Mechanical equipment fault diagnosis method based on improved OfficientNetV2 network

Through the improved EfficientNetV2 network, the CA coordinate attention mechanism and lightweight structure are used to solve the accuracy and efficiency of traditional mechanical equipment fault diagnosis methods in complex data processing, and achieve higher fault diagnosis accuracy and generalization .

CN118918436APending Publication Date: 2024-11-08SHAOXING ZHILIAN TECH CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411007353.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Traditional mechanical equipment fault diagnosis methods are difficult to meet actual needs when processing complex data and diversified fault modes, and rely on expert experience and complex signal processing.

Method used

Using the improved EfficientNetV2 network, the vibration signal is converted into time-frequency images through the Markov transfer field, and the number of channels and stacking times are reduced in the EfficientNetV2-S network, and the SE channel attention is replaced as the CA coordinate attention mechanism to build a lightweight model to improve the accuracy and generalization of fault diagnosis.

Benefits of technology

The improved EfficientNetV2 network significantly improves accuracy and generalization in mechanical equipment fault diagnosis, reduces calculation costs, is suitable for the classification of complex fault characteristics, and shows excellent performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918436A_ABST
    Figure CN118918436A_ABST
Patent Text Reader

Abstract

The invention discloses a mechanical equipment fault diagnosis method based on an improved OfficientNetV2 network, and the method comprises the following steps: obtaining a vibration signal of mechanical equipment, constructing a data set, and converting an original signal into a time-frequency image through a Markov transfer field; dividing the generated time-frequency image into a training set and a test set; the method comprises the following steps of: constructing an improved OfficientNetV2 network model; inputting the training set into an improved OfficientNetV2 network model for training, and storing an optimal model after the training is finished; and inputting the test set into the trained model for fault classification. According to the method, the structure of the OfficientNetV2-S network is lightened, meanwhile, a CA coordinate attention mechanism is adopted in the MBConv module to replace an original SE channel attention mechanism, the improved OfficientNetV2 network reduces the model complexity, reduces the calculation cost, can be better suitable for the fault diagnosis and classification problem of mechanical equipment, and improves the fault diagnosis and classification efficiency of the mechanical equipment. And the method has higher accuracy and generalization in the face of relatively complex fault features.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to a mechanical equipment fault diagnosis method based on an improved EfficientNetV2 network. Background Art

[0002] With the development of modern industry, mechanical equipment is widely used in various industrial scenarios, such as motors, fans, pumps, machine tools, production lines, etc. Its performance and life directly affect the stability and safety of the entire system. However, since mechanical equipment is subjected to complex loads and environmental conditions during operation, it is prone to wear, fatigue, poor lubrication and other faults. Therefore, timely and accurate detection and diagnosis of mechanical equipment failures is of great significance to ensure the normal operation of industrial systems.

[0003] Traditional mechanical equipment fault diagnosis methods mainly include spectrum analysis, time domain analysis, envelope analysis, etc. These methods rely on expert experience and signal processing technology, require a clear understanding of fault characteristics, and usually require complex preprocessing of the collected signals. With the increasing complexity of industrial equipment, traditional methods are unable to cope with large amounts of data and complex fault modes, and the accuracy and efficiency of diagnosis are difficult to meet actual needs.

[0004] Deep learning is a machine learning technology that simulates the human brain neural network for data analysis. It has powerful feature extraction and classification capabilities and can automatically learn and extract fault features from large amounts of data without the need for complex signal preprocessing. Deep learning has achieved remarkable success in areas such as image recognition, speech recognition, and natural language processing, and these successful experiences have promoted its application in the industrial field. Although the mechanical equipment fault diagnosis method based on deep learning has significant advantages in improving diagnostic accuracy and efficiency, there are still some problems and challenges in data processing, model training, and practical applications. Summary of the invention

[0005] In order to solve the deficiencies of the prior art, the present invention proposes a mechanical equipment fault diagnosis method based on an improved EfficientNetV2 network. Through the improved EfficientNetV2 network structure, it has higher accuracy and generalization when dealing with fault diagnosis problems.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network includes the following steps:

[0008] Step 1: Obtain the vibration signal of the mechanical equipment, construct a data set, and transform the original signal into a time-frequency image through the Markov transfer field;

[0009] Step 2: Divide the generated time-frequency image into training set and test set;

[0010] Step 3: Build an improved EfficientNetV2 network model;

[0011] Step 4: Input the training set into the improved EfficientNetV2 network model for training, update the network parameters through back propagation, and save the optimal model after training.

[0012] Step 5: Input the test set into the trained model for fault classification.

[0013] Furthermore, in step 1, the size of the time-frequency image obtained by the Markov transition field is 128 pixels*128 pixels.

[0014] Furthermore, in step 2, the time-frequency image is divided into a training set and a test set according to a 4:1 ratio.

[0015] Furthermore, the specific contents of step 3 are as follows:

[0016] The present invention builds an improved EfficientNetV2 network based on the EfficientNetV2-S network model. First, the number of channels of the Fused-MBConv modules of the first three layers is reduced to 16, 32, and 48, and the number of stacking is reduced to 1, 2, and 2, respectively. The number of channels of the MBConv modules of the last three layers is reduced to 96, 112, and 192, and the number of stacking is reduced to 2, 3, and 4, respectively. At the same time, the 3×3 convolution kernel in the MBConv module is replaced with a 5×5 convolution kernel.

[0017] Secondly, the SE channel attention in the MBConv module is replaced by the CA coordinate attention mechanism; the CA coordinate attention mechanism can capture the feature space position information while paying attention to the relationship between the channels of the feature matrix, thereby enhancing the model's ability to extract key information of complex fault feature matrices.

[0018] Furthermore, in step 4, a cross entropy loss function is adopted, and the Adam optimizer is used to optimize the network model.

[0019] Furthermore, the global average pooling of the CA coordinate attention mechanism is decomposed into two groups of operations; for a given input X, two pooling kernels of size (H, 1) and (1, W) are used on each channel to perform encoding calculations in the horizontal and vertical directions respectively, and the output of the cth channel with a height of h can be expressed as:

[0020]

[0021] The output of the cth channel with width w can be expressed as:

[0022]

[0023] The two structures obtained from the above transformation Splice them together, and then operate the spliced ​​feature map through a 1×1 convolution transformation function, the formula is:

[0024] f=δ(F1([z h ,z w ]))

[0025] Among them [z h ,z w ] represents the splicing of spatial dimensions, F1 is a 1×1 convolution transformation function, δ is a nonlinear activation function, f∈R C / r×(H+W) Represents the feature map obtained by encoding the spatial information in the horizontal and vertical directions, where r is the scaling rate; split f into two independent horizontal and vertical feature maps, one with a width of 1 and the other with a height of 1. The two feature maps are transformed by a 1×1 convolution:

[0026] g h =σ(F h (f h ))

[0027] gw = σ(F w (f w ))

[0028] where f h ∈R C / r×H and f w ∈R C / r×W These are the two feature maps decomposed from f, F h and F w are two 1×1 convolution transformation functions, σ is the Sigmoid activation function; the two outputs g h and g w They are expanded and used as attention weights respectively, and multiplied with the input X to obtain the output of the CA attention mechanism.

[0029] The beneficial effects of the present invention are as follows:

[0030] The present invention makes the EfficientNetV2-S network structure lightweight, and uses the CA coordinate attention mechanism to replace the original SE channel attention mechanism in the MBConv module, and proposes a mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network. The improved EfficientNetV2 network reduces the model complexity and the computational cost, and can be better applied to the mechanical equipment fault diagnosis classification problem, and has higher accuracy and generalization when facing more complex fault characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a main flow chart of the present invention.

[0032] Figure 2 It is a schematic diagram of the CA coordinate attention mechanism of the present invention.

[0033] Figure 3 Schematic diagram of the CA-MBConv module of the present invention.

[0034] Figure 4 This is a schematic diagram of the improved EfficientNetV2 network structure of the present invention.

[0035] Figure 5 This is the confusion matrix diagram of this example on data set A.

[0036] Figure 6 It is an overall schematic diagram of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0038] refer to Figure 1 and Figure 6 As shown in the figure, the mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network includes the following steps:

[0039] Step 1: Obtain the rolling bearing vibration signal, construct a data set, and transform the original signal into a time-frequency image through the Markov transfer field;

[0040] Specifically, this embodiment uses the Case Western Reserve University bearing data set, which is divided into four categories: normal baseline data, 12kHz drive-end bearing fault data, 48kHz drive-end bearing fault data, and fan-end bearing fault data. The experimental equipment works under four different conditions of motor load: 0HP, 1HP, 2HP, and 3HP. The bearing faults include three types: inner ring, ball, and outer ring. Each fault type includes three fault diameters: 0.007 inches, 0.014 inches, and 0.021 inches. This example constructs three data sets A, B, and C according to the motor load, corresponding to three different working conditions of motor load of 1HP, 2HP, and 3HP, respectively. Each data set uses normal baseline data and 48kHz drive-end bearing fault data. There are a total of 10 categories of normal types and fault types, which are divided according to the length of each sample of 512. 500 samples are selected for each category and labeled. These samples are converted into 128 pixels * 128 pixels of time-frequency images through the Markov transfer field. The details of data sets A, B, and C are shown in Table 1 below.

[0041] Table 1 Datasets under different working conditions

[0042] Dataset Motor load Speed / rpm Sample length Sample size A 1HP 1772 512 5000 B 2HP 1750 512 5000 C 3HP 1730 512 5000

[0043] Step 2: Divide the generated time-frequency image into training set and test set;

[0044] Specifically, in this example, all time-frequency images are shuffled and then divided into training set and test set in a ratio of 4:1.

[0045] Step 3: Build an improved EfficientNetV2 network model;

[0046] Specifically, this example is based on the EfficientNetV2-S network model. First, the number of channels of the Fused-MBConv modules of the first three layers are reduced to 16, 32, and 48, respectively, and the number of stacking times is reduced to 1, 2, and 2, respectively; the number of channels of the MBConv modules of the last three layers are reduced to 96, 112, and 192, and the number of stacking times is reduced to 2, 3, and 4, respectively; at the same time, the 3×3 convolution kernel in the MBConv module is replaced with a 5×5 convolution kernel.

[0047] Secondly, the SE channel attention in all MBConv modules is replaced by CA coordinate attention. The global average pooling of the CA coordinate attention mechanism is decomposed into two groups of operations; for a given input X, two pooling kernels of size (H, 1) and (1, W) are used on each channel to perform encoding calculations in the horizontal and vertical directions respectively, and the output of the cth channel with a height of h can be expressed as:

[0048]

[0049] The output of the cth channel with width w can be expressed as:

[0050]

[0051] The two structures obtained from the above transformation Splice them together, and then operate the spliced ​​feature map through a 1×1 convolution transformation function, the formula is:

[0052] f=δ(F1([z h ,z w ]))

[0053] Among them [z h ,z w ] represents the splicing of spatial dimensions, F1 is a 1×1 convolution transformation function, δ is a nonlinear activation function, f∈RC / r×(H+W) Represents the feature map obtained by encoding the spatial information in the horizontal and vertical directions, where r is the scaling rate; split f into two independent horizontal and vertical feature maps, one with a width of 1 and the other with a height of 1. The two feature maps are transformed by a 1×1 convolution:

[0054] g h =σ(F h (f h ))

[0055] g w =σ(F w (f w ))

[0056] where f h ∈R C / r×H and f w ∈R C / r×W These are the two feature maps decomposed from f, F h and F w are two 1×1 convolution transformation functions, σ is the Sigmoid activation function; the two outputs g h and g w They are expanded and used as attention weights respectively, and multiplied with the input X to get the output of the CA attention mechanism. The CA coordinate attention can be referred to Figure 2 .

[0057] CA-MBConv module can be found at Figure 3 The improved EfficientNetV2 network structure can be referenced Figure 4 The improved EfficientNetV2 network structure parameters are shown in Table 2 below.

[0058] Table 2 Improved EfficienNetV2 network structure parameters

[0059]

[0060]

[0061] Step 4: Input the training set into the improved EfficientNetV2 network model for training, update the network parameters through back propagation, and save the optimal model after training.

[0062] Specifically, this example uses the cross entropy loss function and uses the Adam optimizer to optimize the network model. The batch size is set to 32, the number of training times is set to 100, and the learning rate is set to 0.01. The model is randomly initialized, the network parameters are updated through back propagation, and the best model is saved after training is completed.

[0063] Step 5: Input the test set into the trained model for fault classification.

[0064] Specifically, this example loads the best model saved, inputs the test set, and analyzes the model performance through the evaluation indicators of classification result accuracy, confusion matrix, precision, recall rate, and F1 score.

[0065] Based on the technical solution provided in the above example, the following results are obtained

[0066] Table 3 Performance indicators of the improved EfficienNetV2 network model on the A test set

[0067] Category Label Accuracy Recall F1 score Sample size 0 1.00 1.00 1.00 46 1 1.00 1.00 1.00 43 2 0.88 0.88 0.88 57 3 0.98 0.96 0.97 50 4 0.92 0.81 0.86 57 5 1.00 0.98 0.99 54 6 0.89 0.93 0.91 58 7 1.00 1.00 1.00 50 8 0.76 0.85 0.80 41 9 0.98 1.00 0.99 44

[0068] In this example, Table 3 shows the precision, recall, and F1 score of each category label obtained by inputting the A test set into the improved EfficientNetV2 network model. Figure 5 The confusion matrix on the A test set is shown. From Table 3, we can see that the precision, recall and F1 scores of most category labels are close to 1.00, which shows that the improved EfficienNetV2 network model shows good performance in the classification results of samples.

[0069] Table 4 Accuracy of the model under different data sets

[0070] Model Dataset A Dataset B Dataset C Inception-v3 86.32 85.27 86.18 VGG16 86.76 86.10 85.94 ResNet-50 89.27 89.04 90.43 EfficientNet-B4 82.92 82.61 83.34 Improved EfficientNetV2 93.80 93.45 94.40

[0071] Table 5 Accuracy of the model under different working conditions

[0072]

[0073] In this example, Table 4 shows the accuracy of the improved EfficientNetV2 network and other image classification network models under different data sets. It can be seen from the table that the accuracy of the improved EfficientNetV2 network on the A, B, and C test sets is higher than that of the other four network models. Table 5 shows the accuracy of the improved EfficientNetV2 network and other image classification network models under different working conditions. AB in the table means that the model is trained using the A training set and the model is tested using the test set B. It can be seen from the table that the improved EfficientNetV2 network has better fault diagnosis capabilities under different working conditions than the other four network models.

[0074] It can be concluded that the improved EfficientNetV2 network model has a higher accuracy in rolling bearing fault diagnosis, and shows excellent performance in the face of more complex fault characteristics and different working conditions.

Claims

1. A mechanical equipment fault diagnosis method based on an improved EfficientNetV2 network, characterized in that: The steps include: Step 1: Construct a data set: collect equipment vibration signals and transform the original signals into time-frequency images through Markov transfer fields; Step 2: Divide the generated time-frequency image into training set and test set; Step 3: Build an improved EfficientNetV2 network model; Step 4: Input the training set into the improved EfficientNetV2 network model for training, and save the optimal model parameters after training. Step 5: Input the test set into the trained model for fault classification.

2. The mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network according to claim 1 is characterized in that: In the step 1, the size of the time-frequency image obtained by the Markov transition field is 128 pixels*128 pixels.

3. The mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network according to claim 1 is characterized in that: In step 2, the time-frequency image is divided into a training set and a test set according to a 4:1 ratio, and labels are set.

4. The mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network according to claim 1 is characterized in that: The specific contents of step 3 are as follows: Based on the EfficientNetV2-S network model, the number of channels of the Fused-MBConv modules of the first three layers is reduced to 16, 32, and 48, respectively, and the number of stacking times is reduced to 1, 2, and 2, respectively. The number of channels of the MBConv modules of the last three layers is reduced to 96, 112, and 192, and the number of stacking times is reduced to 2, 3, and 4, respectively. At the same time, the 3×3 convolution kernel in the MBConv module is replaced with a 5×5 convolution kernel. Secondly, the SE channel attention in the MBConv module is replaced by the CA coordinate attention mechanism.

5. The mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network according to claim 1 is characterized in that: In step 4, a cross entropy loss function is adopted, and the Adam optimizer is used to optimize the network model.

6. The mechanical equipment fault diagnosis method based on the improved EfficientNetV2 network according to claim 4 is characterized in that: The global average pooling of the CA coordinate attention mechanism is decomposed into two groups of operations; for a given input X, two pooling kernels of size (H, 1) and (1, W) are used on each channel to perform encoding calculations in the horizontal and vertical directions respectively, and the output of the cth channel with a height of h is expressed as: The output of the cth channel with width w is expressed as: The two structures obtained from the above transformation Splice them together, and then operate the spliced ​​feature map through a 1×1 convolution transformation function, the formula is: f=δ(F1([z h ,z w ])) Among them [z h ,z w ] represents the splicing of spatial dimensions, F1 is a 1×1 convolution transformation function, δ is a nonlinear activation function, f∈R C / r×(H+W) represents the feature map obtained by encoding the spatial information in the horizontal and vertical directions, and r is the scaling rate; split f into a horizontal independent feature map and a vertical independent feature map, where the width of one feature map is 1 and the height of the other is 1. The two feature maps are transformed by a 1×1 convolution respectively: g h =σ(F h (f h )) g w =σ(F w (f w )) Among them, f h ∈R C / r×H and f w ∈R C / r×W are two independent feature maps decomposed from f, F h and F w are two 1×1 convolution transformation functions, σ is the Sigmoid activation function; the two outputs g h and g w They are expanded and used as attention weights respectively, and multiplied with the input X to obtain the output of the CA attention mechanism.

Citation Information

Cited By

  • Health monitoring data anomaly detection method integrating improved attention mechanism and residual network structure

    CN121170523A