Fault diagnosis method based on parameter optimization and deep learning
By combining the DE-GWO algorithm to optimize the hyperparameters of 1DCNN and adding an attention mechanism to 1DCNN, the problem of insufficient hyperparameter selection dependence on experience and fault feature expression capabilities in the existing technology is solved, and the accuracy and generalization capabilities of fault diagnosis are significantly improved.
Patent Information
- Application Number
- CN202311417544.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-30
- Publication Date
- 2025-05-02
AI Technical Summary
When using one-dimensional convolutional neural network (1DCNN) for fault diagnosis, hyperparameter selection depends on experience and has weak ability to express fault features, resulting in poor diagnostic results.
Combining differential evolution (DE) and gray wolf optimization (GWO) algorithms, a DE-GWO group intelligent optimization algorithm is formed, which is used to optimize the hyperparameters of 1DCNN and an attention mechanism is added to 1DCNN to enhance the ability to extract fault feature information.
By optimizing hyperparameters and enhancing feature extraction capabilities, the accuracy of fault diagnosis and model generalization capabilities are significantly improved, and the accurate identification and diagnosis efficiency of equipment faults is improved.
Smart Images
Figure CN119917849A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fault monitoring and diagnosis, and specifically is a fault diagnosis method based on parameter optimization and deep learning. Background Art
[0002] In the industrial field, equipment fault diagnosis is extremely important to ensure the safety and stable operation of industrial production. Rolling bearings in mechanical equipment are one of the most easily damaged parts due to the harsh working environment and complex working conditions. According to statistics, mechanical failures caused by rolling bearings account for about 30% of mechanical failures of rotating equipment and about 20% of mechanical failures of gearboxes. Rolling bearing condition monitoring and fault diagnosis technology is an important link to understand the performance status of bearings and timely discover potential faults. In addition, it can effectively improve the operation management level and maintenance efficiency of mechanical equipment, thereby significantly improving the economic benefits of enterprises.
[0003] In recent years, with the advancement of science and technology, computer performance has been greatly improved. Machine learning and deep learning algorithms have also been used by experts and scholars in the field of fault diagnosis. Among them, one-dimensional convolutional neural network (1DCNN) has attracted much attention due to its powerful feature extraction ability. However, when 1DCNN is actually used for fault diagnosis, hyperparameters are often selected based on experience, and the ability to express fault features is also weak, which cannot make 1DCNN play the best effect. Therefore, the two optimization algorithms of DE and 1DCNN are combined to optimize the hyperparameters of the 1DCNN model, and an attention mechanism is added to 1DCNN to highlight more useful fault feature information. Then, the optimized 1DCNN is used to establish a fault diagnosis model to diagnose rolling bearing faults. Summary of the invention
[0004] In view of the shortcomings of the prior art, the present invention provides a fault diagnosis method based on parameter optimization and deep learning. Firstly, an attention mechanism is added to 1DCNN for improvement to enhance the fault feature information. Then, a swarm intelligence algorithm DE-GWO based on differential evolution (DE) and grey wolf optimization (GWO) is proposed and used for the hyperparameter optimization of 1DCNN. Finally, the improved 1DCNN and hyperparameter optimization are used in the fault diagnosis of rolling bearings.
[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:
[0006] Based on the fault diagnosis method of parameter optimization and deep learning, the following steps are performed to obtain the optimized and trained fault diagnosis model:
[0007] 1) Obtain rolling bearing fault data and labels and build a data set;
[0008] 2) Organically combine the differential evolution algorithm with the grey wolf optimization algorithm to form the DE-GWO intelligent optimization algorithm;
[0009] 3) Improve the one-dimensional convolutional neural network by adding an attention mechanism;
[0010] 4) Use the DE-GWO algorithm to optimize the hyperparameters in the one-dimensional convolutional neural network and build a fault diagnosis algorithm model;
[0011] 5) Use the training set data to train the fault diagnosis model to identify the characteristics of different fault types;
[0012] 6) Input the training set data into the trained fault diagnosis model to predict the fault type.
[0013] The step 1 includes: one-hot encoding the fault label, combining it with the fault vibration data to form a data set, and dividing it into a training set and a test set in proportion.
[0014] The swarm intelligence optimization algorithm DE-GWO in step 2 is obtained by the following steps:
[0015] Step 2.1: Population initialization
[0016] The algorithm population initialization uses real number coding to randomly generate an initial population of size [N, D]; N represents the number of individuals in the population, D represents the number of decision variables, and the decision variables represent individual genes; the initial population is expressed as:
[0017] X={x 1,D ,x 2,D ,…,x N,D} (1)
[0018] In the formula, X represents the population, x N,D Represents individuals in a population;
[0019] Step 2.2: Mutation Operation
[0020] The mutation operation in the intelligent optimization algorithm is completed by GWO; first, the fitness of each individual in the population is calculated; the individuals with the first three optimal solutions are defined as alpha wolves, represented by α, β and δ respectively, and the other individuals are gray wolves. The gray wolves simulate the hunting process according to the positions of the three alpha wolves and update their own hunting positions; the mathematical formula of this hunting process is as follows:
[0021] D α =|C1X α (t)-X(t)| (2)
[0022] D β =|C2X β(t)-X(t)| (3)
[0023] D δ =|C3X δ (t)-X(t)| (4)
[0024] X1=X α (t)-A1D α (5)
[0025] X2=X β (t)-A2D β (6)
[0026] X3=X δ (t)-A3D δ (7)
[0027]
[0028] Where t is the population iteration round, C1, C2, and C3 represent swing factors, which are usually floating point data types and are randomly set between 0 and 2; D α , D β , D δ represents the distance between the alpha, beta and delta leaders and the gray wolf respectively; A1, A2 and A3 represent convergence factors, which determine whether the gray wolf moves to the position of the alpha leader; X α (t), X β (t), X δ (t) are the positions of the alpha, beta and delta leaders in the tth round; X1, X2, X3 are intermediate variables, and X(t+1) represents the new position of the individual in the population;
[0029] Step 2.3: Crossover Operation
[0030] A vector of offspring used to generate diversity, leading to structural differences in the population;
[0031] Update the crossover individuals as follows:
[0032]
[0033] Among them, u j,i,g represents the jth gene of the i-th individual in the g-th generation crossover population, v j,i,g represents the jth gene of the i-th individual in the g-th generation mutant population, x j,i,g represents the jth gene of the i-th individual in the g-th generation population. Cr is the crossover operator, which is usually a floating-point data type randomly set between 0 and 1.
[0034] Step 2.4: Select an action
[0035] Update the next generation population using a greedy selection operation:
[0036]
[0037] Among them, x i,g+1 is the next generation individual, u i,g is the individual after the crossover operation, x i,g It is the parent individual before mutation and crossover operations, and f represents the fitness function.
[0038] The fault diagnosis model in step 3 includes an input layer, a convolutional layer, a Dropout layer, a pooling layer, an attention layer, a fully connected layer and an output layer connected in sequence; and is used to extract enhanced fault feature information.
[0039] In step 4, the DE-GWO algorithm is used to optimize the number of convolution kernels, convolution step size, pooling layer step size, drop rate, learning rate and batch size in the one-dimensional convolutional neural network.
[0040] In step 5, the training set data is input into the one-dimensional convolutional neural network fault diagnosis model after parameter optimization for training to learn the characteristics of different fault types of the equipment.
[0041] The training is terminated when the set number of times is reached or the loss function converges.
[0042] In step 6, the test set is input into the trained one-dimensional convolutional neural network fault diagnosis model to test the fault diagnosis accuracy of the model.
[0043] The present invention has the following beneficial effects and advantages:
[0044] 1. The method of the present invention can effectively improve the accuracy of fault diagnosis by combining parameter optimization algorithm and deep learning model.
[0045] 2. By optimizing model parameters, it can better adapt to different fault conditions and improve the generalization ability of the model.
[0046] 3. In addition, this method has the advantages of high degree of automation and easy operation, and can be widely used in fault diagnosis of various equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is a flow chart of the fault diagnosis method of the present invention.
[0048] Figure 2 This is the algorithm flow chart of DE-GWO
[0049] Figure 3 This is the model structure diagram of 1DCNN DETAILED DESCRIPTION
[0050] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.
[0051] like Figure 1 Shown is a flow chart of the method of the present invention.
[0052] The fault diagnosis method based on parameter optimization and deep learning first collects the vibration data and labels of the equipment in different operating states to form the original data, and then divides the original data into training data and test data according to the ratio of 8:2. Then, the 1DCNN is improved by adding attention mechanism and other methods. At this stage, the improved 1DCNN hyperparameters are optimized to find the optimal hyperparameter combination. After training the improved model with the training data, the model is tested with the test data and finally applied to the actual fault diagnosis.
[0053] The specific steps of the present invention are as follows:
[0054] Step 1: Collect vibration data of rolling bearings in different operating states, including normal operating data and data of different fault types, and assign different labels to different operating states, such as normal state label 0, inner ring fault label 1, outer ring fault label 2, and rolling element fault label 3. For each fault type, 1024 vibration data and corresponding operating labels are intercepted from the operating data to form a sample. Each fault type has 400 samples, and the training set and test set are divided into a ratio of 8:2.
[0055] The DE-GWO algorithm process in step 2 is as follows Figure 2 As shown, it includes the following steps:
[0056] Step 2.1: Population initialization
[0057] The algorithm population initialization uses real number coding to randomly generate an initial population of size [N, D]; N represents the number of individuals in the population, D represents the number of decision variables, and the decision variables represent individual genes; the initial population is expressed as:
[0058] X={x 1,D ,x 2,D ,…,x N,D} (1)
[0059] In the formula, X represents the population, x N,D Represents individuals in a population;
[0060] Step 2.2: Mutation Operation
[0061] The mutation operation in the intelligent optimization algorithm is completed by GWO; first, the fitness of each individual in the population is calculated; the individuals with the first three optimal solutions are defined as alpha wolves, represented by α, β and δ respectively, and the other individuals are gray wolves. The gray wolves simulate the hunting process according to the positions of the three alpha wolves and update their own hunting positions; the mathematical formula of this hunting process is as follows:
[0062] D α =|C1X α (t)-X(t)| (2)
[0063] D β =|C2X β (t)-X(t)| (3)
[0064] D δ =|C3X δ (t)-X(t)| (4)
[0065] X1=X α (t)-A1D α (5)
[0066] X2=X β (t)-A2D β (6)
[0067] X3=X δ (t)-A3D δ (7)
[0068]
[0069] Where t is the population iteration round, C1, C2, and C3 represent swing factors, which are usually floating point data types and are randomly set between 0 and 2; D α , D β , D δ represents the distance between the alpha, beta and delta leaders and the gray wolf respectively; A1, A2 and A3 represent convergence factors, which determine whether the gray wolf moves to the position of the alpha leader; X α (t), X β (t), X δ (t) are the positions of the alpha, beta and delta leaders in the tth round; X1, X2, X3 are intermediate variables, and X(t+1) represents the new position of the individual in the population;
[0070] Step 2.3: Crossover Operation
[0071] A vector of offspring used to generate diversity, leading to structural differences in the population;
[0072] Update the crossover individuals as follows:
[0073]
[0074] Among them, u j,i,g represents the jth gene of the i-th individual in the g-th generation crossover population, v j,i,g represents the jth gene of the i-th individual in the g-th generation mutant population, x j,i,g represents the jth gene of the i-th individual in the g-th generation population. Cr is the crossover operator, which is usually a floating-point data type randomly set between 0 and 1.
[0075] Step 2.4: Select an action
[0076] Update the next generation population using a greedy selection operation:
[0077]
[0078] Among them, x i,g+1 is the next generation individual, u i,g is the individual after the crossover operation, x i,g It is the parent individual before mutation and crossover operations, and f represents the fitness function.
[0079] The 1DCNN model with the attention mechanism in step 3 is as follows Figure 3 As shown in the figure, the model consists of an input layer, a convolutional layer, a Dropout layer, a pooling layer, an attention layer, a fully connected layer, and an output layer. According to the attention mechanism, features that have a significant impact on the results will be given greater weights. Adding the attention mechanism between the pooling layer and the fully connected layer can effectively enhance the fault feature information.
[0080] In step 4, the DE-GWO algorithm is used to optimize the hyperparameters in the one-dimensional convolutional neural network, such as the number of convolution kernels, convolution step size, pooling layer step size, dropout rate, learning rate, and batch size, which have a great influence on the algorithm effect. The optimized combination of hyperparameters is input into the 1DCNN to build a fault diagnosis model.
[0081] In step 5, the training set data is input into the one-dimensional convolutional neural network fault diagnosis model after parameter optimization for training to learn the characteristics of different fault types of the equipment.
[0082] In step 6, the test set is input into the trained one-dimensional convolutional neural network fault diagnosis model to test the fault diagnosis accuracy of the model.
[0083] In practical applications, on-site bearing data is collected, input into the optimized and trained fault diagnosis model, and the fault type label results are automatically output.
[0084] The above preferred embodiments are only typical examples of the disclosed embodiments. Any equivalent replacement or equivalent changes and modifications made according to the conceptual scheme of this embodiment are within the scope of protection required by this patent.
Claims
1. A fault diagnosis method based on parameter optimization and deep learning, characterized in that: Perform the following steps to obtain the optimized and trained fault diagnosis model: 1) Obtain rolling bearing fault data and labels and build a data set; 2) Organically combine the differential evolution algorithm with the grey wolf optimization algorithm to form the DE-GWO intelligent optimization algorithm; 3) Improve the one-dimensional convolutional neural network by adding an attention mechanism; 4) Use the DE-GWO algorithm to optimize the hyperparameters in the one-dimensional convolutional neural network and build a fault diagnosis algorithm model; 5) Use the training set data to train the fault diagnosis model to identify the characteristics of different fault types; 6) Input the training set data into the trained fault diagnosis model to predict the fault type.
2. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: The step 1 includes: one-hot encoding the fault label, combining it with the fault vibration data to form a data set, and dividing it into a training set and a test set in proportion.
3. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: The swarm intelligence optimization algorithm DE-GWO in step 2 is obtained by the following steps: Step 2.1: Population initialization The algorithm population initialization uses real number coding to randomly generate an initial population of size [N, D]; N represents the number of individuals in the population, D represents the number of decision variables, and the decision variables represent individual genes; the initial population is expressed as: X={x 1,D ,x 2,D ,…,x N,D } (1) In the formula, X represents the population, x N,D Represents individuals in a population; Step 2.2: Mutation Operation The mutation operation in the intelligent optimization algorithm is completed by GWO; first, the fitness of each individual in the population is calculated; the individuals with the first three optimal solutions are defined as alpha wolves, represented by α, β and δ respectively, and the other individuals are gray wolves. The gray wolves simulate the hunting process according to the positions of the three alpha wolves and update their own hunting positions; the mathematical formula of this hunting process is as follows: D α =|C1X α (t)-X(t)| (2) D β =|C2X β (t)-X(t) (3) D δ =|C3X δ (t)-X(t)| (4) X1=X α (t)-A1D α (5) X2=X β (t)-A2D β (6) X3=X δ (t)-A3D δ (7) Where t is the population iteration round, C1, C2, and C3 represent swing factors, which are usually floating point data types and are randomly set between 0 and 2; D α , D β , D δ represents the distance between the alpha, beta and delta leaders and the gray wolf respectively; A1, A2 and A3 represent convergence factors, which determine whether the gray wolf moves to the position of the alpha leader; X α (t), X β (t), X δ (t) are the positions of the alpha, beta and delta leaders in the tth round; X1, X2, X3 are intermediate variables, and X(t+1) represents the new position of the individual in the population; Step 2.3: Crossover operation A vector of offspring used to generate diversity, leading to structural differences in the population; Update the crossover individuals as follows: Among them, u j,i,g represents the jth gene of the i-th individual in the g-th generation crossover population, v j,i,g represents the jth gene of the i-th individual in the g-th generation mutant population, x j,i,g represents the jth gene of the i-th individual in the g-th generation population. Cr is the crossover operator, which is usually a floating-point data type randomly set between 0 and 1. Step 2.4: Select an action Update the next generation population using a greedy selection operation: Among them, x i,g+1 is the next generation individual, u i,g is the individual after the crossover operation, x i,g It is the parent individual before mutation and crossover operations, and f represents the fitness function.
4. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: The fault diagnosis model in step 3 includes an input layer, a convolutional layer, a Dropout layer, a pooling layer, an attention layer, a fully connected layer and an output layer connected in sequence; and is used to extract enhanced fault feature information.
5. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: In step 4, the DE-GWO algorithm is used to optimize the number of convolution kernels, convolution step size, pooling layer step size, drop rate, learning rate and batch size in the one-dimensional convolutional neural network.
6. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: In step 5, the training set data is input into the one-dimensional convolutional neural network fault diagnosis model after parameter optimization for training to learn the characteristics of different fault types of the equipment.
7. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: The training is terminated when the set number of times is reached or the loss function converges.
8. The fault diagnosis method based on parameter optimization and deep learning according to claim 1, characterized in that: In step 6, the test set is input into the trained one-dimensional convolutional neural network fault diagnosis model to test the fault diagnosis accuracy of the model.