Elevator cross-bearing fault diagnosis method and system

Abstract

The invention discloses a cross-bearing fault diagnosis method and system for an elevator, and relates to the technical field of bearing fault diagnosis. Comprising the steps of 1, building a multi-dimensional space-time collaborative neural network which comprises a domain invariant feature extraction module, a multi-dimensional space-time feature fusion module and a bearing fault classification module; wherein the domain invariant feature extraction module comprises a multi-scale convolution layer and a collaborative normalization layer, and turning to the step 2; the invention innovatively provides a multi-dimensional space-time collaborative neural network model, designs a collaborative normalization method on the basis of a traditional multi-scale convolutional neural network model, and reduces covariant offset and overfitting phenomena through mixed normalization of batch data and single sample data; meanwhile, the multi-dimensional spatial-temporal feature fusion module adaptively captures the importance of different features in local cross-channel information interaction and learning space, and the ability of the model to extract invariant features of a bearing fault domain is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a hoist span bearing fault diagnosis method and system. Background Art

[0002] The mine hoisting system is the core link in mining operations, responsible for transporting ore, waste rock, personnel and equipment from underground to the surface. These systems usually operate under extreme working conditions, such as high loads, harsh climates and complex geological conditions, which makes their key components, especially the spindle bearings at both ends of the hoist, which not only bear the entire torque of the hoist, but also face great wear and failure risks as an important part of winding the hoist wire rope. The failure of the spindle bearing will not only cause the hoist to stop, but may also cause serious safety accidents. Therefore, it is very important to perform effective fault diagnosis and predictive maintenance on the spindle bearings of mine hoists.

[0003] At present, the bearing fault diagnosis at both ends of the mine hoist main shaft still faces a series of challenges:

[0004] 1. Multi-condition operating environment: The mining hoisting system operates under different speed conditions, load requirements and operating modes, resulting in variable operating conditions of the bearings;

[0005] 2. Difficulty in data collection: In mining hoisting systems, due to the harsh environment and complex equipment layout, it is very difficult to fully monitor the bearing status and collect data;

[0006] 3. Model generalization capability requirements: Existing fault diagnosis models often rely on information interaction between multiple source domains and labeled data. When the source domain information is reduced or even only one source domain is available, how to extract key domain-invariant features from samples to achieve efficient generalization learning of the model still faces huge challenges;

[0007] 4. Complexity of feature extraction: The fault characteristics of hoist bearings may behave differently at different time scales and frequencies. Traditional feature extraction methods are difficult to fully capture these complex universal features. Summary of the invention

[0008] The purpose of the present invention is to solve the problems mentioned in the above background technology and to provide a hoist span bearing fault diagnosis method and system.

[0009] The purpose of the present invention can be achieved through the following technical solutions:

[0010] In a first aspect, the present invention provides a method for diagnosing a hoist cross bearing fault, comprising the following steps:

[0011] Step 1: Build a multi-dimensional spatiotemporal collaborative neural network, including a domain-invariant feature extraction module, a multi-dimensional spatiotemporal feature fusion module, and a bearing fault classification module; wherein the domain-invariant feature extraction module includes a multi-scale convolution layer and a collaborative normalization layer, and then proceed to step 2;

[0012] Step 2: Collect vibration signal data under different working conditions, and obtain source domain and target domain data sets after preprocessing; establish a bearing database for each working condition in the source domain, and divide the bearing database into a training set and a test set, and then proceed to step 3:

[0013] Step 3: Build a multi-scale convolution layer. The initial input feature signal X passes through the multi-scale convolution layer to obtain the multi-scale convolution features X1, X2, ..., X5 of X, and then proceed to step 4;

[0014] Step 4: Build a collaborative normalization layer, input the multi-scale convolution features X1, X2, ..., X5 into the collaborative normalization layer, obtain the normalized results Z1, Z2, ..., Z5 of the features, and proceed to step 5;

[0015] Step 5: Build a multi-dimensional spatiotemporal feature fusion module, input the normalized results Z1, Z2, ..., Z5 into the multi-dimensional spatiotemporal feature fusion module, realize the adaptive extraction of important features in channels and spaces, complete the feature extraction of Layer1, and then obtain multi-scale features MF1, MF2..., MF5 through Layer2, Layer3, ... Layer5 based on steps 3 to 5, and proceed to step 6;

[0016] Step 6: Implement the cascade of multi-scale features MF1, MF2, ..., MF5 along the channel, confirm the final fusion feature MF, and build a bearing fault classification module. Input the fusion feature MF into the fault classification module to realize the bearing fault classification in the source domain, and then proceed to step 7;

[0017] Step 7: Repeat the training of the overall network model based on steps 3 to 6 to obtain a trained overall network model, and input the test set of the target domain into the trained overall network model to obtain the test set accuracy.

[0018] As a preferred implementation of the present invention, in step 2, vibration signal data of different working conditions are collected, and after preprocessing, source domain and target domain data sets are obtained respectively; a bearing database for each working condition in the source domain is established, and the bearing database is divided into a training set and a test set, specifically:

[0019] Based on the public PU bearing dataset, bearing datasets under multiple different working conditions are collected. For each working condition, m sample data are taken to establish a sample database in the source domain and the target domain. The sample database is divided into a training set and a test set. The sample ratio of the training set and the test set is 8:2.

[0020] As a preferred embodiment of the present invention, the multi-scale convolution layer includes 5 convolution layers with different convolution kernel sizes and 1 wide convolution layer with a convolution kernel size of 64, specifically:

[0021] On the PU dataset, the initial input feature signal X passes through the multi-scale convolution layer to obtain the multi-scale convolution features X1, X2, ..., X5 of X, that is, the convolution features under different convolution kernel sizes:

[0022] X1 = conv1(conv(X));

[0023] X2 = conv2(conv(X));

[0024] X3 = conv3(conv(X));

[0025] X4 = conv4(conv(X));

[0026] X5 = conv5(conv(X));

[0027] Where X is the input feature signal, conv is a wide convolutional layer, and conv1, conv2, conv3, conv4, and conv5 are convolutional layers of different scales.

[0028] As a preferred implementation of the present invention, a collaborative normalization layer is constructed, and the multi-scale convolutional features X1, X2, ..., X5 obtained in step 3 are input to the collaborative normalization layer to obtain normalized results Z1, Z2, ..., Z5 of the features. Taking the multi-scale convolutional feature X1 as the input of the collaborative normalization layer as an example, the specific steps are as follows:

[0029] Step 4-1: Split each channel equally into two sub-channels C1 and C2, which can be achieved by:

[0030] X1 c1 =X1[:,0:C / / 2,:,:];

[0031] X1 c2 =X1[:,C / / 2:c,:,:];

[0032] Step 4-2, in channel C1, Divide the data into several groups and perform BN operations, and update the parameters of the data by group; specifically:

[0033]

[0034] in, represents the mean value obtained by performing batch normalization in channel C1, represents the variance value obtained by performing batch normalization in channel C1, Represents the scale transformation and offset of the batch normalization result in channel C1;

[0035] Step 4-3, in channel C2, Based on the IN strategy, the mean and standard deviation of each sample in the H and W dimensions are solved, and the N and C dimensions are retained, that is, the mean and standard deviation are only calculated within the channel, specifically:

[0036]

[0037]

[0038] in, represents the mean value obtained by performing batch normalization in channel C2, represents the variance value obtained by performing batch normalization in channel C1, represents the scale transformation and offset of the instance normalization result in channel C2; I represents instance normalization;

[0039] Step 4-4, recombining the normalized sub-channel results to obtain the final output Z:

[0040] Z1=[Y B，1 , Y I，1 , Y B，2 , Y I，2 ,......,Y B，C / / 2 ,Y I，C / / 2 ];

[0041] Among them, BN is batch normalization, IN is instance normalization, is the eigenvalue of each sample in channel C1 in dimensions N, H, and W, is the eigenvalue of each sample in channel C2 in the H and W dimensions, γc and βc are learnable parameters, ε is a constant, and [.] represents the concatenation along the channel dimension.

[0042] As a preferred implementation of the present invention, a multidimensional spatiotemporal feature fusion module is constructed, and the normalized results Z1, Z2, ..., Z5 are input into the multidimensional spatiotemporal feature fusion module to realize adaptive extraction of important features in channels and spaces, complete the feature extraction of Layer1, and then obtain multi-scale features MF1, MF2 ..., MF5 through Layer2, Layer3, ... Layer5 in sequence based on steps three to five. Taking the normalized result Z1 as the input of the multidimensional spatiotemporal feature fusion module as an example, specifically:

[0043] Step 5-1: For a given input feature Z1, perform maximum pooling and average pooling operations to obtain two features Im and Ia, then concatenate Im and Ia and input them into the channel attention branch and the spatial attention branch respectively;

[0044] Im=max(Z1);

[0045]

[0046] Both Sc and Ss are the results of feature concatenation of Im and Ia;

[0047] Step 5-2: In the channel attention branch, the feature Sc is input into the channel weight generation unit, and two one-dimensional convolution layers with convolution kernel sizes of k1 and k2 are adaptively determined by the number of input channels. Then, the feature Sc is input into the above two one-dimensional convolution layers respectively, and the deep features of Sc are adaptively extracted. Finally, after passing through the Softmax layer, its attention weights Wc1 and Wc2 in the channel branch are obtained;

[0048] Wc1=σ[Conv1(Sc,max(1,[(γ1×C+b1) / 2]))];

[0049] Wc2=σ[Conv2(Sc,max(1,[(γ2×C+b2) / 2]))];

[0050] Among them, γ and b are hyperparameters, [.] represents an odd number rounded up, Conv1 and Conv2 are one-dimensional 1*1 convolutions, and σ is the Softmax activation function;

[0051] Step 5-3: In the spatial attention branch, the feature Ss is input into the spatial weight generation unit. The feature Ss passes through two 7*7 convolutional layers to obtain the initial feature S in the spatial branch. s1 and S s1' Then, the linear layer FC1, Relu activation function and linear layer FC2 are used to realize the dimension increase and dimension reduction of the features, and the important features S of the samples in the spatial branch are obtained. s2 and S s2' , and finally the attention weights Ws1 and Ws2 in the spatial branch are obtained through the Softmax activation function;

[0052] Ws1=σ(FC2(η(FC1(Conv3(Ss)))));

[0053] Ws2=σ(FC2(η(FC1(Conv4(Ss)))));

[0054] Among them, Conv3 and Conv4 are 7*7 convolutions;

[0055] Step 5-4: Summarize the channel weights and spatial weights to determine the important parts of the features between channels and spaces, specifically:

[0056] MF1 1 =(Wc1+Ws1)×Im+(Wc2+Ws2)×Ia;

[0057] Complete the feature extraction of Layer1 and obtain feature MF1 1 , then feature MF1 1 Based on steps 3 to 5, the multi-scale feature MF1 is obtained through layer2, Layer3, ... Layer5 layers in sequence.

[0058] As a preferred implementation of the present invention, in step 6, the multi-scale features MF1, MF2, ..., MF5 are cascaded along the channel, the final fusion feature MF is confirmed, and a bearing fault classifier is built. The fusion feature MF obtained in step 6 is input into the fault classifier to realize fault classification in the source domain, specifically:

[0059] MF=Concate[MF1, MF2, MF3, MF4, MF5];

[0060] OutPut = σ(Fc(MF));

[0061] Combined with the cross entropy loss function to achieve bearing fault classification:

[0062] L=-(ylog(OutPut)+(1-y)log(1-OutPut));

[0063] Among them, Fc is the fully connected layer, σ is the Softmax activation function, Output is the output classification result of the model, L is the classification loss, and y is the fault label in the source domain.

[0064] As a preferred implementation of the present invention, in step 7, the overall network model is repeatedly trained based on steps 3 to 6 to obtain a trained overall network model, and the test set of the target domain is input into the trained overall network model to obtain the test set accuracy and feature visualization results, specifically:

[0065] Step 7-1, select the learning rate, optimizer, training batch size and number of training rounds for the overall network model;

[0066] Step 7-2, input the training data set obtained in step 2 into the set bearing fault diagnosis model to obtain a trained cross-bearing fault diagnosis model;

[0067] Step 7-3: Input the test data under different working conditions in the target domain into the trained cross-bearing fault diagnosis model to realize cross-working condition fault diagnosis of the bearing.

[0068] In a second aspect, the present invention provides a hoist cross-bearing fault diagnosis system, including a domain invariant feature extraction module, a multi-dimensional spatiotemporal feature fusion module and a bearing fault classification module;

[0069] The domain invariant feature extraction module is composed of a multi-scale convolution layer and a collaborative normalization layer, which is used to realize the adaptive extraction and fusion of important features of input samples; the multi-scale convolution layer is composed of a small convolution layer with different convolution kernel sizes and a wide convolution layer, which obtains the feature relationship between samples of different scales and the long-term relationship characteristics of all input samples; the collaborative normalization layer simultaneously learns the statistical characteristics between channels and the independent characteristics between instances, and reduces covariate shift and overfitting by hybrid normalization of batch data and single sample data;

[0070] The multi-dimensional spatiotemporal feature fusion module innovatively replaces the traditional (maximum / average) pooling module to adaptively capture local cross-channel information interaction; a spatial weight generation unit is designed in its spatial branch to adaptively learn the importance of different features in the space;

[0071] The bearing fault classification module consists of a fully connected layer and a Softmax activation function, which is used to input the fused convolutional features into the bearing fault classifier and output the classification result of the bearing fault.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] 1. The present invention only uses data from a single source domain to extract universal domain-invariant features, and can achieve cross-operating condition intelligent fault diagnosis of bearings without relying on any target domain information.

[0074] 2. Based on the traditional multi-scale convolutional neural network model, the present invention combines the advantages of universal feature extraction of the multi-dimensional spatiotemporal feature fusion module and the advantage of collaborative normalization to reduce the internal covariate offset of the sample. The proposed method has an average accuracy of 98.74% in 6 different cross-operating conditions (cross-speed, cross-load) tasks of the PU bearing data set. Compared with some mainstream domain generalization methods, it significantly improves the accuracy and model generalization ability in cross-operating condition bearing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0076] Figure 1 is a flow chart of the method;

[0077] Figure 2 It is a framework diagram of the multi-dimensional spatiotemporal collaborative neural network model;

[0078] Figure 3 Schematic diagram of co-normalization;

[0079] Figure 4 This is the framework diagram of the multi-dimensional spatiotemporal feature fusion module. DETAILED DESCRIPTION

[0080] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] See also Figure 1-Figure 4 As shown, the present invention provides a method for diagnosing a hoist span bearing fault, comprising the following steps:

[0082] Step 1: Build a network of cross-bearing fault diagnosis model, including domain-invariant feature extraction module, multi-dimensional spatiotemporal feature fusion module and bearing fault classification module; the domain-invariant feature extraction module includes multi-scale convolution layer and collaborative normalization layer.

[0083] Step 2: Collect vibration signal data under different working conditions. After preprocessing, obtain source domain and target domain data sets respectively. Establish a bearing database for each working condition in the source domain, and divide the database into a training set and a test set. The ratio of the number of bearings in the training set and the test set is 8:2, as follows:

[0084] Based on the public bearing dataset, bearing datasets under multiple different working conditions are collected. n sample data are taken for each working condition to establish an image database in the source domain and the target domain. The database is divided into a training set and a test set. The sample ratio of the training set and the test set is 8:2.

[0085] Step 3: The multi-scale convolution layer contains 5 convolution layers with different convolution kernel sizes and 1 wide convolution layer with a convolution kernel size of 64, as follows:

[0086] On the PU dataset, the initial input feature signal X passes through the multi-scale convolution layer to obtain the multi-scale convolution features X1, X2…, X5 of X, that is, the convolution features under different convolution kernel sizes:

[0087] X1=conv1(conv(X)) (1)

[0088] X2=conv2(conv(X)) (2)

[0089] X3=conv3(conv(X)) (3)

[0090] X4=conv4(conv(X)) (4)

[0091] X5=conv5(conv(X)) (5)

[0092] Among them, X is the input feature signal, conv is a wide convolutional layer, and conv1, conv2, conv3, conv4, and conv5 are convolutional layers of different scales.

[0093] Step 4: Build a collaborative normalization layer, input the multi-scale convolutional features X1, X2, ..., X5 obtained in step 3 into the collaborative normalization layer, and obtain the normalized results Z1, Z2, ..., Z5 of the features. Take the multi-scale convolutional feature X1 as the input of collaborative normalization as an example, as follows:

[0094] Step 4-1: Split each channel equally into two sub-channels C1 and C2. This can be achieved in the following way:

[0095] X1 c1 =X1[:,0:C / / 2,:,:](6)

[0096] X1 c2 =X1[:,C / / 2:c,:,:](7)

[0097] Step 4-2, in channel C1, The data is divided into several groups and the BN operation is performed. The parameters of the data are updated by group. A group of data jointly determines the direction of the gradient, thereby reducing the randomness of the descent. The details are as follows:

[0098]

[0099] Step 4-3, in channel C2, Based on the IN strategy, the data mean and standard deviation of each sample in the H and W dimensions are solved, and the N and C dimensions are retained, that is, the mean and standard deviation are only calculated within the channel, as follows:

[0100]

[0101] Step 4-4, recombining the normalized sub-channel results to form the final output Z:

[0102] Z1=[Y B，1 , Y I，1 , Y B，2 , Y I，2 ,......,Y B，C / / 2 ,Y I，C / / 2 ] (14)

[0103] Among them, BN is batch normalization, IN is instance normalization, is the eigenvalue of each sample in channel C1 in dimensions N, H, and W, is the characteristic value of each sample in channel C2 in the H and W dimensions, γc and βc are learnable parameters, and ε is a small constant, which is 10 here. -5 or 10 -8 , [.] indicates concatenation along the channel dimension.

[0104] Step 5: Build a multi-dimensional spatio-temporal feature fusion module, input the normalized results Z1, Z2, ..., Z5 into the multi-dimensional spatio-temporal feature fusion module, realize the adaptive extraction of important features in channels and spaces, and obtain multi-scale features MF1, MF2, ..., MF5. Take feature Z1 as the input of the multi-dimensional spatio-temporal feature fusion module as an example, the details are as follows:

[0105] Step 5-1: For a given input feature Z1, perform maximum pooling and average pooling operations to obtain two features Im and Ia, then concatenate Im and Ia and input them into the channel attention branch and spatial attention branch respectively.

[0106] Im=max(Z1) (15)

[0107]

[0108] Among them, Sc represents the input of the channel attention branch, and Ss represents the input of the spatial attention branch.

[0109] Step 5-2: In the channel attention branch, the feature Sc is input into the channel weight generation unit. First, two one-dimensional convolution layers with convolution kernel sizes of k1 and k2 are adaptively determined by inputting the number of channels. Then, the feature Sc is input into the above two one-dimensional convolution layers respectively to adaptively extract the deep features of Sc. Finally, after passing through the Softmax layer, its attention weights Wc1 and Wc2 in the channel branch are obtained. The details are as follows:

[0110] Wc1=σ[Conv1(Sc,max(1,[(γ1×C+b1) / 2]))](18)

[0111] Wc2=σ[Conv2(Sc,max(1,[(γ2×C+b2) / 2]))](19)

[0112] Among them, γ and b are hyperparameters, usually set to γ ​​= 2 and b = 1. [.] represents an odd number rounded up, Conv1 and Conv2 are one-dimensional 1*1 convolutions, and σ is the Softmax activation function.

[0113] Step 5-3: In the spatial attention branch, similarly, the feature Ss is input into the spatial weight generation unit. First, the feature Ss passes through two 7*7 convolutional layers to obtain the initial features Ss1 and Ss2 in the spatial branch. s1' Then, the linear layer FC1, Relu activation function and linear layer FC2 are used to realize the dimension increase and dimension reduction of the features, and the important features Ss2 and Ss2 of the samples in the spatial branch are obtained. s2' Finally, the attention weights Ws1 and Ws2 in the spatial branch are obtained through the Softmax activation function. The details are as follows:

[0114] Ws1=σ(FC2(η(FC1(Conv3(Ss)))))(20)

[0115] Ws2=σ(FC2(η(FC1(Conv4(Ss)))))(21)

[0116] Among them, Conv3 and Conv4 are 7*7 convolutions.

[0117] Step 5-4: Summarize the channel weights and spatial weights to determine the important parts of the features between channels and spaces, as follows:

[0118] MF1 1 =(Wc1+Ws1)×Im+(Wc2+Ws2)×Ia(22)

[0119] Complete the feature extraction of Layer1 and obtain feature MF1 1 , then feature MF1 1 Based on steps 3 to 5, the multi-scale feature MF1 is obtained through layer2, Layer3, ... Layer5 layers in sequence.

[0120] Among them, Output represents the fused feature. Since the sum of the feature weights of Im and Ia is equal to 1, the useful part between the features Im and Ia is retained, and the useless part is discarded, thereby achieving effective feature fusion.

[0121] Step 6: Cascade the multi-scale features MF1, MF2…, MF5 along the channel, confirm the final fusion feature MF, and build a bearing fault classifier. Input the fusion feature MF obtained in step 6 into the fault classifier to realize fault classification in the source domain. The details are as follows:

[0122] MF=Concate[MF1, MF2, MF3, MF4, MF5](23)

[0123] OutPut = σ(Fc(MF))(24)

[0124] Combined with the cross entropy loss function to achieve bearing fault classification:

[0125] L=-(ylog(OutPut)+(1-y)log(1-OutPut))(25)

[0126] Among them, Fc is the fully connected layer, σ is the Softmax activation function, Output is the output classification result of the model, L is the classification loss, and y is the fault label in the source domain.

[0127] In step 7, the overall network model is repeatedly trained based on steps 3-6 to obtain a trained overall network model, and the test set of the target domain is input into the trained overall network model to obtain the test set accuracy and feature visualization results. The details are as follows:

[0128] Step 7-1, select the learning rate, optimizer, training batch size and number of training rounds for the overall network model;

[0129] Step 7-2, input the training data set obtained in step 2 into the set bearing fault diagnosis model to obtain a trained cross-bearing fault diagnosis model;

[0130] Step 7-3: Input the test data under different working conditions in the target domain into the trained cross-bearing fault diagnosis model to realize cross-working condition fault diagnosis of the bearing.

[0131] A system for implementing a fault diagnosis method for a hoist cross-bearing includes a domain-invariant feature extraction module, a multi-dimensional spatiotemporal feature fusion module, and a bearing fault classification module; the domain-invariant feature extraction module includes a multi-scale convolution layer and a collaborative normalization layer, as follows:

[0132] The multi-scale convolution layer consists of small convolution layers with different convolution kernel sizes and a wide convolution layer to obtain the feature relationship between samples of different scales and the long-term features of all input samples;

[0133] The co-normalization layer learns the statistical characteristics between channels and the independent characteristics between instances at the same time, and reduces covariate shift and overfitting by normalizing the mixed batch data and single sample data;

[0134] The multi-dimensional spatiotemporal feature fusion module designs a channel weight generation unit in its channel dimension to adaptively capture local cross-channel information interaction; and designs a spatial weight generation unit in its spatial branch to adaptively learn the importance of different features in the space;

[0135] The bearing fault classifier consists of a fully connected layer and a Softmax activation function. The fused convolutional features are input into the bearing fault classifier to output the classification results of the bearing fault.

[0136] Example 1

[0137] The present invention provides a hoist span bearing fault diagnosis method, which, when used in practice, comprises the following steps:

[0138] Step 1: Build a multi-dimensional spatiotemporal collaborative neural network, including a domain-invariant feature extraction module, a multi-dimensional spatiotemporal feature fusion module, and a bearing fault classification module; the domain-invariant feature extraction module includes a multi-scale convolution layer and a collaborative normalization layer;

[0139] Step 2: Collect vibration signal data under different working conditions. After preprocessing, obtain source domain and target domain data sets respectively. Establish a bearing database for each working condition and divide the database into a training set and a test set. The ratio of the number of bearings in the training set and the test set is 8:2, as follows:

[0140] Based on the public PU bearing dataset, bearing datasets under 4 different working conditions are collected. 200 and 50 sample data are taken for each working condition respectively. Sample databases in the source domain and target domain are established, and the database is divided into training set and test set. The sample ratio of training set and test set is 8:2.

[0141] Step 3: The multi-scale convolution layer contains 5 convolution layers with different convolution kernel sizes and 1 wide convolution layer with a convolution kernel size of 64, as follows:

[0142] On the PU dataset, the initial input feature signal X passes through the multi-scale convolution layer to obtain the multi-scale convolution features X1, X2…, X5 of X, that is, the convolution features under different convolution kernel sizes:

[0143] X1=conv1(conv(X)) (1)

[0144] X2=conv2(conv(X)) (2)

[0145] X3=conv3(conv(X)) (3)

[0146] X4=conv4(conv(X)) (4)

[0147] X5=conv5(conv(X)) (5)

[0148] Among them, X is the input feature signal, co is a wide convolutional layer, conv1, conv2, conv3, conv4, conv are convolutional layers of different scales;

[0149] Step 4: Build a collaborative normalization layer, input the multi-scale convolutional features X1, X2…, X5 obtained in step 3 into the collaborative normalization layer, and obtain the normalized results Z1, Z2…, Z5 of the features. Take the multi-scale convolutional feature X1 as the input of the collaborative normalization layer as an example, as follows:

[0150] Step 4-1: Split each channel equally into two sub-channels C1 and C2. The details are as follows:

[0151] X1 c1 =X1[:,0:C / / 2,:,:](6)

[0152] X1 c2 =X1[:,C / / 2:c,:,:](7)

[0153] Step 4-2, in channel C1, Divide the data into several groups and perform BN operation, update the parameters of the data by group, and a group of data jointly determines the direction of the gradient, thereby reducing the randomness of the descent. The details are as follows:

[0154]

[0155]

[0156] Step 4-3, in channel C2, Based on the IN strategy, the mean and standard deviation of each sample in the H and W dimensions are solved, and the N and C dimensions are retained, that is, the mean and standard deviation are only calculated within the channel, as follows:

[0157]

[0158] Step 4-4, recombining the normalized sub-channel results to obtain the final output Z:

[0159] Z1=[Y B，1 , Y I，1 , Y B，2 , Y I，2 ,......,Y B,C / / 2 ,Y I,C / / 2 ] (14)

[0160] Among them, BN is batch normalization, IN is instance normalization, is the eigenvalue of each sample in channel C1 in dimensions N, H, and W, is the characteristic value of each sample in channel C2 in the H and W dimensions, γc and βc are learnable parameters, and ε is a small constant, which is 10 here. -5 or 10 -8 , [.] indicates concatenation along the channel dimension.

[0161] Step 5: Build a multi-dimensional spatio-temporal feature fusion module, input the normalized results Z1, Z2…, Z5 into the multi-dimensional spatio-temporal feature fusion module, realize the adaptive extraction of important features in channels and spaces, and obtain multi-scale features MF1, MF2…MF5. Take feature Z1 as the input of the multi-dimensional spatio-temporal feature fusion module as an example, as follows:

[0162] Step 5-1: For a given input feature Z1, perform maximum pooling and average pooling operations to obtain two features Im and Ia, then concatenate Im and Ia and input them into the channel attention branch and spatial attention branch respectively.

[0163] Im=max(Z1)(15)

[0164]

[0165] Among them, Sc represents the input of the channel attention branch, and Ss represents the input of the spatial attention branch.

[0166] Step 5-2: In the channel attention branch, the feature Sc is input into the channel weight generation unit. First, two one-dimensional convolution layers with convolution kernel sizes of k1 and k2 are adaptively determined by inputting the number of channels. Then, the feature Sc is input into the above two one-dimensional convolution layers respectively to adaptively extract the deep features of Sc. Finally, after passing through the Softmax layer, its attention weights Wc1 and Wc2 in the channel branch are obtained. The details are as follows:

[0167] Wc1=σ[Conv1(Sc,max(1,[(γ1×C+b1) / 2]))](18)

[0168] Wc2=σ[Conv2(Sc,max(1,[(γ2×C+b2) / 2]))](19)

[0169] Among them, γ and b are hyperparameters, usually set to γ ​​= 2 and b = 1. [.] represents an odd number rounded up, Conv1 and Conv2 are one-dimensional 1*1 convolutions, and σ is the Softmax activation function.

[0170] Step 5-3: In the spatial attention branch, similarly, the feature Ss is input into the spatial weight generation unit. First, the feature Ss passes through two 7*7 convolutional layers to obtain the initial features Ss1 and Ss2 in the spatial branch. s1' Then, the linear layer FC1, Relu activation function and linear layer FC2 are used to realize the dimension increase and dimension reduction of the features, and the important features Ss2 and Ss2 of the samples in the spatial branch are obtained. s2'Finally, the attention weights Ws1 and Ws2 in the spatial branch are obtained through the Softmax activation function. The details are as follows:

[0171] Ws1=σ(FC2(η(FC1(Conv3(Ss)))))(20)

[0172] Ws2=σ(FC2(η(FC1(Conv4(Ss)))))(21)

[0173] Among them, Conv3 and Conv4 are 7*7 convolutions.

[0174] Step 5-4: Summarize the channel weights and spatial weights to determine the important parts of the features between channels and spaces, as follows:

[0175] MF1=(Wc1+Ws1)×Im+(Wc2+Ws2)×Ia(22)

[0176] Complete the feature extraction of Layer1 and obtain feature MF1 1 , then feature MF1 1 Based on steps 3 to 5, the multi-scale feature MF1 is obtained through layer2, Layer3, ... Layer5 layers in sequence.

[0177] Among them, Output represents the fused feature. Since the sum of the feature weights of Im and Ia is equal to 1, the useful part between the features Im and Ia is retained, and the useless part is discarded, thereby achieving effective feature fusion.

[0178] Step 6: Cascade the multi-scale features MF1, MF2…, MF5 along the channel, confirm the final fusion feature MF, and build a bearing fault classifier. Input the fusion feature MF obtained in step 6 into the fault classifier to realize fault classification in the source domain. The details are as follows:

[0179] MF=Concate[MF1, MF2, MF3, MF4, MF5](23)

[0180] OutPut = σ(Fc(MF))(24)

[0181] Combined with the cross entropy loss function to achieve bearing fault classification:

[0182] L=-(ylog(OutPut)+(1-y)log(1-OutPut))(25)

[0183] Among them, Fc is the fully connected layer, σ is the Softmax activation function, Output is the output classification result of the model, L is the classification loss, and y is the fault label in the source domain.

[0184] In step 7, the overall network model is repeatedly trained based on steps 3-6 to obtain a trained overall network model, and the test set of the target domain is input into the trained overall network model to obtain the test set accuracy and feature visualization results. The details are as follows:

[0185] Step 7-1, select the learning rate, optimizer, training batch size and number of training rounds for the overall network model;

[0186] Step 7-2, input the training data set obtained in step 2 into the set bearing fault diagnosis model to obtain a trained cross-bearing fault diagnosis model;

[0187] Step 7-3: Input the test data under different working conditions in the target domain into the trained cross-bearing fault diagnosis model to realize cross-working condition fault diagnosis of the bearing.

[0188] The method proposed in the present invention uses the python programming language and pytorch framework to build a network framework on an Intel(R) Core(TM) i3-7100 CPU@3.90GHz host to conduct relevant experiments. The learning rate of the training network is set to 1×e -4 , the optimization strategy selects Momentum, uses the Adam optimization algorithm, its impulse coefficient weight decay = 0.005, the batch-size is set to 64, and the learning rate decay factor is 0.2.

[0189] In order to demonstrate the superior performance of the algorithm of the present invention, the present invention selects some popular CNN-based bearing cross-operating fault diagnosis algorithms in recent years as comparison models. Taking the publicly available “PU dataset” of Padern University in Germany as an example, the comparison experimental results are shown in Table 1.

[0190] Table 1 Comparative experimental results of different methods on PU bearing database

[0191]

[0192] From the experimental results, it can be seen that the average accuracy of the method of the present invention reaches 98.74%, which exceeds the other five mainstream domain generalization models. Compared with other methods, the method of the present invention has better generalization ability, more stable results of cross-condition learning, and stronger deep feature extraction and global feature capture capabilities compared to other network models, especially in the generalization learning of working condition 0 to 1, the accuracy rate reaches 99.81%.

[0193] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for diagnosing a hoist span bearing fault, characterized in that: The following steps are involved: Step 1: Build a multi-dimensional spatiotemporal collaborative neural network, including a domain-invariant feature extraction module, a multi-dimensional spatiotemporal feature fusion module, and a bearing fault classification module; wherein the domain-invariant feature extraction module includes a multi-scale convolution layer and a collaborative normalization layer, and then proceed to step 2; Step 2: Collect vibration signal data under different working conditions, and obtain source domain and target domain data sets after preprocessing; establish a bearing database for each working condition in the source domain, and divide the bearing database into a training set and a test set, and then proceed to step 3: Step 3: Build a multi-scale convolution layer. The initial input feature signal X passes through the multi-scale convolution layer to obtain the multi-scale convolution features X1, X2, ..., X5 of X, and then proceed to step 4; Step 4: Build a collaborative normalization layer, input the multi-scale convolution features X1, X2, ..., X5 into the collaborative normalization layer, obtain the normalized results Z1, Z2, ..., Z5 of the features, and proceed to step 5; Step 5: Build a multi-dimensional spatiotemporal feature fusion module, input the normalized results Z1, Z2, ..., Z5 into the multi-dimensional spatiotemporal feature fusion module, realize the adaptive extraction of important features in channels and spaces, complete the feature extraction of Layer1, and then obtain multi-scale features MF1, MF2..., MF5 through Layer2, Layer3, ... Layer5 based on steps 3 to 5, and proceed to step 6; Step 6: Implement the cascade of multi-scale features MF1, MF2, ..., MF5 along the channel, confirm the final fusion feature MF, and build a bearing fault classification module. Input the fusion feature MF into the fault classification module to realize the bearing fault classification in the source domain, and then proceed to step 7; Step 7: Repeat the training of the overall network model based on steps 3 to 6 to obtain a trained overall network model, and input the test set of the target domain into the trained overall network model to obtain the test set accuracy.

2. A hoist span bearing fault diagnosis method according to claim 1, characterized in that: In step 2, vibration signal data of different working conditions are collected and preprocessed to obtain source domain and target domain data sets respectively; a bearing database for each working condition in the source domain is established, and the bearing database is divided into a training set and a test set, specifically: Based on the public PU bearing dataset, bearing datasets under multiple different working conditions are collected. For each working condition, m sample data are taken to establish sample databases in the source domain and target domain, and the sample database is divided into a training set and a test set.

3. A method for diagnosing faults of hoist span bearings according to claim 1, characterized in that: In step 3, the multi-scale convolution layer includes 5 convolution layers with different convolution kernel sizes and 1 wide convolution layer with a convolution kernel size of 64, specifically: On the PU dataset, the initial input feature signal X passes through the multi-scale convolution layer to obtain the multi-scale convolution features X1, X2, ..., X5 of X, that is, the convolution features under different convolution kernel sizes: X1 = conv1(conv(X)); X2 = conv2(conv(X)); X3 = conv3(conv(X)); X4 = conv4(conv(X)); X5 = conv5(conv(X)); Where X is the input feature signal, conv is a wide convolutional layer, and conv1, conv2, conv3, conv4, and conv5 are convolutional layers of different scales.

4. A method for diagnosing a hoist span bearing fault according to claim 1, characterized in that: In step 4, a collaborative normalization layer is constructed, and the multi-scale convolutional features X1, X2, ..., X5 obtained in step 3 are input to the collaborative normalization layer to obtain the normalized results Z1, Z2, ..., Z5 of the features. Taking the multi-scale convolutional feature X1 as the input of the collaborative normalization layer as an example, the specific steps are as follows: Step 4-1: Split each channel equally into two sub-channels C1 and C2, which can be achieved by: X1 c1 =X1[:,0:C / / 2,:,:]; X1 c2 =X1[:,C / / 2:c,:,:]; Step 4-2, in channel C1, Divide the data into several groups and perform BN operations, and update the parameters of the data by group; specifically: Step 4-3, in channel C2, Based on the IN strategy, the mean and standard deviation of each sample in the H and W dimensions are solved, and the N and C dimensions are retained, that is, the mean and standard deviation are only calculated within the channel, specifically: Step 4-4, recombining the normalized sub-channel results to obtain the final output Z: Z1=[Y B，1 ,Y I，1 ,Y B，2 ,Y I，2 ,......,Y B，C / / 2 ,Y I，C / / 2 ]; Among them, BN is batch normalization, IN is instance normalization, is the eigenvalue of each sample in channel C1 in dimensions N, H, and W, is the characteristic value of each sample in channel C2 in the H and W dimensions, and are all learnable parameters, ε is a constant, and [.] represents concatenation along the channel dimension.

5. A method for diagnosing faults of hoist span bearings according to claim 1, characterized in that: In step 5, a multi-dimensional spatiotemporal feature fusion module is built, and the normalized results Z1, Z2, ..., Z5 are input into the multi-dimensional spatiotemporal feature fusion module to realize the adaptive extraction of important features in channels and spaces, complete the feature extraction of Layer1, and then obtain multi-scale features MF1, MF2..., MF5 through layer2, Layer3, ... Layer5 in sequence based on steps 3 to 5. Taking the normalized result Z1 as the input of the multi-dimensional spatiotemporal feature fusion module as an example, specifically: Step 5-1: For a given input feature Z1, perform maximum pooling and average pooling operations to obtain two features Im and Ia, then concatenate Im and Ia and input them into the channel attention branch and the spatial attention branch respectively; Im=max(Z1); Step 5-2: In the channel attention branch, the feature Sc is input into the channel weight generation unit, and two one-dimensional convolution layers with convolution kernel sizes of k1 and k2 are adaptively determined by the number of input channels. Then, the feature Sc is input into the above two one-dimensional convolution layers respectively, and the deep features of Sc are adaptively extracted. Finally, after passing through the Softmax layer, its attention weights Wc1 and Wc2 in the channel branch are obtained; Wc1=σ[Conv1(Sc,max(1,[(γ1×C+b1) / 2]))]; Wc2=σ[Conv2(Sc,max(1,[(γ2×C+b2) / 2]))]; Among them, γ and b are hyperparameters, [.] represents an odd number rounded up, Conv1 and Conv2 are one-dimensional 1*1 convolutions, and σ is the Softmax activation function; Step 5-3: In the spatial attention branch, the feature Ss is input into the spatial weight generation unit. The feature Ss passes through two 7*7 convolutional layers to obtain the initial feature S in the spatial branch. s1 and S s1' Then, the linear layer FC1, Relu activation function and linear layer FC2 are used to realize the dimension increase and dimension reduction of the features, and the important features S of the samples in the spatial branch are obtained. s2 and S s2' , and finally the attention weights Ws1 and Ws2 in the spatial branch are obtained through the Softmax activation function; Ws1=σ(FC2(η(FC1(Conv3(Ss))))); Ws2=σ(FC2(η(FC1(Conv4(Ss))))); Among them, Conv3 and Conv4 are 7*7 convolutions; Step 5-4: Summarize the channel weights and spatial weights to determine the important parts of the features between channels and spaces, specifically: MF1 1 =(Wc1+Ws1)×Im+(Wc2+Ws2)×Ia; Complete the feature extraction of Layer1 and obtain feature MF1 1 , then feature MF1 1 Based on steps 3 to 5, the multi-scale feature MF1 is obtained through layer2, Layer3, ... Layer5 layers in sequence.

6. A method for diagnosing faults of hoist span bearings according to claim 1, characterized in that: In step 6, the multi-scale features MF1, MF2, ..., MF5 are cascaded along the channel to confirm the final fusion feature MF, and a bearing fault classifier is built. The fusion feature MF obtained in step 6 is input into the fault classifier to realize fault classification in the source domain, specifically: MF=Concate[MF1, MF2, MF3, MF4, MF5]; OutPut = σ(Fc(MF)); Combined with the cross entropy loss function to achieve bearing fault classification: L=-(ylog(OutPut)+(1-y)log(1-OutPut)); Among them, Fc is the fully connected layer, σ is the Softmax activation function, Output is the output classification result of the model, L is the classification loss, and y is the fault label in the source domain.

7. A method for diagnosing faults of hoist span bearings according to claim 1, characterized in that: In step 7, the overall network model is repeatedly trained based on steps 3 to 6 to obtain a trained overall network model, and the test set of the target domain is input into the trained overall network model to obtain the test set accuracy and feature visualization results, specifically: Step 7-1, select the learning rate, optimizer, training batch size and number of training rounds for the overall network model; Step 7-2, input the training data set obtained in step 2 into the set bearing fault diagnosis model to obtain a trained cross-bearing fault diagnosis model; Step 7-3: Input the test data under different working conditions in the target domain into the trained cross-bearing fault diagnosis model to realize cross-working condition fault diagnosis of the bearing.

8. A hoist span bearing fault diagnosis system, characterized in that A hoist cross-bearing fault diagnosis method applied to any one of claims 1 to 7, comprising a domain invariant feature extraction module, a multi-dimensional spatiotemporal feature fusion module and a bearing fault classification module; The domain invariant feature extraction module is composed of a multi-scale convolution layer and a collaborative normalization layer, which is used to realize the adaptive extraction and fusion of important features of input samples; the multi-scale convolution layer is composed of a small convolution layer with different convolution kernel sizes and a wide convolution layer, which obtains the feature relationship between samples of different scales and the long-term relationship features of all input samples; The co-normalization layer simultaneously learns the statistical characteristics between channels and the independent characteristics between instances, and reduces covariate shift and overfitting by normalizing the mixed batch data and single sample data; The multi-dimensional spatiotemporal feature fusion module innovatively replaces the traditional (maximum / average) pooling module and designs a channel weight generation unit in its channel dimension to adaptively capture local cross-channel information interaction; A spatial weight generation unit is designed in its spatial branch to adaptively learn the importance of different features in the space; The bearing fault classification module consists of a fully connected layer and a Softmax activation function, which is used to input the fused convolutional features into the bearing fault classifier and output the classification result of the bearing fault.

Images