Deep and wide fault diagnosis method based on original vibration signal

By adopting deep separable convolutional networks and width learning systems in industrial environments, combined with residual stacking learning mechanisms, the problem of limited computing resources is solved, fast and accurate fault diagnosis and real-time response are achieved, and the diagnostic performance of edge devices is improved .

CN118606799BActive Publication Date: 2024-11-08JIANGSU HE MICROELECTRONICS TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411082944.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-08
Estimated Expiration
2044-08-08

AI Technical Summary

Technical Problem

The existing deep learning methods have limited computing resources in industrial environments, resulting in long and in real-time fault diagnosis time, which cannot meet industrial production needs.

Method used

The deep-width fault diagnosis method based on the original vibration signal is adopted, and the depth-separable convolutional network is constructed, combining the width learning system and the residual stacking learning mechanism to reduce the amount of network parameters and improve the calculation efficiency and response speed.

Benefits of technology

It realizes fast and accurate fault diagnosis on edge devices, enhances real-time response, and facilitates real-time network updates through incremental learning mechanisms, improving diagnostic performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118606799B_ABST
    Figure CN118606799B_ABST
Patent Text Reader

Abstract

The present invention discloses a deep-width fault diagnosis method based on original vibration signals, which is aimed at the situation where the computing power of edge gateway devices is insufficient or the real-time requirements of fault diagnosis are high. The method includes: obtaining original vibration signal data and label values. Preprocessing the original vibration signal, dividing the training set and the test set. Inputting the preprocessed training set into the first layer of deep separable convolutional network. Inputting the test set to obtain the predicted output and the residual output. Using the residual output as the training set input of the second layer of deep separable convolutional network to obtain the second predicted output. Adding the first layer predicted output to the second layer predicted output to obtain the predicted output. The deep-width fault diagnosis method based on original vibration signals of the present invention better realizes intelligent fault diagnosis under the conditions of insufficient computing power of edge gateway devices and high real-time requirements of fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of rotating machinery fault diagnosis, and in particular relates to a deep width fault diagnosis method based on original vibration signals. Background Art

[0002] In modern industry, the diagnosis of mechanical equipment is very important. Reliable fault diagnosis can significantly reduce economic losses and improve the safety of the operation process. As deep learning has achieved promising results in several fields that traditionally rely on professional knowledge, neural networks continue to show their potential in areas such as pattern recognition. Neural networks can extract fault features from complex mechanical vibration signals and distinguish the fault type based on the extracted fault features. This end-to-end approach can greatly improve the speed of fault diagnosis and is expected to be applied to real production. Therefore, many intelligent fault diagnosis methods combined with deep learning have been widely studied and attracted the attention of more and more scholars.

[0003] However, the current mainstream deep learning methods often require a large number of samples, a long training time, and high-configuration computing hardware. However, in industrial environments, the equipment is slow in executing complex algorithms and processing large amounts of data, resulting in long task execution times and affecting industrial production efficiency. The memory capacity of the equipment cannot meet the current task's processing requirements for large-scale data sets, which may lead to insufficient memory and make the diagnostic method impossible to implement in real production. Due to insufficient computing power, there is a delay in the time it takes for the equipment to receive instructions and respond, which does not meet real-time requirements. This contradicts the requirements of the current mainstream deep learning methods and is therefore difficult to implement in industry.

[0004] Therefore, in the context of limited computing resources, achieving rapid and accurate diagnosis of rolling bearing faults has far-reaching national strategic significance. By establishing an effective fault diagnosis system, equipment loss can be reduced and production efficiency can be improved, thereby promoting the country's industrial production and economic development. In an environment with limited resources, efficient use of computing resources is an important task. By designing a resource-saving fault diagnosis system, the utilization efficiency of computing resources can be improved, which is in line with the strategic goal of national sustainable development. Summary of the invention

[0005] In view of the deficiencies in the above technologies, the present invention proposes a deep-width fault diagnosis method based on the original vibration signal in an experimental scenario based on real production. This method constructs a deep-width fault diagnosis network model based on the original vibration signal and can be solved quickly to alleviate the problem of insufficient computing power and insufficient real-time response of edge devices.

[0006] To achieve the purpose of the present invention, the technical solution adopted by the present invention is: a deep width fault diagnosis method based on the original vibration signal, comprising the following steps:

[0007] The first step is to obtain the vibration signal data of the bearing and the corresponding fault type label value of each category of fault; pre-process the vibration signal and split the test set, the first-layer training set and the second-layer training set;

[0008] The second step is to build the first-layer depth-separable convolutional network, input the first-layer training set into the first-layer depth-separable convolutional network for training, save the parameters after the training is completed, and obtain the first-layer model;

[0009] The third step is to input the test set data into the trained first-layer model to obtain the first-layer prediction output and residual output; the residual output is used as the training set input of the second-layer deep separable convolutional network to train the second-layer deep separable convolutional network;

[0010] The fourth step is to add the prediction output of the first layer of the deep separable convolutional network and the prediction output of the second layer of the deep separable convolutional network to obtain the final prediction output, completing the construction of the deep and wide fault diagnosis network;

[0011] Step 5: Sample the new vibration data and input it into the deep and wide fault diagnosis network to output the fault type.

[0012] Furthermore, the construction and training process of the first layer of the depthwise separable convolutional network is as follows:

[0013] Input the first layer training set into the first layer of depth-separable convolutional network;

[0014] The input data first passes through the depthwise separable convolution blocks with convolution kernels of different scales to extract different features. The depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Each depthwise separable convolution block has several depthwise separable convolutions.

[0015] Every two depth-wise separable convolution blocks undergo a regularization operation; the BLS network is designed by Group feature nodes and The group enhancement node consists of the following structure:

[0016] First, depthwise separable convolution The generated The group feature nodes are:

[0017] ,

[0018] in is the input data, and is the weight within the feature node, The set of group feature nodes is represented as ;at the same time, A collection of group feature nodes Solved by pseudo-inverse, the objective function is:

[0019] ,

[0020] in Represents the label corresponding to the input data, take , the objective function optimal problem is set as regular ,value Represents the sum of square weights Constraints;

[0021] Second, The group enhancement node consists of the original convolution Generate, that is:

[0022] ,

[0023] in and is the weight in enhancement, and all enhanced nodes are represented as ;

[0024] Next, Group feature nodes and The set of group enhancement nodes is output through the original convolution dimension reduction , as the input of the second layer of depth-wise separable convolutional network;

[0025] Finally, Group feature nodes and The set of group enhancement nodes passes through the depth-separable convolution classification layer to obtain the desired output :

[0026] ,

[0027] in is the weight of the output layer of the first depthwise separable convolutional network, is the bias of the output layer of the first depthwise separable convolutional network;

[0028] After the first layer of depth-separable convolutional network model training is completed, the model parameters are saved for calling during the second layer of depth-separable convolutional network model training.

[0029] Furthermore, the training set construction method of the second layer of the depth-separable convolutional network is as follows:

[0030] The second layer training set is input into the trained first layer depth separable convolutional network, and the original label of the first layer depth separable convolutional network is Converted to one-hot encoding, and the actual output of the first layer of the depth-separable convolutional network Subtract to get the label , as the fitting target of the second layer of the depth-wise separable convolutional network;

[0031] The label , the dimension reduction output of the feature node and enhanced node set , and the expected output of the first layer of the depthwise separable convolutional network At the same time, the training set of the second-layer deep separable convolutional network is constructed.

[0032] Furthermore, the second-layer depth-separable convolutional network is composed of a stacked BLS network, and the construction and training process is as follows:

[0033] The input data of the second layer of depth-separable convolutional network is represented as , that is, the dimension reduction output ; Output data is the output of each BLS block ,have Stacked BLS blocks The structure is as follows:

[0034] First, formulate the block as follows:

[0035] ,

[0036] in , For the The output of each block, weight and Randomly generated, ,function is the identity function, that is ; and is the feature convolution, express and Composite mapping of ;

[0037] Second, the sum of the outputs of all stacked BLS blocks approximates the desired output ; Adjacent stacked BLS blocks are combined through residuals, where each stacked BLS block has the same structure, The expected output of each block is:

[0038] ;

[0039] For training data , the first stacked BLS block is constructed by setting and generate,

[0040] Right now , where the weight and Randomly generated, weighted and Through the matrix Calculate, that is:

[0041] ;

[0042] The first stacked BLS structure conversion output is fed into the next single stacked BLS block as training data, which is represented as , the expected output of the second stacked BLS block For input label With the first stacked BLS block output The difference, that is , the parameters of each stacked BLS block model are incrementally updated based on training each stacked BLS block; specifically, there are:

[0043] ;

[0044] The weight and Randomly generated; for Increment of stacked BLS blocks, ,have:

[0045] ,

[0046] The weight and Randomly generated, and:

[0047] .

[0048] The beneficial effect of the present invention is that by introducing a width learning system, a deep separable convolution and a residual stacking learning mechanism, the problem of insufficient computing power of edge devices is alleviated, the real-time response of edge devices is enhanced, and an incremental learning mechanism is provided to facilitate real-time updates of the diagnostic network. Experiments show that this method can improve the diagnostic performance of edge devices.

[0049] The deep width fault diagnosis method based on the original vibration signal and capable of rapid solution of the present invention constructs a deep width fault diagnosis network which is a deep separable convolutional neural network, enhanced nodes and a deep separable convolution classification layer. The input of the deep width fault diagnosis network is the original vibration signal. The training sample is trained by a deep separable convolution segmentation method to eliminate the shift variability problem of the original vibration signal.

[0050] The deep width fault diagnosis network reduces the number of network parameters by using deep separable convolution to replace the original convolution. The deep width fault diagnosis network replaces the fully connected layer of the original network by using deep separable convolution to form the convolution classification layer, which greatly reduces the number of network parameters. The deep width fault diagnosis network replaces the convolution kernel of the original convolution network by using wide convolution kernel to construct deep separable convolution, which enhances the feature extraction capability of one-dimensional vibration signals. The deep width fault diagnosis network can increase enhanced nodes horizontally, and the number of enhanced nodes can be incrementally modified according to the equipment situation to adapt to the different computing capabilities of different equipment. The deep width fault diagnosis network uses the concatenation of the feature nodes and enhanced nodes after dimensionality reduction as the input of the second layer, and the fitting target is the difference between the target label value and the predicted label value of the first layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0053] In the present invention, considering the problem of insufficient computing power and insufficient real-time response of edge devices, a deep and wide fault diagnosis network model based on the original vibration signal and capable of rapid solution is designed. First, considering the requirements of deep learning for the number of samples, training time and computing configuration, the present invention introduces the idea of ​​Board Learning System. The advantage of BLS is that it can quickly complete the training of the diagnostic model on the basis of ensuring certain nonlinear feature extraction capabilities and incremental capabilities. This is mainly because the structure used by BLS is relatively simple, without too many layers and nodes. In addition, BLS also has the ability of incremental learning, can quickly reconstruct the model and adapt to new data, avoiding the time and cost of retraining the entire model, and can make the model adapt to new data faster. Secondly, in order to better mine the potential information of features and improve the performance of the diagnostic network, the present invention adds a deep separable convolutional network with a simpler structure to BLS. On the basis of directly using the original signal, the network parameters are reduced, and compared with the original BLS, the ability of the convolutional BLS network to mine the potential information of features is improved. Finally, the present invention adopts a stacked structure to allow the network to adapt to devices with different computing capabilities according to the number of stacked layers.

[0054] The deep width fault diagnosis method based on the original vibration signal of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0055] Step 1: Get the original vibration signal of the bearing.

[0056] Time domain vibration signal As input, It is a signal of a working condition of a type of fault, containing Vibration signal data points; is the total number of categories of different working conditions under different faults; Fault type label for each class of faults.

[0057] Step 2: Construction of training set.

[0058] The collected original vibration signal is overlapped and sampled, and the sliding window method is used to divide the original vibration signal into data of the same length. The test set, the first-layer training set, and the second-layer training set are constructed in a ratio of 7:2:1.

[0059] Step 3: Construction and training of the first layer of depth-wise separable convolutional network.

[0060] The first layer training set constructed in the second step is input into the first layer of the depthwise separable convolutional network.

[0061] Although convolutional neural networks (CNNs) are often described as black-box models, the computational process in the network can still be represented by mathematical formulas.

[0062] The input data first passes through the depthwise separable convolution blocks with convolution kernels of different scales to extract different features. The depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Each depthwise separable convolution block has several depthwise separable convolutions. Every two depthwise separable convolution blocks will undergo a regularization operation, which can effectively prevent the model from overfitting. The multi-scale feature extraction module is designed to capture local and global features in the input data in order to more fully understand the input and improve the neural network's ability to understand and express the input data.

[0063] ,

[0064] ,

[0065] Formula (1) is the computational cost of standard convolution, and formula (2) is the computational cost of depthwise separable convolution. The computational cost multiplication depends on the number of input channels. , number of output channels , since the present invention uses one-dimensional convolution, the kernel size is , the feature map size is From the formula, we can see that the number of parameters of depthwise separable convolution is much smaller than that of common convolution.

[0066] The BLS network can quickly complete the training of the diagnostic model while ensuring a certain degree of nonlinear feature extraction and incremental capabilities. At the same time, the structure is relatively simple, without too many layers and nodes. Assume that the training sample is , is the associated label, where represents the number of categories. The BLS designed by Group feature nodes and The group enhancement node consists of the following structure:

[0067] First, depthwise separable convolution The generated The group feature nodes are:

[0068] ,

[0069] in is the input data, and the weights in the feature nodes are as follows and , usually initialized with some random samples from a given distribution. The set of group feature nodes is represented as .

[0070] at the same time, A collection of group feature nodes It can also be solved by pseudo-inverse,

[0071] ,

[0072] in , is the canonical norm regularization, Represents the label corresponding to the input data. , the optimal problem shown in formula (4) is set as regular , will have better generalization performance. Represents the sum of square weights This solution is equivalent to the ridge regression theory, which is obtained by or Adding a positive number to the diagonal of gives an approximation to the Moore-Penrose generalized inverse. In theory, if , the inverse problem degenerates into a least squares problem and leads to the solution being the original pseudo-inverse. On the other hand, if , the solution is severely constrained and tends to 0. Therefore, we have:

[0073] ,

[0074] in is the identity matrix. Specifically, we can conclude that The pseudo-inverse have:

[0075] .

[0076] Second, The group enhancement node consists of the original convolution Generate, that is:

[0077] ,

[0078] Similarly, the weights in the enhancement and Initialized using random samples drawn from a specific distribution. All augmented nodes are represented as .

[0079] Finally, Group feature nodes and The set of group enhancement nodes is output through the original convolution dimension reduction , as the input of the second layer of depth-separable convolutional network. Afterwards, Group feature nodes and The collection of group enhancement nodes passes through the depth-wise separable convolution classification layer to obtain a desired result The resulting equation is:

[0080] ,

[0081] in is the weight of the output layer of the first depthwise separable convolutional network, is the bias of the output layer of the first depthwise separable convolutional network.

[0082] After the first layer of depth-separable convolutional network model training is completed, the model parameters are saved for calling during the second layer of depth-separable convolutional network model training.

[0083] Step 4: Construction of the second layer deep separable convolutional network training set.

[0084] Call the first layer of the deep separable convolutional network trained in the third step, input the second layer training set constructed in the second step, and convert the original label of the first layer of the deep separable convolutional network Converted to one-hot encoding, and the actual output of the first layer of the depth-separable convolutional network Subtract to get the fitting target of the second layer of depth-separable convolutional network , combining the dimension reduction output of the feature node and the enhanced node set , and the expected output of the first layer of the depthwise separable convolutional network Construct a training set for the second layer of the depthwise separable convolutional network.

[0085] Step 5: Construction and training of the second layer of depth-wise separable convolutional network.

[0086] The training set constructed in the fourth step is input into the second-layer depth-separable convolutional network, which is composed of stacked BLS networks. The input data of the second-layer depth-separable convolutional network is represented as , which is the dimensionality reduction output of the fourth step ; The output data is the output of each BLS block, with Stacked BLS blocks The structure of is as follows.

[0087] First, in order to separate the individual stacked BLS blocks from the original BLS, the blocks are reformulated as follows.

[0088] ,

[0089] in , For the The output of each block, weight and Randomly generated, ,function is the identity function, that is ; and is the feature convolution, express and Composite mapping. and ,for , different values ​​can be selected according to the complexity of the modeling task. Without loss of generality, the Random convolution and Feature convolution Similarly, Feature convolution and Random convolution The subscript is also omitted. Represents a generalized function feature maps, represents the generalized function that generates the enhanced nodes, express and Composite mapping.

[0090] Second, to fully utilize the output of each stacked BLS block, the sum of the outputs of all stacked BLS blocks approximates the desired output ; Adjacent stacked BLS blocks are combined through residuals, where each stacked BLS block has the same structure, The expected output of each block is:

[0091] ;

[0092] For any given training data , the first stacked BLS block is constructed by setting and generate,

[0093] Right now , where the weight and Randomly generated, weighted and Through the matrix Calculate, that is:

[0094] ;

[0095] In fact, the above process generates a standard structure of the original BLS.

[0096] Generate a new dataset according to equation (12) to train the subsequent stacked BLS blocks. The transformation output of the first stacked BLS structure is is fed into the next single stacked BLS block as training data, which is represented as , the expected output of the second stacked BLS block For input label With the first stacked BLS block output The difference, that is , the parameters of each stacked BLS block model are incrementally updated based on training each stacked BLS block; specifically, there are:

[0097] ;

[0098] The weight and Randomly generated; for Increment of stacked BLS blocks, ,have:

[0099] ,

[0100] The weight and Randomly generated, and:

[0101] .

[0102] The prediction output of the first layer of the deep separable convolutional network is added to the prediction output of the second layer of the deep separable convolutional network to obtain the final prediction output, completing the construction of the deep and wide fault diagnosis network.

[0103] Step 6: Troubleshooting.

[0104] The new vibration data obtained by sampling is input into the deep and wide fault diagnosis network to output the health status type.

Claims

1. A deep width fault diagnosis method based on original vibration signals, characterized in that: The following steps are involved: The first step is to obtain the vibration signal data of the bearing and the corresponding fault type label value of each category of fault; pre-process the vibration signal and split the test set, the first-layer training set and the second-layer training set; The second step is to build the first-layer depth-separable convolutional network, input the first-layer training set into the first-layer depth-separable convolutional network for training, save the parameters after the training is completed, and obtain the first-layer model; The construction and training process of the first layer of depth-wise separable convolutional network is as follows: Input the first layer training set into the first layer of depth-separable convolutional network; The input data first passes through the depthwise separable convolution blocks with convolution kernels of different scales to extract different features. The depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Each depthwise separable convolution block has several depthwise separable convolutions. Every two depth-separable convolution blocks undergo a regularization operation; the designed BLS network consists of n groups of feature nodes and m groups of enhancement nodes, with the following structure: First, the i-th group of feature nodes generated by the depthwise separable convolution φi is: Where Z is the input data, G ei and β ei is the weight within the feature node, and the set of n groups of feature nodes is expressed as At the same time, the set Z of n groups of feature nodes n Solved by pseudo-inverse, the objective function is: Where l represents the label corresponding to the input data, σ1=σ2=u=v=2 is taken, the optimal problem of the objective function is set to regular l2, and the value λ represents the constraint on the sum of square weights G; Second, the jth group of enhanced nodes consists of the original convolution ξ j Generate, that is: in and is the weight in enhancement, and all enhanced nodes are represented as Next, the set of n groups of feature nodes and m groups of enhanced nodes are reduced in dimension through the original convolution and output R1 as the input of the second layer of depth-separable convolutional network; Finally, the set of n groups of feature nodes and m groups of enhancement nodes are passed through the depth-separable convolution classification layer to obtain the desired output Y: Among them G m is the weight of the output layer of the first depth-wise separable convolutional network, β m is the bias of the output layer of the first depthwise separable convolutional network; After the first layer of depth-separable convolutional network model training is completed, the model parameters are saved for calling during the second layer of depth-separable convolutional network model training; Step 3. The training set construction method of the second layer of the depth-separable convolutional network is as follows: Input the second-layer training set into the trained first-layer depth-separable convolutional network, convert the original label y of the first-layer depth-separable convolutional network into one-hot encoding, and subtract it from the actual output t1 of the first-layer depth-separable convolutional network to obtain the label y-t1 as the fitting target of the second-layer depth-separable convolutional network; The label y-t1, the dimension reduction output R1 of the feature node and the enhanced node set, and the expected output Y of the first layer of the deep separable convolutional network are constructed together to obtain a training set of the second layer of the deep separable convolutional network; The second layer of the depth-separable convolutional network is composed of a stacked BLS network, and the construction and training process is as follows: The input data of the second-layer depth-separable convolutional network is represented as x, that is, the dimension reduction output R1; the output data is the output t of each BLS block. i , with n stacked BLS blocks K i The structure is as follows: First, formulate the block as follows: where i = 2, ..., n, t i is the output of the i-1th block, weight and Randomly generated, v i =g(t i-1 ), function g is the identity function, that is, v i =t i-1 ; and is the feature convolution, express and Composite mapping of ; Secondly, the sum of the outputs of all stacked BLS blocks approximates the expected output y-t1; adjacent stacked BLS blocks are combined through residuals, where each stacked BLS block has the same structure, and the expected output of the i-th block is: For the training data {(v i ,y i ), i = 1, ..., n}, the first stacked BLS block is generated by setting y1 = y-t1 and v1 = R1, i.e. The weight and Randomly generated, weighted and Through the matrix Calculate, that is: The conversion output t2 of the first stacked BLS structure is fed into the next separate stacked BLS block as training data, which is represented as t3=g(t2). The expected output y2 of the second stacked BLS block is the difference between the input label y-t1 and the output t2 of the first stacked BLS block, that is, y2=y-t1-t2. The parameters of each stacked BLS block model are incrementally updated based on the training of each stacked BLS block; specifically, there are: The weight and Randomly generated; for the increment of the i-1th stacked BLS block, i=2,...,n, we have: The weight and Randomly generated, and: The fourth step is to add the prediction output of the first layer of the deep separable convolutional network and the prediction output of the second layer of the deep separable convolutional network to obtain the final prediction output, completing the construction of the deep and wide fault diagnosis network; Step 5: Sample the new vibration data and input it into the deep and wide fault diagnosis network to output the fault type.