CN117390371A

CN117390371A - Bearing fault diagnosis method, device and equipment based on convolutional neural network

Info

Publication number: CN117390371A
Application number: CN202311346992.9A
Authority: CN
Inventors: 郭江; 袁方; 高纪远; 伊同强; 师勇杰; 柯轶铭; 汪沛
Original assignee: Wuhan University WHU
Current assignee: Wuhan University WHU
Priority date: 2023-10-18
Filing date: 2023-10-18
Publication date: 2024-01-12

Abstract

The invention relates to the technical field of bearing fault diagnosis and discloses a bearing fault diagnosis method, device and equipment based on a convolutional neural network. The method comprises the steps of collecting bearing vibration signals, preprocessing, constructing a sample data set, and dividing the sample data set into a training set and a testing set; entering a model pre-training part, and adopting an Ebola algorithm to perform structural optimization on a residual convolution neural network with an attention mechanism to obtain an optimal structure of the network; after the optimal structure is obtained, performing formal training and evaluation on the residual convolution neural network with the attention mechanism for formal training according to the training set and the testing set to obtain a target bearing fault diagnosis model; and inputting the real-time bearing vibration signal into a target bearing fault diagnosis model to obtain a bearing fault detection result. In the invention, an Ebola algorithm is adopted in the pre-training stage, so that the accuracy of model processing of classification tasks with high complexity is improved. In complex signals, the improved residual connection and multi-headed attention mechanism can effectively capture critical information.

Description

Bearing fault diagnosis method, device and equipment based on convolutional neural network

Technical Field

The invention relates to the technical field of bearing fault diagnosis, in particular to a bearing fault diagnosis method, device and equipment based on a convolutional neural network.

Background

In modern industrial production, rotary machines are widely used in various fields such as electric power, traffic, manufacturing, etc. While bearings serve as critical components in rotary machines, bearing the important functions of supporting and transferring loads. However, long-term high-speed operation and severe operating environments may cause failure problems such as wear, abrasion, cracking, etc. of the bearing. Once the bearing fails, the normal operation of the equipment is affected, production interruption and accidents can be caused, and the production safety and economy are seriously affected. Therefore, it is important to perform timely and accurate fault diagnosis on the bearing state.

The traditional bearing fault diagnosis method mainly relies on experience judgment and normalized maintenance detection, and utilizes vibration analysis, frequency domain analysis and other modes to identify bearing faults. Although these methods are still widely used, such conventional failure diagnosis modes rely on a priori knowledge and have certain limitations in accuracy and real-time.

With the development of artificial intelligence and machine learning techniques, more and more researchers have begun to explore the application of these advanced techniques in bearing failure diagnosis. In this context, one-dimensional convolution algorithm, an important technique for deep learning, has shown great potential in bearing fault diagnosis. The one-dimensional convolution algorithm can effectively extract the characteristics of the time sequence signals and capture the local modes and rules in the signals. By performing one-dimensional convolution processing on fault signals such as bearing vibration signals, the fault signals can be converted into more representative characteristic representations, so that the bearing state can be accurately judged.

For example, rolling bearing fault diagnosis is performed based on a one-dimensional convolutional neural network, and feature extraction and fault diagnosis are performed on bearing signals by using a one-dimensional convolutional algorithm. However, conventional one-dimensional convolutional neural networks can only capture local, shallow features, and ignore deeper, more discriminating features. Furthermore, due to the complexity of rolling bearing data, one-dimensional convolutional neural networks (1D-CNN) are prone to overfitting on small data sets, while in deep networks gradients tend to vanish or explode during back propagation, resulting in model training difficulties. The approach to attention mechanisms, while emphasizing important features, is more complex, increasing storage requirements and computational complexity. Traditional one-dimensional convolutional neural networks or methods for fusing attention mechanisms have certain limitations in processing complex and variable fault signals.

The foregoing is provided merely for the purpose of facilitating understanding of the technical solutions of the present invention and is not intended to represent an admission that the foregoing is prior art.

Disclosure of Invention

The invention mainly aims to provide a bearing fault diagnosis method, device and equipment based on a convolutional neural network, and aims to solve the technical problems that the accuracy and instantaneity of the traditional bearing fault diagnosis mode and the deep learning method are limited to a certain extent when complex and changeable fault signals are processed.

In order to achieve the above object, the present invention provides a bearing fault diagnosis method based on convolutional neural network, comprising:

collecting bearing vibration signals of a bearing, preprocessing the bearing vibration signals, constructing a sample data set according to the preprocessed bearing vibration signals, and dividing the sample data set into a bearing vibration signal training set and a bearing vibration signal testing set;

entering a model pre-training part, and adopting an Ebola algorithm to perform structural optimization on a residual convolution neural network with an attention mechanism so as to obtain an optimal structure of the network;

after the optimal structure is obtained, performing formal training and evaluation on the residual convolution neural network with the attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal testing set so as to obtain a target bearing fault diagnosis model;

And inputting the real-time bearing vibration signal into the target bearing fault diagnosis model to obtain a bearing fault detection result.

In some embodiments, the entering the model pre-training part performs structural optimization on the residual convolution neural network with the attention mechanism by adopting an ebola algorithm to obtain an optimal structure of the network, and the method comprises the following steps:

constructing a one-dimensional convolutional neural network;

adding a residual structure, a multi-head self-attention module and a fully-connected classification layer to the one-dimensional convolutional neural network to obtain a residual convolutional neural network with an attention mechanism;

performing structural optimization on the residual convolution neural network with the attention mechanism based on an Ebola algorithm and the bearing vibration signal training set so as to obtain an optimal structure of the network; the adaptability of the Ebola algorithm is evaluated according to the performance of the network on the bearing vibration signal test set, and the optimal structure of the model is obtained after parameter searching is completed.

In some embodiments, the residual convolutional neural network with attention mechanism comprises 22 layers; wherein,

the first layer to the fourth layer are network structures of the one-dimensional convolutional neural network, and the one-dimensional convolutional neural network comprises a convolutional layer, a batch normalization layer, an activation function layer and a pooling layer;

The fifth layer to the twelfth layer comprise two convolution layers, two batch normalization layers, two activation function layers and two pooling layers;

the tenth layer to the nineteenth layer are the residual structures, and the residual structures comprise two convolution layers, two batch normalization layers, two activation functions and one Dropout layer;

the twentieth layer is the multi-head self-attention module;

a twenty-first global average pooling layer;

and the twenty-second layer is the full-connection classification layer.

In some embodiments, the structural optimization of the residual convolutional neural network with attention mechanism based on ebola algorithm and the bearing vibration signal training set to obtain an optimal structure of the network comprises:

according to an Ebola algorithm and the bearing vibration signal training set, carrying out initial parameter adjustment on the residual convolution neural network with the attention mechanism; wherein the initial parameters include a learning rate, a number of convolution kernels, a size, and a dropout rate;

the optimal parameters are searched and selected by iterative and adaptive processes to obtain an optimal structure of the network.

In some embodiments, the initial parameter adjustment of the residual convolutional neural network with attention mechanism according to ebola algorithm and the bearing vibration signal training set comprises:

Generating an initialization population according to an Ebola algorithm;

setting parameters of each layer of the residual convolution neural network with the attention mechanism to be trained according to the initialized population; the parameters of each layer comprise the size of convolution kernels, the number of convolution kernels, the Dropout ratio and the learning rate.

In some embodiments, the searching and selecting the optimal parameters through the iterative and adaptive process to obtain the optimal structure of the network includes:

inputting the bearing vibration signal training set into a residual convolution neural network with an attention mechanism to be trained for training so as to complete one iteration;

when the iteration times of the residual convolutional neural network with the attention mechanism to be trained meet the preset iteration times, obtaining the residual convolutional neural network with the attention mechanism after pre-training;

and determining optimal parameters according to the parameters of each layer of the pre-trained residual convolution neural network with the attention mechanism so as to obtain an optimal structure of the network.

In some embodiments, the fitness of the ebola algorithm is evaluated according to the performance of the network on the bearing vibration signal test set, and the obtaining of the optimal structure of the model after the parameter search is completed includes:

Evaluating a residual convolution neural network with an attention mechanism according to the bearing vibration signal test set to obtain an evaluation result;

calculating the complement of the test accuracy according to the evaluation result, and taking the complement of the test accuracy as fitness;

performing position movement by adopting an Ebola algorithm;

the residual convolutional neural network with the attention mechanism is retrained and evaluated using the new parameter set until the optimal structure of the model is obtained.

In some embodiments, after the optimal structure is obtained, performing formal training and evaluation on a residual convolution neural network with an attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal testing set to obtain a target bearing fault diagnosis model, where the method includes:

after the optimal structure is obtained, inputting the bearing vibration signal training set into a residual convolution neural network with an attention mechanism for formal training to train so as to complete one iteration;

obtaining a target bearing fault diagnosis model when the iteration times of the residual convolution neural network with the attention mechanism for formal training meet the preset iteration times;

and performing performance evaluation on the target bearing fault diagnosis model according to the bearing vibration signal test set.

In addition, in order to achieve the above object, the present invention also provides a bearing fault diagnosis device based on a convolutional neural network, including:

the signal preprocessing module is used for collecting bearing vibration signals of the bearing, preprocessing the bearing vibration signals, constructing a sample data set according to the preprocessed bearing vibration signals, and dividing the sample data set into a bearing vibration signal training set and a bearing vibration signal testing set;

the pre-training module is used for entering a model pre-training part, and adopting an Ebola algorithm to perform structural optimization on the residual convolution neural network with the attention mechanism so as to obtain an optimal structure of the network;

the formal training module is used for performing formal training and evaluation on the residual convolution neural network with the attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal testing set after the optimal structure is obtained so as to obtain a target bearing fault diagnosis model;

and the fault detection module is used for inputting a real-time bearing vibration signal into the target bearing fault diagnosis model so as to obtain a bearing fault detection result.

In addition, in order to achieve the above object, the present invention also proposes a bearing fault diagnosis apparatus based on a convolutional neural network, the bearing fault diagnosis apparatus based on a convolutional neural network comprising: the device comprises a memory, a processor and a convolutional neural network-based bearing fault diagnosis program stored on the memory and capable of running on the processor, wherein the convolutional neural network-based bearing fault diagnosis program is configured to realize the convolutional neural network-based bearing fault diagnosis method.

The method comprises the steps of collecting bearing vibration signals of a bearing, and preprocessing the bearing vibration signals; constructing a sample data set according to the preprocessed bearing vibration signals, and dividing the sample data set into a bearing vibration signal training set and a bearing vibration signal testing set; pretraining the residual convolutional neural network with the attention mechanism based on the Ebola algorithm and the bearing vibration signal training set to obtain the pretrained residual convolutional neural network with the attention mechanism; evaluating the adaptability of the pre-trained residual convolution neural network with the attention mechanism according to the bearing vibration signal test set, and determining the optimal model parameters of the residual convolution neural network with the attention mechanism according to the evaluation result of the adaptability; constructing a residual convolution neural network with an attention mechanism according to the optimal model parameters; training and evaluating a residual convolution neural network with an attention mechanism according to the bearing vibration signal training set and the bearing vibration signal testing set so as to obtain a target bearing fault diagnosis model; and inputting the real-time bearing vibration signal into a target bearing fault diagnosis model to obtain a bearing fault detection result. According to the invention, the parameter optimization is carried out on the improved one-dimensional convolutional network, namely the residual convolutional neural network with the attention mechanism, based on the Ebola algorithm in the pre-training stage, so that the key parameters of the residual convolutional neural network with the attention mechanism can be effectively optimized, the parameter tuning efficiency of the model is greatly improved, the performance of the model is further enhanced, and the model shows higher accuracy in processing classification tasks with high complexity. In addition, the improved residual connection and multi-head attention mechanism ensure that the target bearing fault diagnosis model can effectively capture key information even in very complex and noisy bearing vibration signals, and solve the technical problems that the accuracy and instantaneity of the existing traditional bearing fault diagnosis mode and the deep learning method are limited when complex and changeable fault signals are processed.

Drawings

FIG. 1 is a schematic structural diagram of a convolutional neural network-based bearing fault diagnosis device for a hardware operating environment according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart of an embodiment of a method for diagnosing bearing faults based on convolutional neural network;

FIG. 3 is a flow chart of fault diagnosis according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a confusion matrix according to an embodiment of the present invention;

fig. 5 is a block diagram of a bearing fault diagnosis apparatus based on convolutional neural network according to an embodiment of the present invention.

The achievement of the objects, functional features and advantages of the present invention will be further described with reference to the accompanying drawings, in conjunction with the embodiments.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It should be noted that all directional indicators (such as up, down, left, right, front, and rear … …) in the embodiments of the present invention are merely used to explain the relative positional relationship, movement, etc. between the components in a particular posture (as shown in the drawings), and if the particular posture is changed, the directional indicator is changed accordingly.

Furthermore, the description of "first," "second," etc. in this disclosure is for descriptive purposes only and is not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In addition, the technical solutions of the embodiments may be combined with each other, but it is necessary to base that the technical solutions can be realized by those skilled in the art, and when the technical solutions are contradictory or cannot be realized, the combination of the technical solutions should be considered to be absent and not within the scope of protection claimed in the present invention. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

Referring to fig. 1, fig. 1 is a schematic structural diagram of a bearing fault diagnosis device based on a convolutional neural network in a hardware operation environment according to an embodiment of the present invention.

As shown in fig. 1, the convolutional neural network-based bearing fault diagnosis apparatus may include: a processor 1001, such as a central processing unit (Central Processing Unit, CPU), a communication bus 1002, a user interface 1003, a network interface 1004, a memory 1005. Wherein the communication bus 1002 is used to enable connected communication between these components. The user interface 1003 may include a Display, an input unit such as a Keyboard (Keyboard), and the optional user interface 1003 may further include a standard wired interface, a wireless interface. The network interface 1004 may optionally include a standard wired interface, a Wireless interface (e.g., a Wireless-Fidelity (Wi-Fi) interface). The Memory 1005 may be a high-speed random access Memory (Random Access Memory, RAM Memory) or a stable nonvolatile Memory (NVM), such as a disk Memory. The memory 1005 may also optionally be a storage device separate from the processor 1001 described above.

Those skilled in the art will appreciate that the structure shown in fig. 1 does not constitute a limitation of the convolutional neural network-based bearing failure diagnosis apparatus, and may include more or fewer components than shown, or certain components may be combined, or a different arrangement of components.

As shown in fig. 1, an operating system, a network communication module, a user interface module, and a bearing failure diagnosis program based on a convolutional neural network may be included in a memory 1005 as one storage medium.

In the convolutional neural network-based bearing fault diagnosis apparatus shown in fig. 1, the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the convolutional neural network-based bearing fault diagnosis device of the present invention may be disposed in the convolutional neural network-based bearing fault diagnosis device, and the convolutional neural network-based bearing fault diagnosis device invokes the convolutional neural network-based bearing fault diagnosis program stored in the memory 1005 through the processor 1001 and executes the convolutional neural network-based bearing fault diagnosis method provided by the embodiment of the present invention.

The embodiment of the invention provides a bearing fault diagnosis method based on a convolutional neural network, and referring to fig. 2, fig. 2 is a flow chart of an embodiment of the bearing fault diagnosis method based on the convolutional neural network.

As shown in fig. 2, the bearing fault diagnosis method based on the convolutional neural network includes:

step S100: collecting bearing vibration signals of a bearing, preprocessing the bearing vibration signals, constructing a sample data set according to the preprocessed bearing vibration signals, and dividing the sample data set into a bearing vibration signal training set and a bearing vibration signal testing set;

step S200: entering a model pre-training part, and adopting an Ebola algorithm to perform structural optimization on a residual convolution neural network with an attention mechanism so as to obtain an optimal structure of the network;

step S300: after the optimal structure is obtained, performing formal training and evaluation on the residual convolution neural network with the attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal testing set so as to obtain a target bearing fault diagnosis model;

step S400: and inputting the real-time bearing vibration signal into the target bearing fault diagnosis model to obtain a bearing fault detection result.

It should be noted that, the execution body in this embodiment may be a bearing fault diagnosis device based on a convolutional neural network, where the bearing fault diagnosis device based on a convolutional neural network may be a computer device with a data processing function, or may be other devices that may implement the same or similar functions, which is not limited in this embodiment, and in this embodiment, a computer device is described as an example.

Illustratively, collecting a bearing vibration signal of a bearing, and preprocessing the bearing vibration signal; and constructing a sample data set according to the preprocessed bearing vibration signals, and dividing the sample data set into a bearing vibration signal training set and a bearing vibration signal testing set. 7 may be employed: the ratio of 3 divides the sample dataset into a bearing vibration signal training set and a bearing vibration signal testing set. Model training was performed using 70% of the data, parameters of the model were adjusted, and 30% of the data was used as a test set for evaluating the failure diagnosis effect of the model.

In some embodiments, the entering the model pre-training part performs structural optimization on the residual convolution neural network with the attention mechanism by adopting an ebola algorithm to obtain an optimal structure of the network, and the method comprises the following steps: constructing a one-dimensional convolutional neural network; adding a residual structure, a multi-head self-attention module and a fully-connected classification layer to the one-dimensional convolutional neural network to obtain a residual convolutional neural network with an attention mechanism; performing structural optimization on the residual convolution neural network with the attention mechanism based on an Ebola algorithm and the bearing vibration signal training set so as to obtain an optimal structure of the network; the adaptability of the Ebola algorithm is evaluated according to the performance of the network on the bearing vibration signal test set, and the optimal structure of the model is obtained after parameter searching is completed.

It should be noted that, this embodiment designs a target bearing fault diagnosis model with a attention mechanism and a residual convolution network. Referring to fig. 3, the network architecture is first designed to enable the model to focus more on critical fault features by introducing residual connections, while fusing attention mechanisms to enhance feature extraction capabilities. After the network structure design is completed, parameters such as the number, the size, the learning rate and the like of convolution kernels in the network structure are optimized by adopting an Ebola algorithm, so that the pre-training of a network model is realized. After pre-training, model training is performed based on the obtained network structure, so that the model can be ensured to be excellent in various bearing fault scenes.

In some embodiments, the residual convolutional neural network with attention mechanism comprises 22 layers; the first layer to the fourth layer are network structures of the one-dimensional convolutional neural network, and the one-dimensional convolutional neural network comprises a convolutional layer, a batch normalization layer, an activation function layer and a pooling layer; the fifth layer to the twelfth layer comprise two convolution layers, two batch normalization layers, two activation function layers and two pooling layers; the tenth layer to the nineteenth layer are the residual structures, and the residual structures comprise two convolution layers, two batch normalization layers, two activation functions and one Dropout layer; the twentieth layer is the multi-head self-attention module; a twenty-first global average pooling layer; and the twenty-second layer is the full-connection classification layer.

Specifically, the network architecture design:

CBRP layer (convolution-batch normalization-RELU-pooling layer, total 4 layers): the first four layers of the model consist of four convolution-batch normalization-pooling modules. Each module receives the output of the previous layer as input and performs convolution, batch normalization and maximum pooling operations. As the depth of the network increases, the number of output channels of the convolutional layer increases gradually from 64 to 512. The operation of the module can extract and strengthen the input characteristic information.

Wherein the convolution layer performs depth feature extraction, wherein the convolution process can be expressed as in equation (1), in equation (1):representing the weight of the i-th filter in layer i; />Representing the deviation of the i-th filter in layer i; x is x ^l (j) Represents the j-th local area in layer l; />Representing the input of the j-th neuron in the l+1 layer frame i.

Batch normalization: the method aims at improving the stability and speed of network training. In each training batch, the output of each layer is normalized to have a mean value of 0 and a standard deviation of 1. This normalization operation can effectively alleviate the internal covariate offset problem. By Batch Normalization (BN), each layer can be made to operate with a relatively stable data distribution, allowing a higher learning rate to be used, accelerating the training process, and introducing a Batch Normalization (BN) into the network can achieve better performance and faster convergence speed.

Pooling layer: downsampling strategies are utilized to reduce the dimensionality of features and network parameters. The maximum pooling can effectively extract the most important characteristic information, and can be expressed as equation (2). Wherein,represents the ith layer of the first layerThe value of the t-th neuron of the frame. t is in the range [ (j-1) W+1, j W)]Where W is the width of the pooling area. P (P) _i ^l+1 (j) The value of the jth neuron of the ith frame in the 1+1st layer after the max pooling operation is represented.

Residual structure: after the convolution-batch normalization-RELU-pooling layer, a residual structure is included in the network, and the residual block consists of two convolution layers, two Batch Normalization (BN) layers, two RELU activation functions, and one Dropout layer. In forward propagation, the input x of each residual block is processed not only by the convolution, BN, reLU and Dropout layers, but also directly added to the outputs of these layers. The method is an implementation mode of residual connection, information of input x can be directly flowed to output, the expression capacity of a model is improved, and meanwhile, the method is also beneficial to counter-propagation of gradients. In this embodiment, by adding a Dropout layer in the residual connection, a part of hidden layer neurons in the network are randomly discarded while new neurons are constructed, and the total number of neurons is ensured to be unchanged, and the innovative network calculation formula can be expressed as equation (3).

Attention mechanism: after the residual block, the model introduces a multi-headed self-attention module, which focuses and adjusts the features themselves. This module can help the model better understand the dependencies between inputs and further extract and augment the feature information.

Full connection classification layer: the last part of the model is a full connection layer which is used for mapping the advanced features extracted by the multi-head self-attention module to the target category to complete the classification task. The model has previously used a global averaging pooling layer to convert the output of the multi-headed self-attention module into one-dimensional features for input into the fully connected layer.

Specifically, the specific structure of the network structure designed in this embodiment is shown in table 1:

TABLE 1 model network architecture

In some embodiments, structural optimization of the residual convolutional neural network with attention mechanism based on ebola algorithm and the bearing vibration signal training set to obtain an optimal structure of the network comprises: according to an Ebola algorithm and the bearing vibration signal training set, carrying out initial parameter adjustment on the residual convolution neural network with the attention mechanism; wherein the initial parameters include a learning rate, a number of convolution kernels, a size, and a dropout rate; the optimal parameters are searched and selected by iterative and adaptive processes to obtain an optimal structure of the network.

In some embodiments, performing initial parameter adjustment on the residual convolutional neural network with attention mechanism according to an ebola algorithm and the bearing vibration signal training set, comprising: generating an initialization population according to an Ebola algorithm; setting parameters of each layer of the residual convolution neural network with the attention mechanism to be trained according to the initialized population; the parameters of each layer comprise the size of convolution kernels, the number of convolution kernels, the Dropout ratio and the learning rate.

In some embodiments, searching and selecting optimal parameters through an iterative and adaptive process to obtain an optimal structure of the network includes: inputting the bearing vibration signal training set into a residual convolution neural network with an attention mechanism to be trained for training so as to complete one iteration; when the iteration times of the residual convolutional neural network with the attention mechanism to be trained meet the preset iteration times, obtaining the residual convolutional neural network with the attention mechanism after pre-training; and determining optimal parameters according to the parameters of each layer of the pre-trained residual convolution neural network with the attention mechanism so as to obtain an optimal structure of the network.

Specifically, parameter optimization: the model adopts an Ebola algorithm, namely an Ebola parameter optimization algorithm, to carry out initial parameter adjustment in a pre-training stage. The Ebola algorithm is inspired by an Ebola virus propagation mode, and the optimal parameters are searched and selected through iteration and adaptation processes to search the learning rate, the convolution kernel size and the number, so that the model can be quickly converged, and the performance of the final model can be remarkably improved.

The specific implementation of the ebola parameter search algorithm is as follows:

after setting the super parameters, initializing the population. Initialization of the population is performed according to equation (4). Wherein U is _i And L _i Representing the upper and lower bounds of the search parameter, rand (0, 1) represents a random number of 0 to 1. In this embodiment, when learning rate and dropout optimization is performed, U _i And L _i Set to 1 and 0, respectively. When the number of convolution kernels is performed, U _i And L _i Set to 64 and 1, respectively, the second layer is 128 and 64, and the third layer is 256 and 128. U when convolving kernel sizes are performed _i And L _i Set to 10 and 1, respectively.

individual _i ＝L _i +rand(0,1)*(U _i +L _i ) (4)

According to formula (5), the current optimal solution bestS, gBest and cBest in the infected individual set are selected at time t to represent the optimal solutions at time t, namely the global optimal solution and the current optimal solution, respectively, while fitness is the objective function of the problem. Superpropagators and ordinary propagators of ebola viruses are assigned to gBest and cBest, respectively.

The location update for each exposed individual is determined by equation (6), where ρ represents the scale factor of the individual's displacement,and->The updated and original positions at time t+1 and time t are shown, respectively. M (I) is the individual's rate of movement, determined by equations (7) and (8), applied to two different phases of the algorithm search, the development phase and the exploration phase, respectively. The development phase assumes that the infected person moves within a very close distance (srate), while the exploration phase is that the infected person moves beyond the average proximity distance (lrate). The farther an infected person moves, the more individuals may be exposed to the infection. Both cases are controlled by one neighborhood parameter: when the parameter is 0.5 or more, infection may cause large-scale spread; if less than 0.5, the infection is controlled to be in a local range.

M(I)＝srate*rand(0，1)+M(Ind _best ) (7)

M(s)＝lrate*rand(0，1)+M(Ind _best ) (8)

Susceptibility (S), infectious agent (I), inpatient (H), exposer (E), inoculator (V), convalescent (R), funeral (F), isolator (Q) and deceased (D) controlled by ordinary differential equations are updated according to the study method in the medical field. The application of the differentiation is to obtain the rate of change of the quantities S, I, H, R, V, D and Q with time t. The related formulas are formulas (9) to (15). The ebola algorithm assumes that equations (9-15) are scalar functions, each with a floating point number. The rate of change of the susceptible population is first determined and then the number of susceptible populations at time t is calculated from this rate. The same method is used to calculate the individual sets of vectors I, H, R, V, D and Q. The initial conditions are set as S (0) =s0, I (0) =i0, R (0) =r0, D (0) =d0, P (0) =p0, and Q (0) =q0, time t is calculated from after epoch, and δ in equation (14) is the ratio of the funeral. Equation (15) then mimics the isolation rate of ebola infection cases.

The specific search mode is as follows:

first, an initial population is created using equation (4). An exponential case (icase) is then generated as global best (gbest) and current best (cbest). When the number of infectors (I) is greater than 0 in a given number of iterations (epoch), the algorithm continues. A portion of the infected persons is randomly selected to enter quarantine (Q), and the remaining infected persons (fracI) will continue to infect susceptible persons. For each infected person who is still susceptible to infection, a movement rate (posi) is calculated. This rate of movement affects whether it exceeds its surroundings. Based on this rate of movement, each infected person has a certain probability of infecting surrounding susceptible persons, generating a new infected person (newI). Some infected persons enter the hospital (H), some will recover (R), some will vaccinate (V), and some will die (D). Recovered infected individuals may rejoin the susceptible population, while dead infected individuals may be replaced with new susceptible individuals. The algorithm will calculate the optimal solution after each iteration and compare if it is better than the previous globally optimal solution. If better, the globally optimal solution is updated. The algorithm continues until a preset number of iterations or number of infectors is reached of 0. Finally, the algorithm returns the best solution and its corresponding solution.

It can be understood that the embodiment adopts a deep learning model structure optimizing method based on the ebola optimizing search algorithm, and is specially designed for processing bearing vibration signals. Bearing vibration signals are used as key indexes of mechanical health and often carry rich information, so that the health condition and potential failure modes of the bearing can be reflected. In order to extract valuable information from these signals more efficiently, a suitable model structure and corresponding parameter settings are crucial. The core idea of this embodiment is to utilize the ebola optimization search algorithm to perform parameter optimization on the improved one-dimensional convolutional network in the pre-training stage. The Ebola optimization search algorithm is inspired by the propagation mode of the Ebola virus, and the behavior mode of the Ebola virus is mapped to the parameter search process, so that the parameter optimization algorithm with strong global search capability is designed. The Ebola optimization search algorithm can keep good diversity while searching for the optimal solution, so that excessive localization is avoided. The key parameters of the deep learning model in the embodiment, including the learning rate, the number of convolution kernels, the size and the dropout rate, can be effectively optimized by adopting the Ebola optimization search algorithm. Through multiple iterations and searches, an optimized parameter combination is obtained, and the combination provides a one-dimensional convolution network with optimal performance for processing the bearing vibration signals. The introduction of the optimization strategy in the embodiment greatly improves the parameter tuning efficiency of the model, further strengthens the performance of the model, and enables the model to show higher accuracy when processing classification tasks with high complexity.

It should be noted that, in the embodiment, when the network structure is designed, a one-dimensional convolution algorithm, that is, a one-dimensional convolution neural network, is adopted, and the one-dimensional convolution algorithm can effectively extract features of the bearing vibration signal, directly process the signal in the time domain, avoid loss of signal information, and capture local modes and rules in the signal. Through a one-dimensional convolution algorithm, the bearing vibration signal is represented by a more representative characteristic, and more reliable input is provided for subsequent fault diagnosis. The present embodiment adds a residual structure and a multi-headed self-attention module, i.e. the attention mechanism and residual connection are fused into a one-dimensional convolution. The attention mechanism enables the model to focus more on important features related to bearing failure. Through the combination of one-dimensional convolution and an attention mechanism, the capacity is further improved on the characteristic representation, weak characteristics of bearing faults can be more accurately captured, and the accuracy and the robustness of fault diagnosis are enhanced. Residual connection can ensure that gradients are smoothly counter-propagated in the deep network, thereby speeding training and improving model performance. In addition, the embodiment introduces an ebola algorithm, namely an ebola parameter optimization algorithm, in the pre-training stage. In the pre-training stage, the Ebola parameter optimization algorithm can efficiently adjust model parameters, and further improves the accuracy and performance of fault diagnosis. Through the Ebola parameter optimization algorithm, the embodiment can fully optimize the model, adapt to different types of faults, improve the generalization capability of the model, and reduce the training cost.

It can be appreciated that the deep learning model structure for the bearing vibration signal: the vibration signal of the bearing is characterized by comprising important information about the health condition of the bearing in the time domain, the frequency domain and the time-frequency domain. Based on this feature, the present embodiment performs the following key processes and optimizations on the model:

first, improved residual connection processing time-frequency domain features: the bearing vibration signal appears as different pulses and oscillations in the time domain. The introduction of deep network structures helps to capture these complex features. But as the depth of the network increases, the problems of gradient extinction and explosion become a serious problem. By using an improved residual connection, the network is able to propagate gradients more efficiently in the depth structure, ensuring that deep features are effectively learned. The method enables the model to better capture subtle time domain differences of the bearing under different health conditions, thereby improving the accuracy of fault detection and classification.

Second, the multi-head attention mechanism processes the time-frequency domain features: bearing failure typically results in a change in the frequency content of the signal, producing a particular frequency signature. These features may appear as specific patterns or structures in the time-frequency domain. The multi-headed attention mechanism allows the model to focus on and learn these features from different angles or "frequency bins". Each "header" may concern a particular portion of the input signal, such as a particular frequency range or time-frequency structure, and then integrate that information. This mechanism ensures that the model can capture various complex features in the bearing vibration signal, from low frequency oscillations to high frequency pulses.

It should be noted that, compared with the traditional feature extraction and machine learning method for the bearing signals, the fault diagnosis model provided by the embodiment directly learns on the original signals, so that complexity of feature engineering and possible information loss are avoided. The improved residual connection and multi-headed attention mechanism ensures that the model is able to effectively capture critical information even in very complex and noisy bearing vibration signals.

In some embodiments, the fitness of the ebola algorithm is evaluated according to the performance of the network on the bearing vibration signal test set, and the optimal structure of the model is obtained after the parameter search is completed, including: evaluating a residual convolution neural network with an attention mechanism according to the bearing vibration signal test set to obtain an evaluation result; calculating the complement of the test accuracy according to the evaluation result, and taking the complement of the test accuracy as fitness; performing position movement by adopting an Ebola algorithm; the residual convolutional neural network with the attention mechanism is retrained and evaluated using the new parameter set until the optimal structure of the model is obtained.

Note that, the pre-training section:

first, an initialization population is generated: first a set of parameters is randomly generated, including the convolution kernel size, the number of convolution kernels, the Dropout ratio, the learning rate.

Second, the fitness is evaluated: firstly, determining specific parameters (convolution kernel size, number of convolution kernels, ratio of Dropout, learning rate) of each layer by using the initialized parameters; secondly, inputting a bearing vibration signal training set into a network for network training, wherein the specific network structure comprises: conv1D signals pass through a first layer of convolution layer, the layer is mainly used for extracting original input features, the BatchNorm1D performs batch normalization on the convolved features to accelerate model convergence and improve model stability, reLU introduces nonlinearity, so that the model can capture more complex features, maxPool1D uses maximum pooling to reduce feature dimensions, and meanwhile retains the most important information, and 5-12 layers: these layers continue to extract and enhance features, and through multi-layer convolution and pooling, the model can identify more complex and abstract features, layers 13-19: the residual block is added to increase the depth of the network, meanwhile, the Dropout layer can avoid model overfitting, and the multi-head attention is introduced into an attention mechanism, so that the model focuses on important features related to bearing faults, namely GlobalAvgPool1D: the dimension of the features is further reduced through global average pooling, important information is reserved, and the full connection layer: and classifying the final features through a full connection layer, and mapping the features to the categories. And (3) adjusting the network weight according to the comparison between the classified output and the actual label, and completing one round of optimization. This process continues until a predetermined number of iteration cycles is reached. Again, test set evaluation was performed: the test set is used to evaluate the performance of the current network structure. Finally, calculate the fitness: and calculating the fitness according to the evaluation result of the test set. The complement of the test accuracy is selected in this embodiment.

Third, position movement: and (3) according to the fitness result obtained in the last step, performing position movement by using an Ebola algorithm.

Fourth, repeat fitness evaluation: the network structure is again determined using the new parameter set and the model is retrained and evaluated to calculate a new fitness. The process is iterated until the termination condition is met, resulting in optimal model parameters.

In some embodiments, after obtaining the optimal structure, performing formal training and evaluation on the residual convolutional neural network with the attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal testing set to obtain a target bearing fault diagnosis model, where the method includes: after the optimal structure is obtained, inputting the bearing vibration signal training set into a residual convolution neural network with an attention mechanism for formal training to train so as to complete one iteration; obtaining a target bearing fault diagnosis model when the iteration times of the residual convolution neural network with the attention mechanism for formal training meet the preset iteration times; and performing performance evaluation on the target bearing fault diagnosis model according to the bearing vibration signal test set.

The formal training is performed after the pre-training. Formal training part:

First, the optimal parameters obtained by using the ebola algorithm, that is, the optimal model parameters obtained by pre-training, are used to determine specific parameters (such as convolution kernel size, number of convolution kernels, dropout ratio, learning rate) of each layer of the network structure.

Secondly, inputting bearing vibration signals of the bearing vibration signal training set into a network for network training:

conv1D: the signal first passes through a first convolution layer, which is primarily used to extract features of the original input. BatchNorm1D: batch normalization is performed on the convolved features to accelerate model convergence and improve model stability. ReLU: introducing nonlinearities enables the model to capture more complex features. MaxPool1D: maximum pooling is used to reduce the dimensionality of features while retaining the most important information. 5-12 layers: these layers continue to extract and enhance features. By multi-layer convolution and pooling, the model can identify more complex and abstract features. 13-19 layers: these are residual blocks that can increase the depth of the network while the Dropout layer can avoid model overfitting. Multiheadattribute: by introducing a focus mechanism, the model may be more focused on important features related to bearing failure. Globalapgpool 1D: the dimension of the features is further reduced by global averaging pooling and important information is preserved. Full tie layer: and classifying the final features through a full connection layer, and mapping the features to the categories. And (3) adjusting the network weight according to the comparison between the classified output and the actual label, and completing one round of optimization. This process continues until a predetermined number of iteration rounds is reached to obtain a trained residual convolutional neural network with an attention mechanism.

Third, test set evaluation: the performance of the final training model is evaluated using the test set to obtain a target bearing failure diagnosis model.

It should be noted that, in this embodiment, a one-dimensional convolution algorithm, an attention mechanism, an improved residual connection, and an ebola parameter optimization algorithm are fused, in order to further verify the performance of the target bearing fault diagnosis model, a bearing fault data set provided by Case Western Reserve University (CWRU) is selected in this embodiment, and the target bearing fault diagnosis model for bearing vibration signal fault diagnosis provided in this embodiment is verified.

In the CWRU dataset, a fault diagnosis test was performed using a 6205-2RS rolling bearing manufactured by SKF company. At a sampling frequency of 12kHz, the vibration signal of the bearing at 4 conditions (0-3 hp) was recorded. Each working condition examines single point faults of three diameters (0.178 mm, 0.356mm and 0.532 mm) of the ball, the inner race and the outer race, and together with data of a normal state, each working condition has 10 fault types, namely the fault classification and the label value of the CWRU bearing shown in table 2. The original vibration signal of the CWRU dataset was split into non-overlapping samples, one sample every 1024 data points. The number of divided test samples is shown in the CWRU sample number table shown in table 3.

Table 2 CWRU bearing failure classification and tag values

Tag value	Fault type	Severity of disease
			0	Normal state	0
1	Failure of inner race	0.007
			2	Ball failure	0.007
3	Failure of outer race	0.007
			4	Failure of inner race	0.014
5	Ball failure	0.014
			6	Failure of outer race	0.014
7	Failure of inner race	0.021
			8	Ball failure	0.021
9	Failure of outer race	0.021

Table 3 CWRU sample quantity table

Specifically, 7 was employed on the data division: the scale of 3 divides the dataset into a training set and a test set. Training a model by using 70% of data, and adjusting parameters of the model; the remaining 30% of the data was used as a test set for evaluating the fault diagnosis effect of the model. After model training is completed, referring to fig. 4, where fig. 4 is a drawn confusion matrix, it can be clearly seen that all predictions fall on the diagonal, meaning that all samples are correctly classified. The result verifies the effectiveness of the target bearing fault diagnosis model for bearing vibration signal fault diagnosis, and proves that the target bearing fault diagnosis model has good effects in the aspects of model structural design, pre-training parameter selection, data processing and the like.

It should be noted that, in the above-mentioned background art, bearing fault diagnosis can be performed to some extent, but in practical application, these schemes still have some obvious drawbacks and disadvantages. In the background art, the common one-dimensional convolution fault diagnosis method and the one-dimensional convolution fault diagnosis method integrating the attention mechanism have some defects. The common one-dimensional convolution method can not accurately capture weak features of bearing faults during feature extraction, parameter learning is insufficient, accuracy and robustness of fault diagnosis are affected, and meanwhile, interpretability is poor. The approach to attention mechanisms, while emphasizing important features, is more complex, increases storage requirements and computational complexity, and is more sensitive to noise. In contrast, the embodiment introduces a parameter optimization algorithm in the pre-training stage, so that the accuracy and the robustness of fault diagnosis are effectively improved. Therefore, compared with the prior art, the embodiment has more comprehensive consideration on the aspects of characteristic characterization and parameter optimization, can provide a more superior solution for bearing fault diagnosis, reduces production risk and improves production efficiency and equipment reliability.

In this embodiment, parameter optimization is performed on the improved one-dimensional convolutional network, that is, the residual convolutional neural network with the attention mechanism, based on the ebola algorithm in the pre-training stage, so that key parameters of the residual convolutional neural network with the attention mechanism in this embodiment can be effectively optimized, parameter tuning efficiency of the model is greatly improved, performance of the model is further enhanced, and the model shows higher accuracy when processing classification tasks with high complexity. In addition, the improved residual connection and multi-head attention mechanism ensure that the target bearing fault diagnosis model can effectively capture key information even in very complex and noisy bearing vibration signals, and solve the technical problems that the accuracy and instantaneity of the existing traditional bearing fault diagnosis mode and the deep learning method are limited when complex and changeable fault signals are processed. In summary, the embodiment fuses a one-dimensional convolution algorithm, an attention mechanism, improved residual connection and an ebola parameter optimization algorithm, and realizes a more efficient and accurate bearing fault diagnosis method by strengthening characteristic characterization capability and adjusting model parameters.

Referring to fig. 5, fig. 5 is a block diagram illustrating an embodiment of a bearing fault diagnosis apparatus based on a convolutional neural network according to the present invention.

As shown in fig. 5, the bearing fault diagnosis apparatus based on convolutional neural network includes:

the signal preprocessing module 10 is used for collecting bearing vibration signals of the bearing, preprocessing the bearing vibration signals, constructing a sample data set according to the preprocessed bearing vibration signals, and dividing the sample data set into a bearing vibration signal training set and a bearing vibration signal testing set;

the pre-training module 20 is configured to enter a model pre-training part, and perform structural optimization on the residual convolutional neural network with the attention mechanism by adopting an ebola algorithm to obtain an optimal structure of the network;

the formal training module 30 is configured to perform formal training and evaluation on the residual convolutional neural network with the attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal testing set after obtaining the optimal structure, so as to obtain a target bearing fault diagnosis model;

the fault detection module 40 is configured to input a real-time bearing vibration signal into the target bearing fault diagnosis model to obtain a bearing fault detection result.

In addition, technical details not described in detail in the embodiment of the bearing fault diagnosis device based on the convolutional neural network may be referred to the bearing fault diagnosis method based on the convolutional neural network provided in any embodiment of the present invention, which is not described herein.

It should be understood that the foregoing is illustrative only and is not limiting, and that in specific applications, those skilled in the art may set the invention as desired, and the invention is not limited thereto.

It should be noted that the above-described working procedure is merely illustrative, and does not limit the scope of the present invention, and in practical application, a person skilled in the art may select part or all of them according to actual needs to achieve the purpose of the embodiment, which is not limited herein.

Furthermore, it should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or system. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or system that comprises the element.

The foregoing embodiment numbers of the present invention are merely for the purpose of description, and do not represent the advantages or disadvantages of the embodiments.

From the above description of the embodiments, it will be clear to those skilled in the art that the above-described embodiment method may be implemented by means of software plus a necessary general hardware platform, but of course may also be implemented by means of hardware, but in many cases the former is a preferred embodiment. Based on such understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art in the form of a software product stored in a storage medium (e.g. Read Only Memory)/RAM, magnetic disk, optical disk) and including several instructions for causing a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to perform the method according to the embodiments of the present invention.

The foregoing description is only of the preferred embodiments of the present invention, and is not intended to limit the scope of the invention, but rather is intended to cover any equivalents of the structures or equivalent processes disclosed herein or in the alternative, which may be employed directly or indirectly in other related arts.

Claims

1. The bearing fault diagnosis method based on the convolutional neural network is characterized by comprising the following steps of:

2. The method for diagnosing bearing faults based on a convolutional neural network as claimed in claim 1, wherein the entering model pre-training section performs structural optimization on a residual convolutional neural network with an attention mechanism by using an ebola algorithm to obtain an optimal structure of the network, comprising:

Constructing a one-dimensional convolutional neural network;

3. The convolutional neural network-based bearing failure diagnosis method of claim 2, wherein the residual convolutional neural network with attention mechanism comprises 22 layers; wherein,

The twentieth layer is the multi-head self-attention module;

a twenty-first global average pooling layer;

and the twenty-second layer is the full-connection classification layer.

4. The convolutional neural network-based bearing fault diagnosis method of claim 2, wherein the ebola algorithm and the bearing vibration signal training set-based structural optimization of the attentional mechanism-equipped residual convolutional neural network to obtain an optimal structure of the network comprises:

5. The method for diagnosing a bearing failure based on a convolutional neural network of claim 4, wherein said performing initial parameter adjustment on said residual convolutional neural network with attention mechanism based on ebola algorithm and said training set of bearing vibration signals comprises:

generating an initialization population according to an Ebola algorithm;

6. The convolutional neural network-based bearing failure diagnosis method of claim 4, wherein the searching and selecting optimal parameters through iterative and adaptive processes to obtain an optimal structure of the network comprises:

7. The convolutional neural network-based bearing fault diagnosis method of claim 2, wherein the adaptation of the ebola algorithm is evaluated according to the performance of the network on the bearing vibration signal test set, and the obtaining of the optimal structure of the model after the completion of the parameter search comprises:

performing position movement by adopting an Ebola algorithm;

8. The method for diagnosing bearing faults based on the convolutional neural network as claimed in claim 1, wherein after the optimal structure is obtained, performing formal training and evaluation on the residual convolutional neural network with an attention mechanism for formal training according to the bearing vibration signal training set and the bearing vibration signal test set to obtain a target bearing fault diagnosis model, comprising:

9. A convolutional neural network-based bearing fault diagnosis device, comprising:

10. A convolutional neural network-based bearing fault diagnosis apparatus, comprising: a memory, a processor, and a convolutional neural network-based bearing fault diagnosis program stored on the memory and executable on the processor, the convolutional neural network-based bearing fault diagnosis program configured to implement the convolutional neural network-based bearing fault diagnosis method of any one of claims 1-8.