Blower bearing fault diagnosis method based on multi-source feature fusion migration model

Abstract

In order to solve the problem that the blower bearing fault is difficult to diagnose accurately in the urban sewage treatment process, the invention designs a blower bearing fault diagnosis method based on a multi-source feature fusion migration model, and the method comprises the following steps: firstly, establishing a fault diagnosis model based on a multi-source feature fusion strategy; torque and vibration data features under different working conditions are extracted, and fusion features under different working conditions are obtained in combination with a multi-source feature fusion strategy. Secondly, a fault diagnosis model parameter transfer learning algorithm is designed, and the accuracy of model sub-classification diagnosis is guaranteed; and finally, on-line diagnosis of the blower bearing fault model is realized in combination with actually measured data and the fault diagnosis model. The result shows that the model can quickly and accurately carry out fault diagnosis on the blower bearing, so that a foundation is laid for safe and stable operation of the urban sewage treatment blower.

Description

Technical Field

[0001] The present invention designs a blower bearing fault diagnosis method based on a multi-source feature fusion migration model, which realizes accurate diagnosis of blower bearing faults during sewage treatment. It can effectively solve the problem of single data source and lack of available labeled samples for target working conditions during equipment maintenance, making it difficult to perform accurate fault diagnosis. Therefore, it is of great significance for blower bearing fault diagnosis and maintenance. The operating status of blower bearings in sewage treatment process is an important basis for early warning of equipment failures, providing important guidance for predictive maintenance of equipment. Background Art

[0002] With the continuous improvement of my country's industrialization level and the continuous enhancement of environmental awareness, sewage treatment automation technology has ushered in unprecedented development opportunities. In the sewage treatment process, it is extremely important to study the fault diagnosis of blower bearings, because bearings are one of the core components of sewage treatment equipment, and their health status directly affects the operating efficiency and safety of the entire equipment. Bearing failures often lead to unstable equipment operation, reduced energy efficiency, and may even cause serious mechanical accidents and downtime, all of which will bring high maintenance costs and production losses. First, bearing fault diagnosis can help sewage treatment facilities to achieve early identification of faults. Through continuous monitoring and data analysis of bearings, initial abnormal signals can be detected before the fault develops to a serious level, so that timely repairs or replacements can be carried out, avoiding large-scale equipment damage and production interruptions. Secondly, the development of fault diagnosis technology has promoted the transformation of the sewage treatment industry to intelligence and automation. Using advanced sensing technology, the Internet of Things and artificial intelligence methods, such as machine learning and deep learning algorithms, a large amount of monitoring data can be deeply analyzed to improve the accuracy and efficiency of diagnosis. Therefore, the present invention is of great significance in the accurate prediction of the operating status of blower bearings and predictive maintenance of equipment.

[0003] The present invention designs a bearing fault diagnosis method based on a multi-source feature fusion migration model. The method mainly establishes a fault diagnosis model based on a multi-source feature fusion strategy, extracts features of the torque and vibration signals of the bearing respectively, obtains a comprehensive feature representation through a multi-source feature fusion strategy, and uses a transfer learning algorithm to update the classifier parameters and the fusion parameters in the target working conditions, thereby achieving accurate prediction of the bearing operating status. Summary of the invention

[0004] The present invention obtains a blower bearing fault diagnosis method based on a multi-source feature fusion migration model. The method establishes a fault diagnosis model based on a multi-source feature fusion strategy to extract torque and vibration data features under different working conditions, and designs a multi-source feature fusion strategy to overcome the problem of a single data source, thereby obtaining fusion features on the source working condition and the target working condition respectively. Finally, a transfer learning algorithm is used to update the classifier parameters, thereby achieving accurate diagnosis of blower bearing faults.

[0005] The present invention adopts the following technical solutions and implementation steps:

[0006] A blower bearing fault diagnosis method based on a multi-source feature fusion migration model is characterized by: collecting blower bearing data, establishing a fault diagnosis model based on a multi-source feature fusion strategy, designing a classifier parameter update method based on transfer learning, and diagnosing blower bearing faults, including the following steps:

[0007] 1. Collect blower bearing data:

[0008] Taking the blower bearing as the research object, the torque data and vibration data are determined as the characteristic variables for bearing fault diagnosis. The source working condition torque data at time t is collected as the matrix N(t) = [n1(t), n2(t),…, n k (t)] T , T represents matrix transpose, the value range of matrix row number k is [10, 50], and the torque data of the kth row is vector n k (t) = [n k,1 (t),n k,2 (t),…,n k,q (t)], where n k,1 (t) is the value of the first column of the kth row of the matrix. The value range of the column number q of each row vector is [50, 100], and the value range of each value is [-1Nm, 1Nm]. N is Newton, m is meter, and the source working condition vibration data at time t is the matrix V(t) = [v1(t), v2(t), …, v r (t)] T , the value interval of the matrix row number r is [10, 50], and the vibration data of the rth row is the vector v r (t) = [v r,1 (t),v r,2 (t),…,v r,q (t)], where v r,1 (t) is the value of the rth row and the first column of the matrix, and each value ranges from [-1m / s 2 , 1m / s 2 ], s is seconds, and the target torque data collected at time t is the matrix N′(t)=[n′1(t),n′2(t),…,n ′w (t)]T , the value range of the matrix row number w is [10, 50], and the torque data of the wth row is vector n ′w (t) = [n′ w,1 (t),n ′w,2 (t),…,n ′w,q (t)], where n′ w,1 (t) is the value of the wth row and the first column of the matrix. The range of each value is [-1Nm, 1Nm]. The target working condition vibration data at time t is the matrix V′(t)=[v′1(t),v′2(t),…,v′ u (t)] T , the value interval of the matrix row number u is [10, 50], and the vibration data of the uth row is the vector v′ u (t) = [v′ u,1 (t),v′ u,2 (t),…,v ′u,q (t)], where v′ u,1 (t) is the value of the first column of the u-th row of the matrix, and each value ranges from [-1m / s 2 , 1m / s 2 ], the real fault label vector L = [L1,L2,…,L f ] T , the number of values ​​f in the vector L ranges from [10, 50], and each value is fixed to 0 or 1;

[0009] 2. Establish a fault diagnosis model based on multi-source feature fusion strategy:

[0010] The fault diagnosis model based on multi-source feature fusion strategy includes a fusion feature extraction network and a classifier.

[0011] 1) Fusion feature extraction network: The model includes data input layer, feature extraction layer, feature compression layer and fusion layer. The feature extraction layer has 32 convolution kernels, each with a size of 4. The first weight matrix in the feature extraction layer is expressed as is the weight vector corresponding to the i-th convolution kernel at time t, and the first bias vector in the feature extraction layer is expressed as is the bias value corresponding to the i-th convolution kernel at time t, and W is the initial value. 1 The internal value of (1) is 1, and the initial moment b 1 (1) The internal values ​​are all 0;

[0012] Data input layer: This layer has j neurons, which are used to input the blower bearing data matrix into the model row by row. The output of the data input layer is the source torque vector n1(t), n2(t),…, n k (t), source condition vibration vector v1(t), v2(t),…,vr (t), target torque vector n′1(t), n′2(t),…, n ′w (t) and the target operating condition vibration vector V′(t)=[v′1(t),v′2(t),…,v′ u (t)] T ;

[0013] Feature extraction layer: This layer uses convolution kernels to automatically extract local features. The output of the feature extraction layer is expressed as:

[0014]

[0015] Among them, max(0,·) means taking the maximum value compared with 0, z i (t) is the source torque feature vector generated by the i-th convolution kernel at time t, z i ′(t) is the target torque feature vector generated by the i-th convolution kernel at time t, e i (t) is the source vibration feature vector generated by the i-th convolution kernel at time t, e i ′(t) is the target operating condition vibration feature vector generated by the i-th convolution kernel at time t;

[0016] Feature compression layer: This layer compresses the feature vector to get the maximum value. The output of the feature compression layer is expressed as:

[0017] c i (t)=max(0,z i (t)) (5)

[0018] c i ′(t)=max(0,z′ i (t)) (6)

[0019] p i (t)=max(0,e i (t)) (7)

[0020] p i ′(t)=max(0,e′ i (t)) (8)

[0021] Among them, c i (t) is the compressed eigenvalue of the torque eigenvector of the i-th source condition at time t, c ′i (t) is the compressed eigenvalue of the torque eigenvector of the i-th target operating condition at time t, p i (t) is the eigenvalue of the compressed vibration eigenvector of the ith source condition at time t, p ′i (t) is the eigenvalue of the compressed vibration eigenvector of the i-th target operating condition at time t;

[0022] Fusion layer: This layer fuses the torque features and vibration features on the source and target working conditions. The output of the fusion layer is:

[0023] h i (t) = α(t) × c i (t)+(1-α(t))×p i (t) (9)

[0024] h i ′(t)=β(t)×c i ′(t)+(1-β(t))×p i ′(t) (10)

[0025] Among them, h i (t) is the fusion eigenvalue of the i-th source condition at time t. The fusion eigenvector of the source condition is h(t) = [h1(t), h2(t), …, h i (t)],h ′i (t) is the fusion feature value of the i-th target working condition at time t, and the fusion feature vector of the target working condition is h′(t)=[h′1(t),h′2(t),…,h ′i (t)], α(t) is the fusion parameter between the torque characteristic and the vibration characteristic of the source working condition at time t, and the value range of α(t) is [0,1], β(t) is the fusion parameter between the torque characteristic and the vibration characteristic of the target working condition at time t, and the value range of β(t) is [0,1];

[0026] 2) Classifier: It consists of a fully connected layer and a classification layer. The fully connected layer has 32 input neurons, 2 output neurons, and 2 classification layer neurons. The second weight matrix of the fully connected layer is expressed as in and is the weight vector corresponding to the two output neurons at time t, and the second bias vector of the fully connected layer is expressed as in and b 22 (t) is the bias corresponding to the two output neurons at time t, and W is the initial time. 2 (1) The internal elements are all 1, initialize b 2 (1) All internal elements are 0;

[0027] Fully connected layer: This layer integrates the fusion features. The output of the fully connected layer is expressed as:

[0028]

[0029] Among them, y g is the output of the g-th output neuron in the fully connected layer, g = 1, 2;

[0030] Classification layer: This layer converts the output of the gth neuron in the fully connected layer into a predicted probability P g , the output of the classification layer is expressed as:

[0031]

[0032] 3. Design a classifier parameter update method based on transfer learning:

[0033] Algorithm 1: α(t) update strategy:

[0034] ①Define the loss function J(t) at time t as;

[0035] J(t)=-[L t log(P g )+(1-L t )log(1-P g )](13)

[0036] Among them, L t is the corresponding true label, the value is 0 or 1, and log is the natural logarithm;

[0037] ② Input the fusion feature vector h(t) and the true label vector L training parameters of the source working condition, set the current time t = 1, and initialize α(1) = 0.5;

[0038] ③Use the gradient descent algorithm to iteratively optimize the model parameters:

[0039]

[0040] Among them, W 2 (t+1) is the weight matrix at time t+1, b 2 (t+1) is the bias vector at time t+1, η(t) is the learning rate at time t, a(t+1) is the fusion parameter in the source condition at time t+1, is the correction amount of the weight matrix in the fully connected layer, is the correction amount of the bias vector in the fully connected layer, is the correction amount of the fusion parameter a(t);

[0041] ④If t<100, then t=t+1, return to step ③ to continue training. If t=100, or the value of a(t) does not change in 20 consecutive updates, the training is completed and go to step ⑤;

[0042] ⑤ Feature distance measurement: The feature distance value between the fusion features of the source condition and the target condition at time t is defined as D(t) as follows:

[0043]

[0044] Among them, |·| means taking the absolute value;

[0045] Algorithm 2: β(t) update strategy:

[0046] ⑥ Input the source working condition fusion feature vector h(t) and the target working condition fusion feature vector h′(t) to update the parameters, reset the time t=1, and initialize the fusion parameter β(1)=0.5 in the target working condition;

[0047] ⑦The update rule of β(t) is as follows:

[0048]

[0049] Among them, β(t+1) is the fusion parameter in the target working condition at time t+1, is the correction value of the fusion parameter β(t);

[0050] ⑧If t<100, then t=t+1, return to step ⑦ to continue training. If t=100, or the value of β(t) does not change in 20 consecutive updates, the training is completed and go to step ⑨;

[0051] Algorithm 3: Weight Matrix W 2 (t) and the bias vector b 2 (t) Update strategy:

[0052] ⑨ Combine formulas (13) and (17) to define the total loss function at time t, which is calculated as follows:

[0053] J total (t) = λ·D(t)+(1-λ)·[-(L t log(P g )+(1-L t )log(1-P g ))](19)

[0054] Among them, the trade-off parameter λ = 0.5, the reset time t = 1;

[0055] ⑩The update rules for the weight matrix and bias vector at time t+1 are:

[0056]

[0057] If t<100, then t=t+1, return to step ⑩ to continue training; if t=100, the training is completed and go to step 4;

[0058] 4. Diagnose blower bearing failure:

[0059] The trained blower bearing fault diagnosis model is used, and the torque data and vibration data of the source working condition and the target working condition are used as the input of the blower bearing fault diagnosis model. The failure rate of the bearing is calculated according to formulas (1)-(12). If P1 < 0.5, the blower bearing is normal. If P1 > 0.5, the blower bearing is faulty.

[0060] The creativity of the present invention is mainly reflected in:

[0061] (1) In view of the problems of single input data source and low diagnostic accuracy in the process of blower bearing fault diagnosis, the present invention proposes a blower bearing fault diagnosis method based on a multi-source feature fusion migration model, which realizes accurate diagnosis of blower bearing faults;

[0062] (2) The present invention establishes a fault diagnosis model based on a multi-source feature fusion strategy, extracts features from torque data and vibration data respectively, and adopts a multi-source feature fusion strategy to fuse torque features and vibration features under different working conditions to compensate for the singleness of data features. The model parameters are updated using a transfer learning algorithm, thereby improving the robustness and accuracy of the model for diagnosing equipment under different working conditions; BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a graph showing the accuracy of the present invention in diagnosing bearing faults;

[0064] Figure 2 is an error diagram for bearing fault diagnosis according to the present invention;

[0065] Figure 3 It is a confusion matrix diagram for bearing fault diagnosis according to the present invention;

[0066] Figure 4 It is a prediction result diagram of bearing fault diagnosis of the present invention, wherein the green dots represent the actual bearing operation status, and the blue circles represent the bearing operation status predicted by the model; DETAILED DESCRIPTION

[0067] The present invention obtains a blower bearing fault diagnosis method based on a multi-source feature fusion migration model; the torque data and vibration data of the blower bearing are selected as input variables for bearing fault diagnosis, and the operating state (normal or faulty) of the bearing is used as the output variable;

[0068] The experimental data comes from the daily report of equipment analysis of a certain urban sewage treatment plant; after eliminating abnormal experimental samples, 200 groups of available data remain, of which 50 groups are torque data of source working conditions, 50 groups are vibration data of source working conditions, 50 groups are torque data of target working conditions, and the remaining 50 groups are vibration data of target working conditions.

[0069] The present invention adopts the following technical solutions and implementation steps:

[0070] A blower bearing fault diagnosis method based on a multi-source feature fusion migration model is characterized by: collecting blower bearing data, establishing a fault diagnosis model based on a multi-source feature fusion strategy, designing a classifier parameter update method based on transfer learning, and diagnosing blower bearing faults, including the following steps:

[0071] 1. Collect blower bearing data:

[0072] Taking the blower bearing as the research object, the torque data and vibration data are determined as the characteristic variables for bearing fault diagnosis. The source working condition torque data at time t is collected as the matrix N(t) = [n1(t), n2(t),…, n k (t)] T , T represents matrix transpose, the value range of matrix row number k is [10, 50], and the torque data of the kth row is vector n k (t) = [n k,1 (t),n k,2 (t),…,n k,q (t)], where n k,1 (t) is the value of the first column of the kth row of the matrix. The value range of the column number q of each row vector is [50, 100], and the value range of each value is [-1Nm, 1Nm]. N is Newton, m is meter, and the source working condition vibration data at time t is the matrix V(t) = [v1(t), v2(t), …, v r (t)] T , the value interval of the matrix row number r is [10, 50], and the vibration data of the rth row is the vector v r (t) = [v r,1 (t),v r,2 (t),…,v r,q (t)], where v r,1 (t) is the value of the rth row and the first column of the matrix, and each value ranges from [-1m / s 2 , 1m / s 2 ], s is seconds, and the target torque data collected at time t is the matrix N′(t)=[n′1(t),n′2(t),…,n ′w (t)] T , the value range of the matrix row number w is [10, 50], and the torque data of the wth row is vector n ′w (t) = [n′ w,1 (t),n ′w,2 (t),…,n ′w,q (t)], where n′ w,1(t) is the value of the wth row and the first column of the matrix. The range of each value is [-1Nm, 1Nm]. The target working condition vibration data at time t is the matrix V′(t)=[v′1(t),v′2(t),…,v′ u (t)] T , the value interval of the matrix row number u is [10, 50], and the vibration data of the uth row is the vector v′ u (t) = [v′ u,1 (t),v′ u,2 (t),…,v ′u,q (t)], where v′ u,1 (t) is the value of the first column of the u-th row of the matrix, and each value ranges from [-1m / s 2 , 1m / s 2 ], the real fault label vector L = [L1,L2,…,L f ] T , the number of values ​​f in the vector L ranges from [10, 50], and each value is fixed to 0 or 1;

[0073] 2. Establish a fault diagnosis model based on multi-source feature fusion strategy:

[0074] The fault diagnosis model based on multi-source feature fusion strategy includes a fusion feature extraction network and a classifier.

[0075] 1) Fusion feature extraction network: The model includes data input layer, feature extraction layer, feature compression layer and fusion layer. The feature extraction layer has 32 convolution kernels, each with a size of 4. The first weight matrix in the feature extraction layer is expressed as is the weight vector corresponding to the i-th convolution kernel at time t, and the first bias vector in the feature extraction layer is expressed as is the bias value corresponding to the i-th convolution kernel at time t, and W is the initial value. 1 The internal value of (1) is 1, and the initial moment b 1 (1) The internal values ​​are all 0;

[0076] Data input layer: This layer has a total of j neurons, j = p, which are used to input the blower bearing data matrix into the model row by row. The output of the data input layer is the source working torque vector n1(t), n2(t),…, n k (t), source condition vibration vector v1(t), v2(t),…,v r (t), target torque vector n′1(t), n′2(t),…, n ′w (t) and the target operating condition vibration vector V′(t)=[v′1(t),v′2(t),…,v′ u (t)] T ;

[0077] Feature extraction layer: This layer uses convolution kernels to automatically extract local features. The output of the feature extraction layer is expressed as:

[0078]

[0079] Among them, max(0,·) means taking the maximum value compared with 0, z i (t) is the source torque feature vector generated by the i-th convolution kernel at time t, z i ′(t) is the target torque feature vector generated by the i-th convolution kernel at time t, e i (t) is the source vibration feature vector generated by the i-th convolution kernel at time t, e i ′(t) is the target operating condition vibration feature vector generated by the i-th convolution kernel at time t;

[0080] Feature compression layer: This layer compresses the feature vector to get the maximum value. The output of the feature compression layer is expressed as:

[0081] c i (t)=max(0,z i (t)) (26)

[0082] c i ′(t)=max(0,z′ i (t)) (27)

[0083] p i (t)=max(0,e i (t)) (28)

[0084] p i ′(t)=max(0,e′ i (t)) (29)

[0085] Among them, c i (t) is the compressed eigenvalue of the torque eigenvector of the ith source condition at time t, c′i(t) is the compressed eigenvalue of the torque eigenvector of the ith target condition at time t, p i (t) is the compressed eigenvalue of the vibration eigenvector of the ith source condition at time t, and p′i(t) is the compressed eigenvalue of the vibration eigenvector of the ith target condition at time t;

[0086] Fusion layer: This layer fuses the torque features and vibration features on the source and target working conditions. The output of the fusion layer is:

[0087] h i (t) = α(t) × c i (t)+(1-α(t))×p i (t) (30)

[0088] h i ′(t)=β(t)×c i ′(t)+(1-β(t))×p i ′(t) (31)

[0089] Among them, h i (t) is the fusion eigenvalue of the i-th source condition at time t. The fusion eigenvector of the source condition is h(t) = [h1(t), h2(t), …, h i (t)],h ′i (t) is the fusion feature value of the i-th target working condition at time t, and the fusion feature vector of the target working condition is h′(t)=[h′1(t),h′2(t),…,h ′i (t)], α(t) is the fusion parameter between the torque characteristic and the vibration characteristic of the source working condition at time t, and the value range of α(t) is [0,1], β(t) is the fusion parameter between the torque characteristic and the vibration characteristic of the target working condition at time t, and the value range of β(t) is [0,1];

[0090] 2) Classifier: It consists of a fully connected layer and a classification layer. The fully connected layer has 32 input neurons, 2 output neurons, and 2 classification layer neurons. The second weight matrix of the fully connected layer is expressed as in and is the weight vector corresponding to the two output neurons at time t, and the second bias vector of the fully connected layer is expressed as in and b 22 (t) is the bias corresponding to the two output neurons at time t, and W is the initial time. 2 (1) The internal elements are all 1, initialized All internal elements of are 0;

[0091] Fully connected layer: This layer integrates the fusion features. The output of the fully connected layer is expressed as:

[0092]

[0093] Among them, y g is the output of the g-th output neuron in the fully connected layer, g = 1, 2;

[0094] Classification layer: This layer converts the output of the gth neuron in the fully connected layer into a predicted probability P g , the output of the classification layer is expressed as:

[0095]

[0096] 3. Design a classifier parameter update method based on transfer learning:

[0097] Algorithm 1: α(t) update strategy:

[0098] ①Define the loss function J(t) at time t as;

[0099] J(t)=-[L t log(P g )+(1-L t )log(1-P g )](34)

[0100] Among them, L t is the corresponding true label, the value is 0 or 1, and log is the natural logarithm;

[0101] ② Input the fusion feature vector h(t) and the true label vector L training parameters of the source working condition, set the current time t = 1, and initialize α(1) = 0.5;

[0102] ③Use the gradient descent algorithm to iteratively optimize the model parameters:

[0103]

[0104] Among them, W 2 (t+1) is the weight matrix at time t+1, b 2 (t+1) is the bias vector at time t+1, η(t) is the learning rate at time t, a(t+1) is the fusion parameter in the source condition at time t+1, is the correction amount of the weight matrix in the fully connected layer, is the correction amount of the bias vector in the fully connected layer, is the correction amount of the fusion parameter a(t);

[0105] ④If t<100, then t=t+1, return to step ③ to continue training. If t=100, or the value of a(t) does not change in 20 consecutive updates, the training is completed and go to step ⑤;

[0106] ⑤ Feature distance measurement: The feature distance value between the fusion features of the source condition and the target condition at time t is defined as D(t) as follows:

[0107]

[0108] Among them, |·| means taking the absolute value;

[0109] Algorithm 2: β(t) update strategy:

[0110] ⑥ Input the source working condition fusion feature vector h(t) and the target working condition fusion feature vector h′(t) to update the parameters, reset the time t=1, and initialize the fusion parameter β(1)=0.5 in the target working condition;

[0111] ⑦The update rule of β(t) is as follows:

[0112]

[0113] Among them, β(t+1) is the fusion parameter in the target working condition at time t+1, is the correction value of the fusion parameter β(t);

[0114] ⑧If t<100, then t=t+1, return to step ⑦ to continue training. If t=100, or the value of β(t) does not change in 20 consecutive updates, the training is completed and go to step ⑨;

[0115] Algorithm 3: Weight Matrix W 2 (t) and the bias vector b 2 (t) Update strategy:

[0116] ⑨ Combine formulas (13) and (17) to define the total loss function at time t, which is calculated as follows:

[0117] J total (t) = λ·D(t)+(1-λ)·[-(L t log(P g )+(1-L t )log(1-P g ))](40)

[0118] Among them, the trade-off parameter λ = 0.5, the reset time t = 1;

[0119] ⑩The update rules for the weight matrix and bias vector at time t+1 are:

[0120]

[0121] If t<100, then t=t+1, return to step ⑩ to continue training; if t=100, the training is completed and go to step 4;

[0122] 4. Diagnose blower bearing failure:

[0123] The trained blower bearing fault diagnosis model is used, and the torque data and vibration data of the source working condition and the target working condition are used as the input of the blower bearing fault diagnosis model. The bearing failure rate is calculated according to formulas (22)-(33). If P1 < 0.5, the blower bearing is normal. If P1 > 0.5, the blower bearing is faulty.

Claims

1. A blower bearing fault diagnosis method based on a multi-source feature fusion migration model, characterized in that: Collect blower bearing data, establish a fault diagnosis model based on multi-source feature fusion strategy, design a classifier parameter update method based on transfer learning, and diagnose blower bearing faults, including the following steps: 1). Collect blower bearing data: Taking the blower bearing as the research object, the torque data and vibration data are determined as the characteristic variables for bearing fault diagnosis. The source working condition torque data at time t is collected as the matrix N(t) = [n1(t), n2(t),…, n k (t)] T , T represents matrix transpose, the value range of matrix row number k is [10, 50], and the torque data of the kth row is vector n k (t) = [n k,1 (t),n k,2 (t),…,n k,q (t)], where n k,1 (t) is the value of the first column of the kth row of the matrix. The value range of the column number q of each row vector is [50, 100], and the value range of each value is [-1Nm, 1Nm]. N is Newton, m is meter, and the source working condition vibration data at time t is the matrix V(t) = [v1(t), v2(t), …, v r (t)] T , the value interval of the matrix row number r is [10, 50], and the vibration data of the rth row is the vector v r (t) = [v r,1 (t),v r,2 (t),…,v r,q (t)], where v r,1 (t) is the value of the rth row and the first column of the matrix, and each value ranges from [-1m / s 2 , 1m / s 2 ], s is seconds, and the target torque data collected at time t is the matrix N′(t)=[n′1(t),n′2(t),…,n′ w (t)] T , the value range of the matrix row number w is [10, 50], and the torque data of the wth row is the vector n′ w (t) = [n′ w,1 (t),n′ w,2 (t),…,n′ w,q (t)], where n′ w,1 (t) is the value of the wth row and the first column of the matrix. The range of each value is [-1Nm, 1Nm]. The target working condition vibration data at time t is the matrix V′(t)=[v′1(t),v′2(t),…,v′ u (t)] T , the value interval of the matrix row number u is [10, 50], and the vibration data of the uth row is the vector v′ u (t) = [v′ u,1 (t),v′ u,2 (t),…,v′ u,q (t)], where v′ u,1 (t) is the value of the first column of the u-th row of the matrix, and each value ranges from [-1m / s 2 , 1m / s 2 ], the real fault label vector L = [L1,L2,…,L f ] T , the number of values ​​f in the vector L ranges from [10, 50], and each value is fixed to 0 or 1; 2). Establish a fault diagnosis model based on multi-source feature fusion strategy: The fault diagnosis model based on multi-source feature fusion strategy includes a fusion feature extraction network and a classifier. 1) Fusion feature extraction network: The model includes data input layer, feature extraction layer, feature compression layer and fusion layer. The feature extraction layer has 32 convolution kernels, each with a size of 4. The first weight matrix in the feature extraction layer is expressed as i=1,2,…,32, is the weight vector corresponding to the i-th convolution kernel at time t, and the first bias vector in the feature extraction layer is expressed as is the bias value corresponding to the i-th convolution kernel at time t, and W is the initial value. 1 The internal value of (1) is 1, and the initial moment b 1 (1) The internal values ​​are all 0; Data input layer: This layer has j neurons, which are used to input the blower bearing data matrix into the model row by row. The output of the data input layer is the source torque vector n1(t), n2(t),…, n k (t), source condition vibration vector v1(t), v2(t),…,v r (t), target torque vector n′1(t), n′2(t),…, n′ w (t) and the target operating condition vibration vector V′(t)=[v′1(t),v′2(t),…,v′ u (t)] T ; Feature extraction layer: This layer uses convolution kernels to automatically extract local features. The output of the feature extraction layer is expressed as: Among them, max(0,·) means taking the maximum value compared with 0, z i (t) is the source torque feature vector generated by the i-th convolution kernel at time t, z i ′(t) is the target torque feature vector generated by the i-th convolution kernel at time t, e i (t) is the source vibration feature vector generated by the i-th convolution kernel at time t, e i ′(t) is the target operating condition vibration feature vector generated by the i-th convolution kernel at time t; Feature compression layer: This layer compresses the feature vector to get the maximum value. The output of the feature compression layer is expressed as: c i (t)=max(0,z i (t)) (5) c i ′(t)=max(0,z′ i (t)) (6) p i (t)=max(0,e i (t)) (7) p i ′(t)=max(0,e′ i (t)) (8) Among them, c i (t) is the compressed eigenvalue of the torque eigenvector of the i-th source condition at time t, c i ′(t) is the eigenvalue of the torque eigenvector of the i-th target working condition at time t after compression, p i (t) is the eigenvalue of the compressed vibration eigenvector of the ith source condition at time t, p i ′(t) is the compressed eigenvalue of the vibration eigenvector of the i-th target operating condition at time t; Fusion layer: This layer fuses the torque features and vibration features on the source and target working conditions. The output of the fusion layer is: h i (t)=α(t)×c i (t)+(1-α(t))×p i (t) (9) h′ i (t)=β(t)×c′ i (t)+(1-β(t))×p′ i (t) (10) Among them, h i (t) is the fusion eigenvalue of the i-th source condition at time t. The fusion eigenvector of the source condition is h(t) = [h1(t), h2(t), …, h i (t)],h i ′(t) is the fusion eigenvalue of the i-th target condition at time t, and the fusion eigenvector of the target condition is h′(t)=[h′1(t),h′2(t),…,h i ′(t)], α(t) is the fusion parameter between the torque characteristic and the vibration characteristic of the source working condition at time t, and the value range of α(t) is [0,1], β(t) is the fusion parameter between the torque characteristic and the vibration characteristic of the target working condition at time t, and the value range of β(t) is [0,1]; 2) Classifier: It consists of a fully connected layer and a classification layer. The fully connected layer has 32 input neurons, 2 output neurons, and 2 classification layer neurons. The second weight matrix of the fully connected layer is expressed as in and is the weight vector corresponding to the two output neurons at time t, and the second bias vector of the fully connected layer is expressed as in and is the bias corresponding to the two output neurons at time t, and the initial time W 2 (1) The internal elements are all 1, initialize b 2 (1) All internal elements are 0; Fully connected layer: This layer integrates the fusion features. The output of the fully connected layer is expressed as: Among them, y g is the output of the g-th output neuron in the fully connected layer, g = 1, 2; Classification layer: This layer converts the output of the gth neuron in the fully connected layer into a predicted probability P g , the output of the classification layer is expressed as: 3). Design a classifier parameter update method based on transfer learning: α(t) update strategy: ①Define the loss function J(t) at time t as; J(t)=-[L t log(P g )+(1-L t )log(1-P g )] (13) Among them, L t is the corresponding true label, the value is 0 or 1, and log is the natural logarithm; ② Input the fusion feature vector h(t) and the true label vector L training parameters of the source working condition, set the current time t = 1, and initialize α(1) = 0.5; ③Use the gradient descent algorithm to iteratively optimize the model parameters: Among them, W 2 (t+1) is the weight matrix at time t+1, b 2 (t+1) is the bias vector at time t+1, η(t) is the learning rate at time t, a(t+1) is the fusion parameter in the source condition at time t+1, is the correction amount of the weight matrix in the fully connected layer, is the correction amount of the bias vector in the fully connected layer, is the correction amount of the fusion parameter a(t); ④If t<100, then t=t+1, return to step ③ to continue training. If t=100, or the value of a(t) does not change in 20 consecutive updates, the training is completed and go to step ⑤; ⑤ Feature distance measurement: The feature distance value between the fusion features of the source condition and the target condition at time t is defined as D(t) as follows: Among them, |·| means taking the absolute value; β(t) update strategy: ⑥ Input the source working condition fusion feature vector h(t) and the target working condition fusion feature vector h′(t) to update the parameters, reset the time t=1, and initialize the fusion parameter β(1)=0.5 in the target working condition; ⑦The update rule of β(t) is as follows: Among them, β(t+1) is the fusion parameter in the target working condition at time t+1, is the correction value of the fusion parameter β(t); ⑧If t<100, then t=t+1, return to step ⑦ to continue training. If t=100, or the value of β(t) does not change in 20 consecutive updates, the training is completed and go to step ⑨; Weight matrix W 2 (t) and the bias vector b 2 (t) Update strategy: ⑨ Combine formulas (13) and (17) to define the total loss function at time t, which is calculated as follows: J total (t)=λ·D(t)+(1-λ)[-(L t log(P g )+(1-L t )log(1-P g ))] (19) Among them, the trade-off parameter λ = 0.5, the reset time t = 1; ⑩The update rules for the weight matrix and bias vector at time t+1 are: If t<100, then t=t+1, return to step ⑩ to continue training; if t=100, the training is completed and go to step 4; 4). Diagnosis of blower bearing failure: The trained blower bearing fault diagnosis model is used, and the torque data and vibration data of the source working condition and the target working condition are used as the input of the blower bearing fault diagnosis model. The failure rate of the bearing is calculated according to formulas (1)-(12). If P1 < 0.5, the blower bearing is normal. If P1 > 0.5, the blower bearing is faulty.

Images