Intelligent image classification method based on big data

Abstract

The invention discloses an intelligent image classification method based on big data, and relates to the technical field of image classification, and the method comprises the steps: image feature extraction: employing a deep learning algorithm to integrate and classify the extracted features, the method has the advantages that information of two different physical quantities can be integrated into a visual representation by constructing a vibration and temperature data fusion time-frequency diagram, and vibration data generally reflect dynamic characteristics of a mechanical system, such as running balance of a bearing, collision or friction between parts and the like; temperature data is related to energy loss, heat conduction, potential fault heating and the like, associated information of mechanical vibration and heat change in time and frequency can be captured at the same time through fusion of a time-frequency diagram, a more comprehensive view angle is provided for equipment state monitoring and fault diagnosis, and when a fault occurs, the fault diagnosis accuracy is improved. Vibration and temperature signals often change at the same time, and when a bearing has a local abrasion fault, the vibration amplitude may be increased.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and in particular to an intelligent image classification method based on big data. Background Art

[0002] Bearings are widely used in modern industry in the fields of machinery manufacturing, automobile industry, aerospace, etc., and play an important role in mechanical connection, support and friction reduction. With the development of industrial technology, the requirements for mechanical equipment are getting higher and higher, and the application of bearings is becoming more and more extensive. Bearings are one of the most vulnerable parts, and the service status of bearings will directly affect the operation of mechanical equipment. This puts forward higher requirements for bearing status monitoring and fault diagnosis. In engineering practice, once a bearing fails, the mechanical equipment will not operate normally. Therefore, real-time fault monitoring and diagnosis of bearings is necessary;

[0003] Mechanical equipment is accompanied by abnormal vibration changes in the fault state, and the vibration generated by the bearing during operation is nonlinear and non-stationary. When the bearing fails, it is affected by factors such as nonlinearity and friction, and the non-stationary characteristics of its vibration signal will be more obvious. The time-frequency diagram can simultaneously display the information of the signal in the time and frequency domains, which is very effective for analyzing the non-stationary vibration signal generated by the bearing failure. By identifying the time-frequency diagram of the bearing failure, the fault type (such as inner ring failure, outer ring failure, rolling element failure, etc.) and the degree of the fault can be discovered in time;

[0004] Widely used and traditional signal analysis methods, such as the global transformation Fourier transform, are not suitable for processing non-stationary signals. Although the wavelet transform has good time-frequency localization and multi-resolution characteristics for the nonlinear signals generated by bearing vibration, its ability to re-decompose the high-frequency part is weak. The empirical mode decomposition method is suitable for processing nonlinear and non-stationary signal sequences, and there is no need to pre-set any basis functions. However, the "modal aliasing" problem it produces is more prominent, and the endpoint effect is obvious, which has certain limitations in engineering applications. For this reason, we propose an intelligent image classification method based on big data. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent image classification method based on big data.

[0006] In order to solve the problems raised in the above background technology, the present invention provides the following technical solutions: an intelligent image classification method based on big data, including image feature extraction, integrating and classifying the extracted features using a deep learning algorithm, and the specific operation steps of the intelligent image classification method based on big data are as follows:

[0007] Step 1: Use an acceleration sensor and a thermocouple sensor to collect vibration data and temperature data of the bearing when it is working, perform a two-dimensional wavelet transform on the vibration data and temperature data, and construct a time-frequency diagram of the vibration data and temperature data that reflects the bearing status information;

[0008] Step 2: Weightedly fuse the vibration time-frequency graph and the temperature time-frequency graph with the same dimension, and annotate the fused time-frequency graph to form a fused time-frequency graph dataset with accurate labels;

[0009] Step 3: Use the fused time-frequency graph dataset as the training object of the convolutional neural network, and then use the convolutional neural network to test the noise resistance. Add Gaussian white noise to the original vibration data and temperature data to build a convolutional neural network framework with noise resistance.

[0010] Step 4: Use the sparrow optimization algorithm to optimize the convolutional neural network framework, find out the batch sampling number and learning rate parameters of the convolutional neural network framework according to its iteration curve and use them as the evaluation criteria of the loss function;

[0011] Step 5: Build a generator and a discriminator. The generator generates fake features that are different from the real image, and the discriminator distinguishes between real features and fake features.

[0012] Step 6: To verify the convolutional neural network fault diagnosis method, the public bearing fault data set of Case Western Reserve University in the United States is used as a sample to test the convolutional neural network model's ability to identify fault types and degrees.

[0013] As a further solution of the present invention: in the step 1, an acceleration sensor and a thermocouple sensor are installed on the surface of the bearing seat, the sampling frequency of the acceleration sensor is set to 1000Hz-2000Hz, the sampling frequency of the thermocouple sensor = the sampling frequency of the acceleration sensor, the vibration data and temperature data of the bearing are collected in chronological order, the vibration data are decomposed into one-dimensional data by using a two-dimensional discrete wavelet transform and then a one-dimensional discrete wavelet transform is performed, each time a two-dimensional wavelet transform is performed, the original vibration data will be decomposed into four components, namely, a high-frequency component, a vertical component, a horizontal component and a low-frequency component, the original vibration data and temperature data are decomposed into wavelet coefficients at four scales, the wavelet coefficients contain information about the vibration data in time and frequency, and a vibration time-frequency graph TF is constructed using the wavelet coefficients after the two-dimensional wavelet transform. v (t, f).

[0014] As a further solution of the present invention: in the step 1, the temperature data is normalized to a range of 0-1, the normalized temperature data is converted into data with frequency characteristics by differential operation, the temperature difference data is wavelet transformed by discrete wavelet transform, the discrete wavelet transform decomposes the temperature difference data through a low-pass filter and a high-pass filter to obtain approximate coefficients and detail coefficients under four scales, and then a temperature time-frequency diagram TF is constructed with time as the horizontal coordinate and the correspondence between scale and frequency as the vertical coordinate. t (t, f).

[0015] As a further solution of the present invention: in the step 2, the vibration time-frequency diagram and the temperature time-frequency diagram with the same dimension are weighted fused, and the fusion formula is:

[0016]

[0017] Among them, TF fusion (t, f) represents the fused time-frequency graph, H v Represents the information entropy of the vibration data time-frequency diagram, H t The information entropy of the time-frequency graph representing the temperature data is taken as input, and the convolutional neural network model is used to learn the features in the fused time-frequency graph. The extracted fused time-frequency graphs are collected and sorted, and labeled according to categories. The fused time-frequency graph in the normal state is marked as 0, the fused time-frequency graph in the inner circle is marked as 1, and the fused time-frequency graph in the outer circle is marked as 2, forming a fused time-frequency graph dataset with accurate labels.

[0018] As a further solution of the present invention: in the step 3, the entire fused time-frequency graph data set is divided into a training set, a validation set and a test set according to a ratio, and the division ratio is set to 70% for the training set, 15% for the validation set, and 15% for the test set. The GoogleNet light neural network trained using bearing vibration data and temperature data is loaded into the convolutional neural network model for preliminary training, and Gaussian noise with an integer signal-to-noise ratio between 10 and 20 is added to the fused time-frequency graph data in the test set. The test set data with added Gaussian noise is sequentially input into the preliminarily trained convolutional neural network model, and the number of samples simultaneously input into the convolutional neural network model in one training iteration is set to 30-35, and the learning rate is set to 10 -4 The number of iterations is set to 5-15 times, and the accuracy, recall rate and F1 value evaluation indicators of the convolutional neural network model under these noise interferences are calculated. The changes in the performance indicators of the convolutional neural network model with the increase of noise intensity are observed, and the noise-resistant convolutional neural network framework is obtained after training.

[0019] As a further solution of the present invention: the convolutional neural network framework is optimized using the sparrow optimization algorithm, and the iteration curve of the sparrow optimization algorithm is plotted with the number of iterations as the horizontal axis and the fitness value as the vertical axis. By observing the convergence speed of the curve, the number of batch sampling and the learning rate combination found are used as the evaluation criteria of the loss function of the convolutional neural network framework on the validation set, and the fused time-frequency graph data in the training set are input into the constructed convolutional neural network framework in batches. After each batch of data is input, the convolutional neural network framework calculates the difference between the predicted result and the true label according to the preset loss function. During the training process, the fused time-frequency graph with labels is fused with information entropy features and input into the convolutional neural network framework, and the output result is obtained through forward propagation. The calculation formula of fused information entropy is:

[0020]

[0021] Among them, H combined represents the fusion information entropy, α represents the weight, n represents the number of energy value division intervals in the fusion time-frequency graph, M represents the total number of elements in the fusion time-frequency graph, and n a Represents the number of elements in the vibration time-frequency diagram, n b Represents the number of elements in the temperature time-frequency graph division interval, and then uses the back propagation algorithm to update the various parameters of the model. The whole process lasts for 50-100 training rounds until the set number of training rounds is reached and the training is stopped.

[0022] As a further solution of the present invention: in the step five, a generator and a discriminator are constructed, a real image is input into the discriminator, and the loss of the discriminator judging the real image as real is calculated. Then, the generator is allowed to generate a fake image, and the fake image is input into the discriminator, and the loss of the discriminator judging the fake image as fake is calculated. The total discriminator loss is the sum of these two parts of the loss. The parameters of the discriminator are updated through back propagation to generate a new random noise vector.

[0023] As a further solution of the present invention: in the step five, a fake image is generated by the generator, and the discriminator judges the generated fake image as a real image. The loss of the generator is calculated according to the output of the discriminator on these fake images, and then the parameters of the generator are updated by back propagation. After 10-15 rounds of training, the generator learns to generate fake images with different features from real images, and the discriminator learns the ability to distinguish between real images and fake images. During the training process, the convolutional neural network model is trained against the generator and the discriminator. The convolutional neural network model attempts to use features for accurate classification, while the generator and the discriminator are continuously optimized to make the fake features more difficult to distinguish.

[0024] As a further solution of the present invention: in step six, a standard bearing fault data set published by Case Western Reserve University in the United States is used. The data set includes inner ring faults, outer ring faults and rolling element faults of the bearing, and also includes vibration data and temperature data under normal bearing operation as a control. The data set is used to test the convolutional neural network model's ability to identify fault types and degrees.

[0025] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. The present invention can integrate the information of two different physical quantities into a visual representation by constructing a fusion time-frequency diagram of vibration and temperature data. Vibration data usually reflects the dynamic characteristics of mechanical systems, such as the running balance of bearings, collision or friction between components, etc. Temperature data is related to energy loss, heat conduction, and potential fault heating. By fusing the time-frequency diagram, the correlation information of mechanical vibration and thermal changes in time and frequency can be captured at the same time, providing a more comprehensive perspective for equipment status monitoring and fault diagnosis. When a fault occurs, the vibration and temperature signals often change at the same time. When a local wear fault occurs in a bearing, the vibration amplitude may increase. At the same time, due to frictional heat generation, the temperature will also rise. The fusion time-frequency diagram can highlight the joint feature changes under these fault conditions, making the performance of the fault features more obvious in the time-frequency diagram. Compared with the use of vibration or temperature time-frequency diagrams alone, the fusion time-frequency diagram can better show the overall picture of the fault features and reduce the possibility of missed diagnosis and misdiagnosis.

[0027] 2. The present invention can decompose vibration and temperature data at different scales through two-dimensional wavelet transform. In the low-frequency part, the general trend of the signal can be captured, such as the overall temperature rise trend of the equipment or the low-frequency component of the vibration. In the high-frequency part, transient changes and detailed information in the signal can be found, such as shock pulses in vibration or rapid temperature fluctuations. This multi-resolution feature helps to explore the intrinsic connection between vibration and temperature data at different scales. In actual environments, both vibration and temperature data may be interfered by noise. The two-dimensional wavelet transform can separate noise components from useful signal components by selecting appropriate wavelet basis functions and decomposition layers. For vibration data, the wavelet transform can remove high-frequency noise and highlight features such as fault impact. For temperature data, it can smooth out measurement noise and highlight the real trend of temperature change. In the process of constructing the fused time-frequency diagram, the vibration and temperature data after denoising by wavelet transform can better show the real signal characteristics and improve the quality of the fused time-frequency diagram. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a flow chart of an intelligent image classification method based on big data in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] The present invention provides an intelligent image classification method based on big data, including image feature extraction, integrating and classifying the extracted features using a deep learning algorithm, and the specific operation steps of the intelligent image classification method based on big data are as follows:

[0031] Step 1: Use an acceleration sensor and a thermocouple sensor to collect vibration data and temperature data of the bearing when it is working, perform a two-dimensional wavelet transform on the vibration data and temperature data, and construct a time-frequency diagram of the vibration data and temperature data that reflects the bearing status information;

[0032] Step 2: Weightedly fuse the vibration time-frequency graph and the temperature time-frequency graph with the same dimension, and annotate the fused time-frequency graph to form a fused time-frequency graph dataset with accurate labels;

[0033] Step 3: Use the fused time-frequency graph dataset as the training object of the convolutional neural network, and then use the convolutional neural network to test the noise resistance. Add Gaussian white noise to the original vibration data and temperature data to build a convolutional neural network framework with noise resistance.

[0034] Step 4: Use the sparrow optimization algorithm to optimize the convolutional neural network framework, find out the batch sampling number and learning rate parameters of the convolutional neural network framework according to its iteration curve and use them as the evaluation criteria of the loss function;

[0035] Step 5: Build a generator and a discriminator. The generator generates fake features that are different from the real image, and the discriminator distinguishes between real features and fake features.

[0036] Step 6: To verify the convolutional neural network fault diagnosis method, the public bearing fault data set of Case Western Reserve University in the United States is used as a sample to test the convolutional neural network model's ability to identify fault types and degrees.

[0037] In one embodiment of the present invention: in step 1, an acceleration sensor and a thermocouple sensor are installed on the surface of the bearing seat, the sampling frequency of the acceleration sensor is set to 1000Hz-2000Hz, the sampling frequency of the thermocouple sensor = the sampling frequency of the acceleration sensor, the vibration data and temperature data of the bearing are collected in chronological order, the vibration data are decomposed into one-dimensional data by using a two-dimensional discrete wavelet transform, and then a one-dimensional discrete wavelet transform is performed. Each time a two-dimensional wavelet transform is performed, the original vibration data will be decomposed into four components, namely, a high-frequency component, a vertical component, a horizontal component and a low-frequency component. The original vibration data and temperature data are decomposed into wavelet coefficients at four scales, and the wavelet coefficients contain information about the vibration data in time and frequency. The wavelet coefficients after the two-dimensional wavelet transform are used to construct a vibration time-frequency graph TF v (t, f).

[0038] In one embodiment of the present invention, in step 1, the temperature data is normalized to a range of 0-1, and the normalized temperature data is converted into data with frequency characteristics by differential operation. The temperature difference data is wavelet transformed by discrete wavelet transform. The discrete wavelet transform decomposes the temperature difference data through a low-pass filter and a high-pass filter to obtain approximate coefficients and detail coefficients under four scales, and then a temperature time-frequency diagram TF is constructed with time as the horizontal coordinate and the correspondence between scale and frequency as the vertical coordinate. t (t, f).

[0039] In one embodiment of the present invention: in step 2, the vibration time-frequency diagram and the temperature time-frequency diagram with the same dimension are weighted fused, and the fusion formula is:

[0040]

[0041] Among them, TF fusion (t, f) represents the fused time-frequency graph, H v Represents the information entropy of the vibration data time-frequency diagram, H t The information entropy of the time-frequency graph representing the temperature data is taken as input, and the convolutional neural network model is used to learn the features in the fused time-frequency graph. The extracted fused time-frequency graphs are collected and sorted, and labeled according to categories. The fused time-frequency graph in the normal state is marked as 0, the fused time-frequency graph in the inner circle is marked as 1, and the fused time-frequency graph in the outer circle is marked as 2, forming a fused time-frequency graph dataset with accurate labels.

[0042] In one embodiment of the present invention: in step 3, the entire fused time-frequency graph data set is divided into a training set, a validation set and a test set according to a ratio, and the division ratio is set to 70% for the training set, 15% for the validation set, and 15% for the test set. The GoogleNet light neural network trained using bearing vibration data and temperature data is loaded into the convolutional neural network model for preliminary training, and Gaussian noise with an integer signal-to-noise ratio between 10 and 20 is added to the fused time-frequency graph data in the test set. The test set data with added Gaussian noise is sequentially input into the preliminarily trained convolutional neural network model, and the number of samples simultaneously input into the convolutional neural network model in one training iteration is set to 30-35, and the learning rate is set to 10 -4 The number of iterations is set to 5-15 times, and the accuracy, recall rate and F1 value evaluation indicators of the convolutional neural network model under these noise interferences are calculated. The changes in the performance indicators of the convolutional neural network model with the increase of noise intensity are observed, and the noise-resistant convolutional neural network framework is obtained after training.

[0043] In one embodiment of the present invention: in step 4, the convolutional neural network framework is optimized using the sparrow optimization algorithm, and the iteration curve of the sparrow optimization algorithm is plotted with the number of iterations as the horizontal axis and the fitness value as the vertical axis. By observing the convergence speed of the curve, the number of batch sampling and the learning rate combination found are used as the evaluation criteria of the loss function of the convolutional neural network framework on the validation set, and the fused time-frequency graph data in the training set are input into the constructed convolutional neural network framework in batches. After each batch of data is input, the convolutional neural network framework calculates the difference between the predicted result and the true label according to the preset loss function. During the training process, the fused time-frequency graph with the label is fused with the information entropy feature and input into the convolutional neural network framework, and the output result is obtained through forward propagation. The calculation formula of the fused information entropy is:

[0044]

[0045] Among them, H combined represents the fusion information entropy, α represents the weight, n represents the number of energy value division intervals in the fusion time-frequency graph, M represents the total number of elements in the fusion time-frequency graph, and n a Represents the number of elements in the vibration time-frequency diagram, n b Represents the number of elements in the temperature time-frequency graph division interval, and then uses the back propagation algorithm to update the various parameters of the model. The whole process lasts for 50-100 training rounds until the set number of training rounds is reached and the training is stopped.

[0046] In one embodiment of the present invention: In step five, a generator and a discriminator are constructed, a real image is input into the discriminator, and the loss of the discriminator judging the real image as real is calculated. Then, the generator is allowed to generate a fake image, and the fake image is input into the discriminator, and the loss of the discriminator judging the fake image as fake is calculated. The total discriminator loss is the sum of these two parts of the loss. The parameters of the discriminator are updated through back propagation to generate a new random noise vector.

[0047] In one embodiment of the present invention: In step five, a fake image is generated by a generator, and a discriminator judges the generated fake image as a real image. The loss of the generator is calculated based on the output of the discriminator on these fake images, and then the parameters of the generator are updated by back propagation. After 10-15 rounds of training, the generator learns to generate fake images with different features from real images, and the discriminator learns the ability to distinguish between real images and fake images. During the training process, the convolutional neural network model is trained against the generator and the discriminator. The convolutional neural network model attempts to use features for accurate classification, while the generator and the discriminator are continuously optimized to make the fake features more difficult to distinguish.

[0048] In one embodiment of the present invention: In step six, a standard bearing fault data set published by Case Western Reserve University in the United States is used. The data set includes inner ring faults, outer ring faults and rolling element faults of the bearing, and also includes vibration data and temperature data under normal bearing operating conditions as a control. The data set is used to test the convolutional neural network model's ability to identify fault types and degrees.

[0049] Embodiment 1: The vibration signal amplitude of the bearing in the normal state is small, and the signal amplitude increases in the fault state. However, the time domain waveforms of the vibration signals in some states are relatively similar. In addition, the interference of the external environment makes it difficult to distinguish the fault state of the bearing. Therefore, it is necessary to extract the characteristics of the vibration signal to provide a data set for subsequent convolutional neural network training. Combined with the time-frequency diagram of the vibration signal, a cmor wavelet basis with good adaptability is selected, and a two-dimensional wavelet transform is performed on each sample vibration signal in the normal state, inner ring fault, outer ring fault and rolling element fault. The time-frequency diagram used for convolutional neural network model training is obtained. Considering the actual operation of the mechanical equipment, some time-frequency diagrams with noise are collected for training the fault diagnosis model of the convolutional neural network.

[0050] Embodiment 2: Since temperature data itself does not have an intuitive frequency characteristic like vibration data, it can be converted into data with more "frequency" characteristics through differential operation. The first-order differential formula is ΔT = T i+1 -T i , where T iis the temperature value at the i-th time point. The differential operation can highlight the temperature change trend, which is similar to the frequency component in the vibration signal. During the development of the bearing fault, the first-order difference in the temperature rise stage may be positive and gradually increase.

[0051] Embodiment 3: After each training round, the time-frequency graph data in the validation set is input into the convolutional neural network model, and the accuracy, recall rate, F1 value and other evaluation indicators of the convolutional neural network model on the validation set are calculated, and the changes in model performance are observed. If the model is found to be overfitting, which is manifested as the accuracy on the training set continues to improve, but the accuracy on the validation set begins to decrease, some measures can be taken to adjust it, such as increasing the regularization term and reducing the network complexity, reducing the number of convolutional layers or fully connected layers, the number of neurons, etc. If underfitting occurs, the accuracy on both the training set and the validation set is low, you can consider increasing the amount of training data, increasing the network complexity, or adjusting hyperparameters such as the learning rate.

[0052] As attached Figure 1 As shown in the figure, the time-frequency diagram of the bearing status information is obtained by using two-dimensional wavelet transform, the time-frequency diagram is used as the training object of the convolutional neural network, Gaussian white noise is added to the original vibration data, a convolutional neural network framework with noise resistance is constructed, and the convolutional neural network framework is optimized using the sparrow optimization algorithm, and the optimal batch sampling number and learning rate parameters of the convolutional neural network framework are selected.

[0053] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

[0054] In the description of the specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0055] The above contents are merely examples and explanations of the present invention. Various modifications or additions to the specific embodiments described or replacements in similar ways by technicians in the technical field shall fall within the protection scope of the present invention as long as they do not deviate from the invention or exceed the scope defined by the claims.

[0056] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent image classification method based on big data, including image feature extraction, integrating and classifying the extracted features using a deep learning algorithm, characterized in that: The specific operation steps of the intelligent image classification method based on big data are as follows: Step 1: Use an acceleration sensor and a thermocouple sensor to collect vibration data and temperature data of the bearing when it is working, perform a two-dimensional wavelet transform on the vibration data and temperature data, and construct a time-frequency diagram of the vibration data and temperature data that reflects the bearing status information; Step 2: Weightedly fuse the vibration time-frequency graph and the temperature time-frequency graph with the same dimension, and annotate the fused time-frequency graph to form a fused time-frequency graph dataset with accurate labels; Step 3: Use the fused time-frequency graph dataset as the training object of the convolutional neural network, and then use the convolutional neural network to test the noise resistance. Add Gaussian white noise to the original vibration data and temperature data to build a convolutional neural network framework with noise resistance. Step 4: Use the sparrow optimization algorithm to optimize the convolutional neural network framework, find out the batch sampling number and learning rate parameters of the convolutional neural network framework according to its iteration curve and use them as the evaluation criteria of the loss function; Step 5: Build a generator and a discriminator. The generator generates fake features that are different from the real image, and the discriminator distinguishes between real features and fake features. Step 6: To verify the convolutional neural network fault diagnosis method, the public bearing fault data set of Case Western Reserve University in the United States is used as a sample to test the convolutional neural network model's ability to identify fault types and degrees.

2. The intelligent image classification method based on big data according to claim 1, characterized in that: In the step 1, an acceleration sensor and a thermocouple sensor are installed on the surface of the bearing seat, the sampling frequency of the acceleration sensor is set to 1000Hz-2000Hz, the sampling frequency of the thermocouple sensor=the sampling frequency of the acceleration sensor, the vibration data and the temperature data of the bearing are collected in chronological order, the vibration data are decomposed into one-dimensional data by using a two-dimensional discrete wavelet transform, and then a one-dimensional discrete wavelet transform is performed. Each time a two-dimensional wavelet transform is performed, the original vibration data will be decomposed into four components, namely, a high-frequency component, a vertical component, a horizontal component, and a low-frequency component. The original vibration data and the temperature data are decomposed into wavelet coefficients at four scales, and the wavelet coefficients contain information about the vibration data in time and frequency. The wavelet coefficients after the two-dimensional wavelet transform are used to construct a vibration time-frequency graph TF v (t, f).

3. The intelligent image classification method based on big data according to claim 2, characterized in that: In the step 1, the temperature data is normalized to a range of 0-1, and the normalized temperature data is converted into data with frequency characteristics by using a differential operation. The temperature difference data is wavelet transformed by using a discrete wavelet transform. The discrete wavelet transform decomposes the temperature difference data through a low-pass filter and a high-pass filter to obtain approximate coefficients and detail coefficients under four scales, and then a temperature time-frequency diagram TF is constructed with time as the horizontal coordinate and the corresponding relationship between scale and frequency as the vertical coordinate. t (t, f).

4. The intelligent image classification method based on big data according to claim 3 is characterized in that: In the step 2, the vibration time-frequency diagram and the temperature time-frequency diagram with the same dimension are weighted fused, and the fusion formula is: Among them, TF fusion (t, f) represents the fused time-frequency graph, H v The information entropy of the time-frequency graph of vibration data, H t The information entropy of the time-frequency graph representing the temperature data is taken as input, and the convolutional neural network model is used to learn the features in the fused time-frequency graph. The extracted fused time-frequency graphs are collected and sorted, and labeled according to categories. The fused time-frequency graph in the normal state is marked as 0, the fused time-frequency graph in the inner circle is marked as 1, and the fused time-frequency graph in the outer circle is marked as 2, forming a fused time-frequency graph dataset with accurate labels.

5. The intelligent image classification method based on big data according to claim 4, characterized in that: In the step 3, the entire fused time-frequency graph data set is divided into a training set, a validation set, and a test set according to a ratio, and the division ratio is set to 70% for the training set, 15% for the validation set, and 15% for the test set. The GoogleNet light neural network trained using the bearing vibration data and the temperature data is loaded into the convolutional neural network model for preliminary training, and Gaussian noise with an integer signal-to-noise ratio between 10 and 20 is added to the fused time-frequency graph data in the test set. The test set data with Gaussian noise added is sequentially input into the preliminarily trained convolutional neural network model. The number of samples simultaneously input into the convolutional neural network model in one training iteration is set to 30-35, and the learning rate is set to 10 -4 The number of iterations is set to 5-15 times, and the accuracy, recall rate and F1 value evaluation indicators of the convolutional neural network model under these noise interferences are calculated. The changes in the performance indicators of the convolutional neural network model with the increase of noise intensity are observed, and the noise-resistant convolutional neural network framework is obtained after training.

6. The intelligent image classification method based on big data according to claim 5, characterized in that: In the step 4, the convolutional neural network framework is optimized using the sparrow optimization algorithm, and the iteration curve of the sparrow optimization algorithm is drawn with the number of iterations as the horizontal axis and the fitness value as the vertical axis. By observing the convergence speed of the curve, the number of batch sampling and the learning rate combination found are used as the evaluation criteria of the loss function of the convolutional neural network framework on the validation set, and the fused time-frequency graph data in the training set are input into the constructed convolutional neural network framework in batches. After each batch of data is input, the convolutional neural network framework calculates the difference between the predicted result and the true label according to the preset loss function. During the training process, the fused time-frequency graph with labels is fused with information entropy features and input into the convolutional neural network framework, and the output result is obtained through forward propagation. The calculation formula of fused information entropy is: Among them, H combined represents the fusion information entropy, α represents the weight, n represents the number of energy value division intervals in the fusion time-frequency graph, M represents the total number of elements in the fusion time-frequency graph, and n a Represents the number of elements in the vibration time-frequency diagram, n b Represents the number of elements in the temperature time-frequency graph division interval, and then uses the back propagation algorithm to update the various parameters of the model. The whole process lasts for 50-100 training rounds until the set number of training rounds is reached and the training is stopped.

7. The intelligent image classification method based on big data according to claim 1, characterized in that: In the step five, a generator and a discriminator are constructed, a real image is input into the discriminator, and the loss of the discriminator judging the real image as real is calculated. Then, the generator generates a fake image, and the fake image is input into the discriminator, and the loss of the discriminator judging the fake image as fake is calculated. The total discriminator loss is the sum of these two losses. The parameters of the discriminator are updated through back propagation to generate a new random noise vector.

8. The intelligent image classification method based on big data according to claim 7, characterized in that: In the step five, a fake image is generated by the generator, and the discriminator judges the generated fake image as a real image. The loss of the generator is calculated according to the output of the discriminator on these fake images, and then the parameters of the generator are updated by back propagation. After 10-15 rounds of training, the generator learns to generate fake images with different features from real images, and the discriminator learns the ability to distinguish between real images and fake images. During the training process, the convolutional neural network model is trained against the generator and the discriminator. The convolutional neural network model attempts to use features for accurate classification, while the generator and the discriminator are continuously optimized to make the fake features more difficult to distinguish.

9. The intelligent image classification method based on big data according to claim 5, characterized in that: In step six, a standard bearing fault data set published by Case Western Reserve University in the United States is used. The data set includes inner ring faults, outer ring faults and rolling element faults of the bearing, and also includes vibration data and temperature data under normal bearing operation as a control. The data set is used to test the convolutional neural network model's ability to identify fault types and degrees.

Images