Method and system for fault diagnosis of rolling bearing

Abstract

The present disclosure discloses a method and system for fault diagnosis of a rolling bearing, and relates to the field of rolling bearing fault diagnosis. The method includes: S101, obtaining a curre

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of rolling bearing fault diagnosis, and in particular, to a method and system for fault diagnosis of a rolling bearing. BACKGROUND

[0002] Rolling bearings are prone to faults such as fracture, wear, fatigue spalling during operation, which lead to property losses, or even casualties, etc. Therefore, timely diagnosis of rolling bearing faults is of great significance for safe production.

[0003] Currently, most fault diagnosis is implemented by selecting single-domain information of signals, such as time domain, frequency domain or time-frequency domain information, for deep learning. However, due to interference of factors such as noise, different fault types have different sensitivities to signals in different domains, which may lead to poor robustness of a diagnosis model. Therefore, it is of great significance to study a method for fault diagnosis based on a multi-feature domain input.

[0004] With the development of computer technology, networks such as a convolutional neural network, a deep belief network, and a long short-term memory network have risen rapidly. The convolutional neural network is widely used due to its features such as strong learning ability, but with the deepening of the network, network degradation may be caused. Therefore, many researchers have tried to add hop connection between convolutional layers, especially DenseNet uses a dense connection mode, which ensures maximization of information between layers in the network, and not only alleviates the learning problem, but also introduces feature reuse. DenseNet has remarkable advantages, but also has obvious disadvantages. Firstly, each layer of the network obtains a feature map from a front layer in a combined manner, without considering interdependence of different channels. Secondly, rolling bearings are susceptible to noise interference during operation, and the use of a deeper network easily introduces noise, which affects an effect of feature extraction. Thirdly, the DenseNet structure has many network layers, a complicated structure, and a low calculation speed, so that it cannot quickly recognize a fault and cannot feed back fault information in a timely manner, resulting in mechanical shutdown and property losses. Fourthly, the network is composed of dense blocks and transition layers by alternate connection. The dense blocks are only responsible for feature extraction, and sizes of feature maps between layers remain unchanged. It is required to reduce dimensions of feature maps at the transition layers, which leads to redundancy in most calculations of dense blocks.

[0005] Therefore, based on the above problems, a novel method or system for fault diagnosis of a rolling bearing is urgently needed to improve precision of fault diagnosis of rolling bearings. SUMMARY

[0006] An objective of the present disclosure is to provide a method and system for fault diagnosis of a rolling bearing, which can improve precision of fault diagnosis of rolling bearings.

[0007] To achieve the above objective, the present disclosure provides the following technical solutions:

[0008] A method for fault diagnosis of a rolling bearing includes:

[0009] obtaining a current one-dimensional time series signal of a rolling bearing;

[0010] determining a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal; and

[0011] determining, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, where the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes: convolutional layers, improved DenseNet, a deep feature fusion unit, a tiled layer, a fully connected layer, and a softmax classification layer, and the diagnosis result includes: a normal state, an inner ring fault, an outer ring fault, or a rolling element fault; where

[0012] a process of determining the diagnosis result by using the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes:

[0013] extracting features from the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal respectively by using the convolutional layers, to obtain corresponding shallow features;

[0014] extracting features from the shallow features respectively by using the improved DenseNet, to obtain corresponding deep features, where the improve DenseNet includes improved dense blocks and transition layers; the improved dense blocks are each implemented by introducing pooling layers and multi-scale convolution kernels into a dense block; and a squeeze-and-excitation module is added after each convolutional layer in the improved dense block;

[0015] performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features; and

[0016] outputting results of the global features sequentially through the tiled layer, the fully connected layer, and the softmax classification layer to obtain the diagnosis result.

[0017] Optionally, the determining a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal specifically includes:

[0018] standardizing the current one-dimensional time series signal;

[0019] performing Fourier transform on the current standardized one-dimensional time series signal to determine the corresponding frequency domain signal; and

[0020] performing envelope spectrum analysis on the current standardized one-dimensional time series signal to determine the corresponding envelope spectrum signal.

[0021] Optionally, before the determining, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, the method further includes:

[0022] obtaining one-dimensional time series data of different state types of the rolling bearing;

[0023] determining a frequency domain signal and an envelope spectrum signal corresponding to the one-dimensional time series data of different state types;

[0024] determining a data set based on the one-dimensional time series data of different state types and the corresponding frequency domain signal and envelope spectrum signal;

[0025] dividing the data set into a training set, a validation set, and a test set;

[0026] training the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion by using the training set; and

[0027] validating and testing the trained rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion by using the validation set and the test set respectively.

[0028] Optionally, a loss function of the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion is a cross-entropy loss function.

[0029] Optionally, the cross-entropy loss function is: q) = p<x ) log q{x )

[0030]

[0031] where mP^ is the cross-entropy loss function, is a true tag value of a type of input data, is a predicted probability value of the jth type of input data, and is a number of classifications.

[0032] Optionally, the performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features specifically includes:

[0033] calculating a weight of each deep feature by multi-level feature dimension reduction;

[0034] determining an importance weight of each deep feature; and

[0035] determining the global features based on the weights of the deep features and the importance weights of the deep features.

[0036] A system for fault diagnosis of a rolling bearing includes: 3

[0037] a one-dimensional time series signal acquisition module, configured to obtain a current one-dimensional time series signal of a rolling bearing;

[0038] a transform domain signal determining module, configured to determine a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal; and

[0039] a diagnosis result determining module, configured to determine, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, where the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes: convolutional layers, improved DenseNet, a deep feature fusion unit, a tiled layer, a fully connected layer, and a softmax classification layer, and the diagnosis result includes: a normal state, an inner ring fault, an outer ring fault, or a rolling element fault; where

[0040] a process of determining the diagnosis result by using the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes:

[0041] extracting features from the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal respectively by using the convolutional layers, to obtain corresponding shallow features;

[0042] extracting features from the shallow features respectively by using the improved DenseNet, to obtain corresponding deep features, where the improve DenseNet includes improved dense blocks and transition layers; the improved dense blocks are each implemented by introducing pooling layers and multi-scale convolution kernels into a dense block; and a squeeze-and-excitation module is added after each convolutional layer in the improved dense block;

[0043] performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features; and

[0044] outputting results of the global features sequentially through the tiled layer, the fully connected layer, and the softmax classification layer to obtain the diagnosis result.

[0045] Optionally, the transform domain signal determining module specifically includes:

[0046] a standardization processing unit, configured to standardize the current one-dimensional time series signal;

[0047] a Fourier transform unit, configured to perform Fourier transform on the current standardized one-dimensional time series signal to determine the corresponding frequency domain signal; and

[0048] an envelope spectrum analysis unit, configured to perform envelope spectrum analysis on the current standardized one-dimensional time series signal to determine the corresponding envelope spectrum signal.

[0049] A system for fault diagnosis of a rolling bearing includes at least one processor, at least one memory, and computer program instructions stored in the memory, where when the computer program instructions are executed by the processor, the method is implemented.

[0050] According to the specific embodiments provided by the present disclosure, the present disclosure discloses the following technical effects:

[0051] In the method and system for fault diagnosis of a rolling bearing according to the present disclosure, multi-domain signals are used to diagnose a fault of a rolling bearing. Compared with a deep learning algorithm using single dimension data, the method solves the problem of single original signal information of rolling bearings. The multi-domain signals enable valuable information to be extracted from a plurality of transform domains simultaneously, and information losses are minimized, thereby providing more important feature information for fault diagnosis, and enabling a model to keep good accuracy under noise and off-design conditions. Improved DenseNet is used to extract deep features, pooling layers are fused with dense blocks, and multi-scale convolution kernels and squeeze-and-excitation modules (SEMs) are introduced, thereby not only improving feature reuse and reducing dimensions of feature maps, but also automatically calculating an importance weight of each feature channel, enhancing useful features of a current task and suppressing useless features, improving accuracy of data classification, achieving a good generalization ability, accelerating network convergence, and improving a fitting ability and off-design diagnosis performance of the model. An attention mechanism method is used to perform weighted fusion on deep features of a plurality of signals to ensure diagnosis precision. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] To describe the technical solutions in the embodiments of the present disclosure or in the prior art more clearly, the accompanying drawings required for the embodiments are briefly described below. Apparently, the accompanying drawings in the following description show merely some embodiments of the present disclosure, and a person of ordinary skill in the art may still derive other accompanying drawings from these accompanying drawings without creative efforts.

[0053] FIG. 1 is a schematic flowchart of a method for fault diagnosis of a rolling bearing according to the present disclosure;

[0054] FIG. 2 is a schematic diagram of a determining process by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion according to the present disclosure;

[0055] FIG. 3 is a schematic diagram of a data set test bed of Case Western Reserve University Bearing Data Center;

[0056] FIG. 4 is a schematic structural diagram of a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion according to the present disclosure;

[0057] FIG. 5 is a schematic structural diagram of improved DenseNet in a method according to the present disclosure;

[0058] FIG. 6 is a schematic diagram of an SEM of improved DenseNet in a method according to the present disclosure; and

[0059] FIG. 7 is a schematic diagram of an attention mechanism feature fusion method in a method according to the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The technical solutions of the embodiments of the present disclosure are clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are merely some rather than all of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0061] An objective of the present disclosure is to provide a method and system for fault diagnosis of a rolling bearing, which can improve precision of fault diagnosis of rolling bearings.

[0062] To make the above-mentioned objective, features, and advantages of the present disclosure clearer and more comprehensible, the present disclosure will be further described in detail below in conjunction with the accompanying drawings and specific implementations.

[0063] As shown in FIGs. 1 and 2, a method for fault diagnosis of a rolling bearing according to the present disclosure includes the following steps.

[0064] S101: Obtain a current one-dimensional time series signal of a rolling bearing.

[0065] SI02: Determine a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal, that is, obtain a plurality of transform domain signals.

[0066] Step SI02 specifically includes:

[0067] standardizing the current one-dimensional time series signal to convert all data points into values between 0 and 1, where a standardization formula is as follows:

[0068] Xmax Xmm , X                                                           Y        X

[0069] where ! is a value of a point in one-dimensional time series data, and and max are a minimum value and a maximum value in the one-dimensional time series data;

[0070] then extending the current standardized one-dimensional time series signal to a plurality of feature domains, and as shown in FIG. 4, performing Fourier transform and envelope spectrum analysis on original one-dimensional time series data to obtain a frequency domain signal and an envelope spectrum signal, where the two signals and a time domain signal form an input of a model;

[0071] performing Fourier transform on the current standardized one-dimensional time series signal to determine the corresponding frequency domain signal; and

[0072] performing envelope spectrum analysis on the current standardized one-dimensional time series signal to determine the corresponding envelope spectrum signal.

[0073] SI03: Determine, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, where the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes: convolutional layers, improved DenseNet, a deep feature fusion unit, a tiled layer, a fully connected layer, and a softmax classification layer, and the diagnosis result includes: a normal state, an inner ring fault, an outer ring fault, or a rolling element fault.

[0074] A process of determining the diagnosis result by using the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes:

[0075] extracting features from the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal respectively by using the convolutional layers, to obtain corresponding shallow features;

[0076] extracting features from the shallow features respectively by using the improved DenseNet, to obtain corresponding deep features, where the improve DenseNet includes improved dense blocks and transition layers; the improved dense blocks are each implemented by introducing pooling layers and multi-scale convolution kernels into a dense block; and a squeeze-and-excitation module is added after each convolutional layer in the improved dense block;

[0077] performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features; where

[0078] a specific process of obtaining the global features includes:

[0079] calculating a weight of each deep feature by multi-level feature dimension reduction;

[0080] determining an importance weight of each deep feature; and

[0081] determining the global features based on the weights of the deep features and the importance weights of the deep features; and

[0082] outputting results of the global features sequentially through the tiled layer, the fully connected layer, and the softmax classification layer to obtain the diagnosis result.

[0083] Before step S103, the method further includes the following steps.

[0084] One-dimensional time series data of different state types of the rolling bearing is obtained. As shown in FIG. 3, the data came from a data set of Case Western Reserve University Bearing Data Center, and driving end bearing state data was used, with a collection frequency of 12 kHz. The above data includes ten bearing state types, that is, a normal state, an inner ring fault, an outer ring fault, and a rolling element fault. Damage diameters of the outer ring fault, the inner ring fault, and the rolling element fault were 7, 14, and 21 mils respectively, and 100,000 data points were collected for each bearing state, totaling 1,000,000 data points. Only one bearing state type existed in this data set at the same time, and there was no concurrency.

[0085] A frequency domain signal and an envelope spectrum signal corresponding to the one-dimensional time series data of different state types are determined.

[0086] 500 samples were randomly taken for each bearing state of an original signal, and each sample had 1,024 points. Then the one-dimensional original signal was preprocessed to obtain a frequency domain signal and an envelope spectrum signal. Because of symmetry of a sampling frequency, sample data points of the frequency domain signal and the envelope spectrum signal each had a length that was only a half of the sample length of the original signal, and there were 512 sample data points. 5,000 samples were taken for each of the three types of signals, totaling 15,000 samples. Sample composition information of time domain signals is shown in Table 1.

[0087] Table 1 Sample composition information of the time series signals Sample type Sample length Number of samples Sample tag Normal 1024 500 0 Inner ring damage (7 mils) 1024 500 1 Outer ring damage (7 mils) 1024 500 2 Rolling element damage (7 mils) 1024 500 3 Inner ring damage (14 mils) 1024 500 4 Outer ring damage (14 mils) 1024 500 5 Rolling element damage (14 mils) 1024 500 6 Inner ring damage (21 mils) 1024 500 7 Outer ring damage (21 mils) 1024 500 8 Rolling element damage (21 mils) 1024 500 9

[0088] A data set is determined based on the one-dimensional time series data of different state types and the corresponding frequency domain signal and envelope spectrum signal.

[0089] The data set is divided into a training set, a validation set, and a test set; and the training set, the validation set, and the test set account for 60%, 20%, and 20% respectively.

[0090] The rolling bearing fault diagnosis model is trained based on improved DenseNet and deep feature fusion by using the training set.

[0091] The training process is as follows:

[0092] First, the convolutional layer adopts 32 large convolution kernels with a size of 1 * 64, and the three types of signals have strides of 16, 8, and 8 respectively, to extract shallow features of signals. Improved DenseNet is used to extract deep features of the signals, and mainly includes four convolutional layers and six pooling layers, where the first three convolutional layers are convolution kernels with sizes of 1 * 3 and 1 * 5 and a stride of 2 and each have 16 convolution kernels; the last convolutional layer has convolution kernels with a size of 1 * 3 and a stride of 2, and has 32 convolution kernels. The network introduces pooling layers and multi-scale convolution kernels into dense blocks, and an SEM is added is added after each convolutional layer, so that the weight of each channel can be adjusted, thereby enhancing the expressive ability of features.

[0093] Second, in the deep feature fusion unit, weighted fusion is performed on deep features of the three types of signals by using the attention mechanism method.

[0094] Then, a fused feature map is flattened into a form of a one-dimensional vector, which is input into a softmax output layer by using the fully connected layer, to obtain a predicted probability value of the data.

[0095] Finally, a cross-entropy loss function is used to calculate an error between a true value and a predicted value, a gradient descent algorithm is used to perform back propagation calculation on the obtained error, parameters of the model are updated, and the above steps are repeated. When a number of training times reaches a maximum number of iterations, the model training is finished, and trained model parameters are saved. A model connection mode is shown in FIG. 4. A specific process is as follows:

[0096] The convolutional layer adopts 32 large convolution kernels with a size of 1 * 64, and the three types of signals have strides of 16, 8, and 8 respectively, to extract shallow features of signals. An expression of the convolution process is as follows:

[0097]

[0098] j is an output of the convolutional layer, 1 is an input of the convolutional layer, Wij is a weight of the convolutional layer, is a bias of the convolutional layer, * is convolution operation, and f is an ReLU nonlinear activation function, with an expression of Rulu(x) = max(0, v)

[0099] Improved DenseNet mainly includes improved dense blocks and SEMs, and is used to extract deep features of signals.

[0100] The improved dense block is improved from the DenseNet network, which is mainly composed of dense blocks and transition layers, where the dense blocks function to extract features and fuse multi-layer features, and the pooling layer functions to reduce a dimension of a feature map, but this network has many disadvantages. Based on this, the present disclosure introduces pooling layers into a dense block, and proposes an improved dense block structure. As shown in FIG. 5, the improved dense block structure mainly includes four convolutional layers and six pooling layers, where the first three convolutional layers are convolution kernels with sizes of 1 x 3 and 1 x 5 and a stride of 2 and each have 16 convolution kernels; and the last convolutional layer has convolution kernels with a size of 1x3 and a stride of 2, and has 32 convolution kernels. This structure has the advantages of the DenseNet network, reduces the complexity of the model, increases the training speed, and enhances the anti-noise ability of the network.

[0101] After being connected to the convolutional layers of the improved dense block, the SEMs are used to adjust channel weights and enhance the interdependence of different channels, which improves off-design diagnosis performance, as shown in FIG. 6.

[0102] The deep feature fusion unit fuses the extracted deep features by using improved DenseNet. A main process of a weighted feature fusion algorithm using the attention mechanism in the present disclosure include Excitation, Proportion, and Fusion. As shown in FIG. 7, a specific process is as follows.

[0103] Excitation: A weight ! of a feature by multi-level feature dimension reduction is calculated.

[0104]

[0105] where hi is a deep feature corresponding to time domain, frequency domain and envelope spectrum signals, i is a number of features, Wj and are a weight and a bias, and is a sigmoid nonlinear activation function. Cl

[0106] Proportion: An importance weight ! of each feature in a feature map is calculated. a

[0107]

[0108] where n is a number of features in each feature map.

[0109] Fusion: Weighted fusion operation is performed on deep features to obtain global features f. f = \aihi

[0110] =i ; and

[0111] a calculation formula of the fully connected layer is as follows:

[0112] / +1

[0113] where is a value of an h th neuron at a (^ + 1 )th layer, h is a value of an h th neuron at an ^th layer, lh and are a weight and a bias of the fully connected layer, and * is an ReLU nonlinear activation function.

[0114] A softmax layer is mainly used to calculate a predicted probability value of a bearing state, and an expression of softmax is as follows: pZm Pi^-- ZgZ"

[0115] «=i P(z ) z

[0116] where ' is a predicted probability value of an m th type, m is a value of an wth output neuron of the fully connected layer, and is a number of classifications.

[0117] The expression of the cross-entropy loss function is as follows: N

[0118] >=1

[0119] where P^x^ is a true tag value of a J th type, is a predicted probability value of the Jth type, and is a number of classifications.

[0120] The validation set is used to validate the trained network model, network model parameters are outputted, and the test set is used to test recognition accuracy of the model.

[0121] In order to test accuracy of the model, the method is compared with DenseNet, ResNet, and multi-scale DenseNet. Network precision is shown in Table 2. It can be seen from the table that the method herein has the highest precision, which reached 99.772%.

[0122] Table 2 Recognition accuracy of different methods / % Method 1 2 3 4 5 Mean ResNet 99.32 99.15 99.30 99.22 99.34 99.133 DenseNet 99.58 99.33 99.74 99.26 99.55 99.492 Multi-scale 99.65 99.57 99.38 99.43 99.66 99.538 DenseNet Present disclosure 99.80 99.72 99.65 99.80 99.89 99.772

[0123] The method according to the present disclosure can effectively fuse various transform domain information, and increase diversity and integrity of information.

[0124] In addition, the improved DenseNet not only simplifies the model and reduces the training time, but also reduces the noise interference, and improves accuracy of rolling bearing fault diagnosis. Multi-layer feature fusion is implemented while multi-domain information fusion is implemented, network convergence is accelerated, and the fitting ability and off-design diagnosis performance of the model are improved.

[0125] Corresponding to the above method, the present disclosure further provides a system for fault diagnosis of a rolling bearing, including:

[0126] a one-dimensional time series signal acquisition module, configured to obtain a current one-dimensional time series signal of a rolling bearing;

[0127] a transform domain signal determining module, configured to determine a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal; and

[0128] a diagnosis result determining module, configured to determine, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, where the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes: convolutional layers, improved DenseNet, a deep feature fusion unit, a tiled layer, a fully connected layer, and a softmax classification layer, and the diagnosis result includes: a normal state, an inner ring fault, an outer ring fault, or a rolling element fault; where

[0129] a process of determining the diagnosis result by using the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion includes:

[0130] extracting features from the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal respectively by using the convolutional layers, to obtain corresponding shallow features;

[0131] extracting features from the shallow features respectively by using the improved DenseNet, to obtain corresponding deep features, where the improve DenseNet includes improved dense blocks and transition layers; the improved dense blocks are each implemented by introducing pooling layers and multi-scale convolution kernels into a dense block; and a squeeze-and-excitation module is added after each convolutional layer in the improved dense block;

[0132] performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features; and

[0133] outputting results of the global features sequentially through the tiled layer, the fully connected layer, and the softmax classification layer to obtain the diagnosis result.

[0134] The transform domain signal determining module specifically includes:

[0135] a standardization processing unit, configured to standardize the current one-dimensional time series signal;

[0136] a Fourier transform unit, configured to perform Fourier transform on the current standardized one-dimensional time series signal to determine the corresponding frequency domain signal; and

[0137] an envelope spectrum analysis unit, configured to perform envelope spectrum analysis on the current standardized one-dimensional time series signal to determine the corresponding envelope spectrum signal.

[0138] To implement the corresponding method of the first embodiment so as to achieve corresponding functions and technical effects, the present disclosure further provides a system for fault diagnosis of a rolling bearing, including at least one processor, at least one memory, and computer program instructions stored in the memory, where when the computer program instructions are executed by the processor, the method is implemented.

[0139] Each embodiment of the description is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and for the same and similar parts between the embodiments, reference may be made to each other. Since the system disclosed in an embodiment corresponds to the method disclosed in another embodiment, the description is relatively simple, and reference may be made to the method description.

[0140] Specific examples are used herein to explain the principles and implementations of the present disclosure. The foregoing description of the embodiments is merely intended to help understand the method of the present disclosure and its core ideas; besides, various modifications may be made by a person of ordinary skill in the art to specific implementations and the scope of application in accordance with the ideas of the present disclosure. In conclusion, the content of the present specification shall not be construed as limitations to the present disclosure.

Claims

1. A method for fault diagnosis of a rolling bearing, comprising:obtaining a current one-dimensional time series signal of a rolling bearing;determining a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal; anddetermining, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, wherein the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion comprises: convolutional layers, improved DenseNet, a deep feature fusion unit, a tiled layer, a fully connected layer, and a softmax classification layer, and the diagnosis result comprises: a normal state, an inner ring fault, an outer ring fault, or a rolling element fault; whereina process of determining the diagnosis result by using the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion comprises:extracting features from the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal respectively by using the convolutional layers, to obtain corresponding shallow features;extracting features from the shallow features respectively by using the improved DenseNet, to obtain corresponding deep features, wherein the improve DenseNet comprises improved dense blocks and transition layers; the improved dense blocks are each implemented by introducing pooling layers and multi-scale convolution kernels into a dense block; and a squeeze-and-excitation module is added after each convolutional layer in the improved dense block;performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features; andoutputting results of the global features sequentially through the tiled layer, the fully connected layer, and the softmax classification layer to obtain the diagnosis result.

2. The method for fault diagnosis of a rolling bearing according to claim 1, wherein the determining a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal specifically comprises:standardizing the current one-dimensional time series signal;performing Fourier transform on the current standardized one-dimensional time series signal to determine the corresponding frequency domain signal; andperforming envelope spectrum analysis on the current standardized one-dimensional time series signal to determine the corresponding envelope spectrum signal.

3. The method for fault diagnosis of a rolling bearing according to claim 1, wherein before the determining, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, the method further comprises:obtaining one-dimensional time series data of different state types of the rolling bearing;determining a frequency domain signal and an envelope spectrum signal corresponding to the one-dimensional time series data of different state types;determining a data set based on the one-dimensional time series data of different state types and the corresponding frequency domain signal and envelope spectrum signal;dividing the data set into a training set, a validation set, and a test set;training the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion by using the training set; andvalidating and testing the trained rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion by using the validation set and the test set respectively.

4. The method for fault diagnosis of a rolling bearing according to claim 3, wherein a loss function of the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion is a cross-entropy loss function.

5. The method for fault diagnosis of a rolling bearing according to claim 4, wherein the cross-entropy loss function is:NH(p, q) = p(Xj) log q(Xj)7=1wherein ^^P’^ is the cross-entropy loss function, is a true tag value of a J th type of input data, is a predicted probability value of the Jth type of input data, and is a number of classifications.

6. The method for fault diagnosis of a rolling bearing according to claim 1, wherein the performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features specifically comprises:calculating a weight of each deep feature by multi-level feature dimension reduction;determining an importance weight of each deep feature; anddetermining the global features based on the weights of the deep features and the importance weights of the deep features.

7. A system for fault diagnosis of a rolling bearing, comprising:a one-dimensional time series signal acquisition module, configured to obtain a current one-dimensional time series signal of a rolling bearing;a transform domain signal determining module, configured to determine a corresponding frequency domain signal and envelope spectrum signal based on the current one-dimensional time series signal; anda diagnosis result determining module, configured to determine, based on the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal, a diagnosis result by using a rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion, wherein the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion comprises: convolutional layers, improved DenseNet, a deep feature fusion unit, a tiled layer, a fully connected layer, and a softmax classification layer, and the diagnosis result comprises: a normal state, an inner ring fault, an outer ring fault, or a rolling element fault; whereina process of determining the diagnosis result by using the rolling bearing fault diagnosis model based on improved DenseNet and deep feature fusion comprises:extracting features from the current one-dimensional time series signal and the corresponding frequency domain signal and envelope spectrum signal respectively by using the convolutional layers, to obtain corresponding shallow features;extracting features from the shallow features respectively by using the improved DenseNet, to obtain corresponding deep features, wherein the improve DenseNet comprises improved dense blocks and transition layers; the improved dense blocks are each implemented by introducing pooling layers and multi-scale convolution kernels into a dense block; and a squeeze-and-excitation module is added after each convolutional layer in the improved dense block;performing weighted feature fusion on the deep features by using an attention mechanism method in the deep feature fusion unit, to obtain global features; andoutputting results of the global features sequentially through the tiled layer, the fully connected layer, and the softmax classification layer to obtain the diagnosis result.

8. The system for fault diagnosis of a rolling bearing according to claim 7, wherein thetransform domain signal determining module specifically comprises:a standardization processing unit, configured to standardize the current one-dimensional time series signal;a Fourier transform unit, configured to perform Fourier transform on the current standardized one-dimensional time series signal to determine the corresponding frequency domain signal; andan envelope spectrum analysis unit, configured to perform envelope spectrum analysis on the current standardized one-dimensional time series signal to determine the corresponding envelope spectrum signal.

9. A system for fault diagnosis of a rolling bearing, comprising at least one processor, at least one memory, and computer program instructions stored in the memory, wherein when the computer program instructions are executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images