[0019]

[0020] In the formula, is an indicator function; when the prediction result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i is the i-th piece of training data in the validation set;

[0021] Use the network parameters that have completed one round of training as the initial parameters for the next round of training, and repeat continuously. When the change in the accuracy of the fault diagnosis network model's prediction of the new validation set after 8 consecutive rounds of training is no more than 0.01%, or the number of training rounds reaches the preset value, complete the training of the second cycle;

[0022] Step 4.5: Use the network parameters that have completed the training of the second cycle as the initial parameters for the next cycle of training, and repeat the process of Step 4.4 until the prediction accuracy of the network model that has completed one cycle of training for the validation set is not less than 99% or reaches the maximum number of cycles, to obtain the fault diagnosis network model that has completed training;

[0023] Step Five, Fault Diagnosis of Rolling Bearings

[0024] Obtain the acceleration vibration signal of a certain duration of the rolling bearing to be diagnosed, and convert it into a two-dimensional RGB image according to the method in Step Two; then input the two-dimensional RGB image into the fault diagnosis network model that has completed training in Step Three, and according to the predicted probabilities of different types output by the network model, select the category corresponding to the maximum probability as the fault type of the rolling bearing to obtain the fault diagnosis result of the rolling bearing.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] (1) Lightweight and high efficiency: This method is based on the SqueezeNet convolutional neural network, which has a small number of model parameters and low computational complexity, is suitable for embedded devices or scenarios with limited resources, and can meet the requirements of real-time fault diagnosis at the same time.

[0027] (2) Accuracy and robustness: Through transfer learning technology, this method fine-tunes on the basis of using a pre-trained model, significantly improving the learning ability and diagnostic accuracy for small sample data sets; especially for rolling bearing faults, this method can accurately identify different fault types and their severity.

[0028] (3) Strong applicability: This method can not only be applied to the fault diagnosis of rolling bearings, but also be extended to the condition monitoring and fault diagnosis tasks of other mechanical equipment, with strong versatility and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of the steps of an embodiment of a rolling bearing fault diagnosis method based on small sample data according to the present invention.

[0030] Figure 2 is a schematic diagram of the structure and principle of the SqueezeNet convolutional neural network.

[0031] Figure 3 is the training situation and verification results of the fault diagnosis network model using the training data of the training set and the training data of the validation set in Embodiment 1 of a rolling bearing fault diagnosis method based on small sample data according to the present invention; in the figure, the light gray curve represents the change result of the accuracy during the training process, the dark gray curve represents the change result of the accuracy after smoothing during the training process, and the dotted line represents the change result of the accuracy during the verification process.

[0032] Figure 4 is the confusion matrix of the prediction results of the fault diagnosis network model trained with the training data of the training set and the training data of the validation set in Embodiment 1 using a rolling bearing fault diagnosis method based on small sample data according to the present invention for the training data of the test set in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The following further illustrates the present invention with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited do not limit the present invention.

[0034] The present invention provides a rolling bearing fault diagnosis method based on small sample data, and the flowchart of this method is as Figure 1 shown, and specifically includes the following steps:

[0035] Step 1: Obtain the characteristic data of the rolling bearing under different fault states and different working conditions, and establish a rolling bearing fault diagnosis data set; in this data set, one piece of data includes the acceleration vibration signal x(t) of a certain duration after preprocessing, and the corresponding fault state is used as a label.

[0036] As an embodiment, the above preprocessing is denoising.

[0037] As an embodiment, the fault states specifically include three types: normal state, inner ring fault, and outer ring fault.

[0038] As an embodiment, the working conditions specifically include the sampling frequency f s , rotational frequency f r , and load load.

[0039] Step 2: Through the continuous wavelet transform method, convert the acceleration vibration signal x(t) in the rolling bearing fault diagnosis data set into a two-dimensional RGB image through time-frequency domain transformation to obtain a training data set;

[0040] The specific steps of Step 2 are as follows: First, divide the acceleration vibration signal x(t) of each piece of data into multiple segments, and the segmentation principle is defined as follows:

[0041]

[0042] In the formula, L interval is the data length of each segment of the signal, L interval = 0.0512f s , f s is the sampling frequency of this piece of data; here, N is the number of segments into which the acceleration vibration signal x(t) of this piece of data can be divided, || represents rounding, and t represents the sampling time point of the acceleration vibration signal. A segment of the acceleration vibration signal x(t) of one piece of data can be expressed as:

[0043] x k (t) = x[L interval (k - 1) + 1:L interval k], k = 1, 2,..., N

[0044] In the formula, x k (t) is the vibration signal of the kth segment of the acceleration vibration signal x(t).

[0045] Then, perform continuous wavelet transform on each segment of the vibration signal respectively to obtain wavelet coefficients:

[0046]

[0047] In the formula, a is the scale factor, representing the compression or expansion of frequency; b is the time translation factor, representing the movement of the wavelet function in time. * represents the operation of taking the complex conjugate, and ψ a,b (t) is a wavelet function, which is the translation and scale change form of the mother wavelet:

[0048]

[0049] Here, CWT(a, b) is the complex coefficient in the time-scale domain, that is, the wavelet coefficient. The above mother wavelet is one of the mother wavelets such as the Mexican hat wavelet, Morlet wavelet, Ricker wavelet, etc.

[0050] Take the absolute value of the wavelet coefficient as the energy distribution of this section of the vibration signal:

[0051] W k (a, b) = |CWT k (a, b)|

[0052] Then, flip and normalize the wavelet coefficients of each section of the vibration signal, and map them to an RGB image:

[0053]

[0054] In the formula, W k,min and W k,max respectively refer to the minimum and maximum values in the wavelet coefficient matrix of this section of the vibration signal.

[0055] Generate a two-dimensional RGB image from the N-section vibration signals of one piece of data, and then scale the generated two-dimensional RGB image to 224×224×3 to obtain one piece of training data, completing the process of converting the acceleration vibration signal x(t) of one piece of data into a two-dimensional RGB image.

[0056] The acceleration vibration signal x(t) of each piece of data in the rolling bearing fault diagnosis dataset is generated into a two-dimensional RGB image through continuous wavelet transform to obtain a training dataset. The size of this two-dimensional RGB image is 224×224×3, which is the same as the size of the input image required by the SqueezeNet convolutional neural network model.

[0057] Step 3. Establish a fault diagnosis network model

[0058] The fault diagnosis network model is an improvement on the pre-trained SqueezeNet convolutional neural network. The specific architecture of the SqueezeNet convolutional neural network is as Figure 2As shown in the figure, it is the prior art, specifically including the first convolutional layer Conv1, the first max pooling layer (Maxpool), the first Fire module Fire2, the second Fire module Fire3, the third Fire module Fire4, the second max pooling layer, the fourth Fire module Fire5, the fifth Fire module Fire6, the sixth Fire module Fire7, the seventh Fire module Fire8, the third max pooling layer, the eighth Fire module Fire9, the second convolutional layer Conv10, the global average pooling layer (Global avgpool), and the classification regression layer (Softmax). The above fifteen parts are connected in sequence. The input data is first input into the first convolutional layer Conv1 and processed through the fifteen parts in sequence. The output of the previous part is the input of the next part. Finally, the classification regression layer outputs the probability that the input data belongs to each category prediction. The category with the largest probability value is selected as the predicted category of the SqueezeNet convolutional neural network for the input data.

[0059] The pre-trained SqueezeNet convolutional neural network refers to a publicly available pre-trained SqueezeNet convolutional neural network whose trainable parameter values are trained on more than one million publicly available images.

[0060] Specifically, the fault diagnosis network model refers to replacing the second convolutional layer Conv10 in the pre-trained SqueezeNet convolutional neural network with a fully connected layer with a 1×1 convolutional kernel. The output channels of this fully connected layer are set to the number of label categories, which is 3 in this embodiment. The trainable parameters of this fully connected layer are initialized to random values, thus obtaining the fault diagnosis network model.

[0061] Step 4: Train the fault diagnosis network model

[0062] Use the pre-trained SqueezeNet convolutional neural network as the starting point of the rolling bearing fault diagnosis model. Use transfer learning and the data in the training set to fine-tune the network to adapt to the fault characteristics of the rolling bearing.

[0063] Step 4.1: Randomly and non-repeatedly divide the training data set in step 2 into a training set and a validation set according to a certain proportion of the number of data, generally set to 6:4. Both the training set and the validation set contain training data under different fault states and different working conditions.

[0064] Step 4.2: Set the initial learning rate to 0.0001, the maximum number of training epochs to 4, and the maximum number of iterations per epoch to 40 times.

[0065] Step 4.3: Initialize the trainable parameters of the fully connected layer of the fault diagnosis network model to random values, and fix the remaining trainable parameters to the pre-trained values of the pre-trained SqueezeNet convolutional neural network; use the two-dimensional RGB image of a piece of training data in the training set in Step 4.1 as the input of the fault diagnosis network model, and the corresponding label ("inner race fault", "outer race fault", and "normal state") as the reference value of the model output. Use cross-entropy loss to calculate the training loss; according to the training loss of a piece of training data, use SGDM (stochastic gradient descent with momentum) as the optimization algorithm to update the trainable parameters of the fully connected layer backward once; use the updated network parameters as the initial parameters when training the next piece of training data, input the next piece of training data in the training set, and repeat continuously until the last piece of data in the training set is trained, completing one round of training of the training set; after each round of training is completed, input the training data in the validation set in Step 4.1 into the network that has completed one round of training, and calculate the accuracy of the current fault diagnosis network model's prediction of the validation set:

[0066]

[0067] wherein, is the indicator function. When the prediction result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i is the i-th piece of training data in the validation set.

[0068] Use the network parameters that have completed one round of training as the initial parameters for the next round of training, and repeat continuously. When the change in the accuracy of the fault diagnosis network model's prediction of the validation set when 8 consecutive rounds of training are completed is no more than 0.01%, or the number of training rounds reaches the preset value, complete the training of the first cycle to obtain the initial training value of the trainable parameters of the fully connected layer.

[0069] Step 4.4: Use the initial training values of the trainable parameters of the fully connected layer of the fault diagnosis network model and the pre-training values of the remaining trainable parameters as the initial values of the network model; repeat the process of dividing the training set and the validation set in Step 4.1 to obtain a new training set and a new validation set; use the two-dimensional RGB image of a piece of training data in the new training set as the input of the fault diagnosis network model, and use its label ("inner race fault", "outer race fault", and "normal state") as the reference value of the model output. Use cross-entropy loss to calculate the training loss; according to the training loss of a piece of training data, use SGDM (stochastic gradient descent with momentum) as the optimization algorithm to update all the trainable parameters of the fault diagnosis network model backward once; use the updated network parameters as the initial parameters when training the next piece of training data, input the next piece of training data in the new training set, and repeat continuously until the last piece of training data in the new training set is trained, completing one round of training of the new training set; after each round of training is completed, input the training data in the new validation set into the network model that has completed one round of training, and calculate the accuracy of the current fault diagnosis network model's prediction of the new validation set:

[0070]

[0071] In the formula, is an indicator function. When the prediction result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i is the i-th piece of training data in the validation set.

[0072] Use the network parameters that have completed one round of training as the initial parameters for the next round of training, and repeat continuously. When the change in the accuracy of the fault diagnosis network model's prediction of the new validation set after 8 consecutive rounds of training is no more than 0.01%, or the number of training rounds reaches the preset value, complete the training of the second cycle.

[0073] Step 4.5: Use the network parameters that have completed the training of the second cycle as the initial parameters for the next cycle of training, and repeat the process of Step 4.4 until the prediction accuracy of the network model that has completed one cycle of training for the validation set is not less than 99% or reaches the maximum number of cycles, obtaining the fault diagnosis network model that has completed training.

[0074] Step Five: Rolling bearing fault diagnosis

[0075] Obtain the acceleration vibration signal of a certain duration of the rolling bearing to be diagnosed, and convert it into a two-dimensional RGB image according to the method in Step Two; then input the two-dimensional RGB image into the fault diagnosis network model that has completed training in Step Three, and select the category corresponding to the maximum probability as the fault type of the rolling bearing according to the prediction probabilities of different types output by the network model, obtaining the fault diagnosis result of the rolling bearing.

[0076] Example 1

[0077] Taking the rolling bearing fault dataset of the Machinery Failure Prevention Technology Society (MFPT) as an example, the technical effect of the rolling bearing fault diagnosis method designed based on small-sample data of the present invention is verified.

[0078] The parameters of the rolling bearings tested in the MFPT dataset are shown in Table 1 below. The units of the ball diameter and pitch diameter are not published in this dataset, which has no impact on the fault diagnosis of rolling bearings.

[0079] Table 1 Rolling bearing parameters

[0080]

[0081] There are a total of 23 datasets under different fault conditions in this dataset. The first 20 datasets are collected from the bearing test bench, among which 3 are in good condition, 3 have outer ring faults under constant load, 7 have outer ring faults under different loads, and 7 have inner ring faults under different loads. The last 3 datasets come from actual rotating machinery, but the specific fault locations are unknown. In this example, only the first 20 datasets collected from the bearing test bench under known conditions are used. The detailed description of the bearing test bench dataset is shown in Table 2 below.

[0082] Table 2 Bearing test bench dataset

[0083]

[0084]

[0085] Each of the above datasets contains information such as the acceleration signal x(t), sampling frequency fs, rotation frequency fr, load, etc. At the same time, it includes four critical frequencies representing different fault locations: the outer ring characteristic frequency BPFO, the inner ring characteristic frequency BPFI, the cage characteristic frequency FTF, and the ball characteristic frequency BSF. According to the rolling bearing parameters included in the dataset, the outer ring characteristic frequency BPFO, the inner ring characteristic frequency BPFI, the cage characteristic frequency FTF, and the ball characteristic frequency BSF can be calculated according to the following formula:

[0086]

[0087] In the above formula, d is the ball diameter, D is the pitch diameter, n is the number of rolling elements, is the bearing contact angle. The calculation results are shown in Table 3 below.

[0088] Table 3 Four critical frequencies representing different fault locations

[0089]

[0090] The vibration signals with serial numbers 1, 4, 7, 8, 9, 10, 14, 15, 16, and 17 in the above dataset are used as the training set in the rolling bearing fault diagnosis process. The vibration signals with serial numbers 2, 5, 11, and 18 are used as the validation set in the rolling bearing diagnosis process. The vibration signals with serial numbers 3, 6, 12, 13, 19, and 20 are used as the test set in the rolling bearing fault process.

[0091] All one-dimensional vibration signals in the dataset are transformed into two-dimensional RGB image signals using continuous wavelet transform. The specific process is as follows:

[0092] First, the acceleration vibration signal x(t) of each piece of data is divided into multiple segments, and the segmentation principle is defined as follows:

[0093]

[0094] where L interval is the data length of each segment of the signal, L interval = 0.0512f s ; N is the number of segments into which the entire vibration signal x(t) can be divided, || represents rounding, and t represents the sampling time point of the acceleration vibration signal. Therefore, according to the data in Table 2, the number of segments into which each dataset is divided can be calculated respectively. Each segment of the acceleration vibration signal x(t) of a piece of data can be expressed as:

[0095] x k (t) = x[L interval (k - 1)+1:L interval k], k = 1, 2, …, N

[0096] where x k (t) is the vibration signal of the k-th segment of the acceleration vibration signal x(t).

[0097] Then, continuous wavelet transform is performed on each segment of the vibration signal respectively:

[0098]

[0099] where a is the scale factor, representing the compression or expansion of frequency; b is the time shift factor, representing the shift of the wavelet function in time. * represents the operation of taking the complex conjugate, and ψ a,b (t) is a wavelet function, which is the translation and scale change form of the mother wavelet:

[0100]

[0101] Here, CWT(a,b) is the complex coefficient in the time-scale domain, that is, the wavelet system. Take the absolute value of the wavelet coefficient as the energy distribution of the signal:

[0102] W k (a,b) = |CWT k (a,b)|

[0103] Then, flip and normalize the wavelet coefficients of each segment of the vibration signal, and map them to an RGB image:

[0104]

[0105] In the formula, W k,min and W k,max respectively refer to the minimum and maximum values in the wavelet coefficient matrix.

[0106] Generate a two-dimensional RGB image from the N segments of vibration signals of one piece of data, and scale the generated two-dimensional RGB image to 224×224×3, completing the process of converting the acceleration vibration signal x(t) of one piece of data into a two-dimensional RGB image. The two-dimensional RGB image has the same size as the input image required by the SqueezeNet convolutional neural network model, and can be used as the input image to train the SqueezeNet convolutional neural network model.

[0107] By performing continuous wavelet transform on each segment of vibration signal in each dataset, the number of two-dimensional RGB image data in the generated training set, validation set, and test set is shown in Table 4.

[0108] Table 4 Number of two-dimensional RGB image data

[0109]

[0110] Use the two-dimensional RGB image data in the training set to train the fault diagnosis network model. The fault diagnosis network model uses an improved pre-trained SqueezeNet convolutional neural network. This pre-trained SqueezeNet convolutional neural network has been trained on more than one million images and has learned rich feature representations. Use the pre-trained SqueezeNet convolutional neural network as the starting point for the rolling bearing fault diagnosis model. Use transfer learning and the data in the training set to fine-tune the network to adapt it to the fault characteristics of the rolling bearing.

[0111] Replace the second convolutional layer Conv10 of the pre-trained SqueezeNet convolutional neural network with a fully connected layer with a 1×1 convolutional kernel, set the output channels of this fully connected layer to 3, and initialize the trainable parameters of this fully connected layer to random values, then the fault diagnosis network model is obtained.

[0112] The fault diagnosis network model is trained in stages using the training set and the validation set. In the first stage, the pre-trained values of the trainable parameters of the pre-trained SqueezeNet convolutional neural network part are frozen, and the trainable parameters of the fully connected layer are trained. In the second stage, based on the first stage, all the trainable parameters of the fault diagnosis network model are trained.

[0113] During the training process, the cross-entropy loss function is adopted, and SGDM (Stochastic Gradient Descent with Momentum) is used as the optimization algorithm. By introducing momentum, the gradient descent is accelerated, and the oscillation during convergence is effectively avoided. The initial learning rate is set to 0.0001. A lower learning rate helps prevent excessive parameter updates, thereby avoiding unstable situations during training and ensuring that the network can converge smoothly. The maximum number of training epochs is set to 4, and the maximum number of training iterations per epoch is 40. After each training epoch, the training data is randomly shuffled to avoid overfitting to the data order and improve the generalization ability of the model. After each training iteration, the validation set is used to validate the model to evaluate whether overfitting occurs in the network.

[0114] Finally, the predicted results are compared with the true results, and their accuracy is calculated.

[0115]

[0116] In the formula, is the indicator function. When the predicted result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i is the i-th training data in the validation set.

[0117] The training situation and validation results of the fault diagnosis network model using the training set training data and the validation set training data are as Figure 3 shown. It can be seen from the results that after four epochs of training, the accuracy of the fault diagnosis network model for rolling bearing fault diagnosis can reach more than 99%.

[0118] The trained model is used to classify the training data in the test set and output the fault types. The classification results include three status classifications: normal status, inner ring fault, and outer ring fault. The confusion matrix is used to visualize the classification results to help analyze the performance of the model in each category. The results are as Figure 4 shown. The results show that the present invention has high accuracy in the rolling bearing fault diagnosis of small sample data.

[0119] Using the technical solution described in the present invention, or a person skilled in the art designing a similar technical solution inspired by the technical solution of the present invention and achieving the above technical effects, all fall within the protection scope of the present invention.

[0120] What is not described in the present invention is applicable to the prior art.

Claims

1. A rolling bearing fault diagnosis method based on small sample data, characterized in that, The method specifically includes the following steps: Step 1: Obtain the characteristic data of the rolling bearing under different fault states and different working conditions, and establish a rolling bearing fault diagnosis data set; in this data set, one piece of data includes the acceleration vibration signal x(t) of a certain duration after preprocessing, and the corresponding fault state is used as the label; Step 2: Through the continuous wavelet transform method, transform the acceleration vibration signal x(t) in the rolling bearing fault diagnosis data set into a two-dimensional RGB image through time-frequency domain transformation to obtain a training data set; Step 3: Establish a fault diagnosis network model The fault diagnosis network model is an improvement on the pre-trained SqueezeNet convolutional neural network; the SqueezeNet convolutional neural network specifically includes a first convolutional layer, a first max pooling layer, a first Fire module, a second Fire module, a third Fire module, a second max pooling layer, a fourth Fire module, a fifth Fire module, a sixth Fire module, a seventh Fire module, a third max pooling layer, an eighth Fire module, a second convolutional layer, a global average pooling layer, and a classification regression layer. These fifteen parts are connected in sequence. The input data is first input into the first convolutional layer and processed through fifteen parts in sequence. The output of the previous part is the input of the next part. Finally, the classification regression layer outputs the probability of the input data belonging to each category prediction. The category with the largest probability value is selected as the prediction category of the SqueezeNet convolutional neural network for the input data; The fault diagnosis network model refers to replacing the second convolutional layer in the pre-trained SqueezeNet convolutional neural network with a fully connected layer with a 1×1 convolutional kernel. The output channel of this fully connected layer is set to the number of label categories, and the trainable parameters of this fully connected layer are initialized to random values, that is, the fault diagnosis network model is obtained; Step 4: Train the fault diagnosis network model Step 4.1: Randomly and non-repeatedly divide the training data set in Step 2 into a training set and a validation set according to the number of data, with a certain proportional amount. Both the training set and the validation set contain training data under different fault states and different working conditions; Step 4.2: Set the initial learning rate to 0.0001, and set the maximum number of training epochs and the maximum number of iterations per epoch; Step 4.3: Initialize the trainable parameters of the fully connected layer of the fault diagnosis network model to random values, and fix the remaining trainable parameters to the pre-trained values of the pre-trained SqueezeNet convolutional neural network; use the two-dimensional RGB image of a piece of training data in the training set in Step 4.1 as the input of the fault diagnosis network model, and the corresponding label as the reference value of the model output. Use cross-entropy loss to calculate the training loss; according to the training loss of a piece of training data, use SGDM as the optimization algorithm to update the trainable parameters of the fully connected layer in the reverse direction once; use the updated network parameters as the initial parameters when training the next piece of training data, input the next piece of training data in the training set, and repeat continuously until the last piece of data in the training set is trained, completing one round of training of the training set; After each round of training is completed, the training data in the validation set in step 4.1 is input into the network that has completed one round of training, and the accuracy of the current fault diagnosis network model's prediction for the validation set is calculated: Wherein, is an indicator function; when the prediction result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i represents the i-th training data in the validation set; The network parameters after one round of training are used as the initial parameters for the next round of training, and this is repeated continuously. When the change in the accuracy of the fault diagnosis network model's prediction for the validation set after 8 consecutive rounds of training is no greater than 0.01%, or when the number of training rounds reaches the preset value, the training of the first cycle is completed, and the initial training values of the trainable parameters of the fully connected layer are obtained; Step 4.4: Use the initial training values of the trainable parameters of the fully connected layer of the fault diagnosis network model and the pre-training values of the other trainable parameters as the initial values of the network model; repeat the process of dividing the training set and the validation set in step 4.1 to obtain a new training set and a new validation set; use the two-dimensional RGB image of a piece of training data in the new training set as the input of the fault diagnosis network model, use its label as the reference value of the model output, and use cross-entropy loss to calculate the training loss; according to the training loss of a piece of training data, use SGDM as the optimization algorithm to update all the trainable parameters of the fault diagnosis network model in the reverse direction once; use the updated network parameters as the initial parameters when training the next piece of training data, input the next piece of training data in the new training set, and repeat continuously until the last piece of training data in the new training set is trained, and one round of training of the new training set is completed; After each round of training is completed, the training data in the new validation set is input into the network model that has completed one round of training, and the accuracy of the current fault diagnosis network model's prediction for the new validation set is calculated: In the formula, is an indicator function; when the prediction result is equal to the true result, the function value is 1, otherwise it is 0. Here, N represents the number of training data in the validation set, and i represents the i-th training data in the validation set; The network parameters after one round of training are used as the initial parameters for the next round of training, and this is repeated continuously. When the change in the accuracy of the fault diagnosis network model's prediction for the new validation set after 8 consecutive rounds of training is no greater than 0.01%, or when the number of training rounds reaches the preset value, the training of the second cycle is completed; Step 4.5: Use the network parameters after the second cycle of training as the initial parameters for the next cycle of training, and repeat the process of step 4.4 until the prediction accuracy of the network model for the validation set after one cycle of training is not less than 99% or the maximum number of cycles is reached, and obtain the fault diagnosis network model that has completed training; Step Five, Rolling Bearing Fault Diagnosis Obtain the acceleration vibration signal of a certain duration of the rolling bearing to be diagnosed, and convert it into a two-dimensional RGB image according to the method in step two; then input this two-dimensional RGB image into the fault diagnosis network model that has completed training in step three, and select the category corresponding to the maximum probability as the fault type of this rolling bearing according to the prediction probabilities of different types output by the network model, and obtain the fault diagnosis result of this rolling bearing.

2. The rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that, In step one, the preprocessing is noise reduction.

3. A rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that, In step one, the fault states specifically include three types: normal state, inner ring fault, and outer ring fault.

4. A rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that, In step one, the working conditions specifically include sampling frequency, rotational frequency, and load.

5. A rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that, The specific content of the second step is as follows: First, the acceleration vibration signal x(t) of each piece of data is divided into multiple segments, and the segmentation principle is defined as follows: where L interval is the data length of each segment of the signal, and L interval = 0.0512f s , where f s is the sampling frequency of this data; N here is the number of segments into which the acceleration vibration signal x(t) of this data can be divided, || represents rounding, and t represents the sampling time point of the acceleration vibration signal; a segment of the acceleration vibration signal x(t) of a piece of data can be expressed as: x k \(l(t)=x[L interval (k - 1)+1:L interval k], k = 1, 2, …, N where x k (t) is the vibration signal of the k-th segment of the acceleration vibration signal x(t); Then, continuous wavelet transform is respectively performed on each segment of the vibration signal to obtain wavelet coefficients: where a is a scale factor representing frequency compression or expansion; b is a time shift factor representing the shift of the wavelet function in time; * represents the operation of taking the complex conjugate, and ψ a,b (t) is a wavelet function, which is the translation and scale change form of the mother wavelet: The absolute value of the wavelet coefficient is taken as the energy distribution of this segment of the vibration signal: W k (a,b) = |CWT k (a,b)| Then, the wavelet coefficients of each segment of the vibration signal are flipped and normalized, and mapped to an RGB image: Where, W k,min and W k,max respectively refer to the minimum value and the maximum value in the wavelet coefficient matrix of the vibration signal of this section; A two-dimensional RGB image is generated from the N segments of vibration signals of one piece of data, and then the generated two-dimensional RGB image is scaled to 224×224×3 to obtain one piece of training data, completing the process of converting the acceleration vibration signal x(t) of one piece of data into a two-dimensional RGB image; The acceleration vibration signal x(t) of each piece of data in the rolling bearing fault diagnosis dataset is subjected to continuous wavelet transform to generate a two-dimensional RGB image, obtaining a training dataset.

6. A rolling bearing fault diagnosis method based on small sample data according to claim 5, characterized in that, The mother wavelet is one of the Mexican hat wavelet, Morlet wavelet, and Ricker wavelet.

7. A rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that In step three, the pre-trained SqueezeNet convolutional neural network refers to the publicly available pre-trained SqueezeNet convolutional neural network whose trainable parameter values are trained on more than one million publicly available images.

8. A rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that, In step four, the ratio of the number of data in the training set to the number of data in the validation set is 6:

4.

9. A rolling bearing fault diagnosis method based on small sample data according to claim 1, characterized in that, In step four, the maximum number of training epochs is 4, and the maximum number of iterations per epoch is 40.

Images