A bearing intelligent fault diagnosis method based on parallel integrated optimization and transfer learning

Abstract

The application discloses a bearing intelligent fault diagnosis method based on parallel integrated optimization and transfer learning. First, based on a signal demodulation method, an adaptive input module is constructed to automatically select an input length from an original vibration signal. Second, a TL network with low-dimensional feature reuse is constructed to realize weight and bias sharing. Third, a parallel integrated optimization loss function is proposed to align data with unknown distribution between a source domain and a target domain. Finally, the performance of the method is verified through two instances of multi-source, unsupervised and cross-domain TL. Results show that the method not only gets rid of the restriction of manual input length setting, but also overcomes the restriction of the optimization function, and is more effective than existing intelligent fault diagnosis models.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and specifically to a bearing intelligent fault diagnosis method based on parallel integrated optimization and transfer learning. Background Technology

[0002] Transfer learning (TL) is an important method for accurately identifying the health status of bearings across domains, ensuring safe mechanical operation. With further research, choosing different neural networks or optimization functions to remodel and improve fault diagnosis performance is becoming a trend. However, these methods and models have not significantly improved performance, especially as the model's generalization ability decreases when the dataset changes. Ensemble learning addresses the problem of low generalization ability. Therefore, there is an urgent need for an intelligent bearing fault diagnosis method that can solve the problem of unknown distribution differences between the source and target domains, thereby improving the generalization ability of the optimization objective. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning. First, based on the signal demodulation method, an adaptive input module is constructed to automatically select the input length from the original vibration signal. Second, a TL network with low-dimensional feature reuse is constructed to achieve weight and bias sharing. Third, a parallel ensemble optimization loss function is proposed to align data with unknown distribution between the source and target domains, thus solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning, comprising the following steps:

[0005] S1. Acquire vibration signals from the source and target domains;

[0006] S2. Adaptive Length Input Preprocessing: An adaptive input length module is constructed to preprocess the vibration signals in the source and target domains to obtain the input length of the vibration signals;

[0007] S3. Construct a CNN-based transfer learning network. Obtain parameters through pre-training with source domain samples and share training parameters with target domain samples. Use five convolutional networks and one deconvolutional network to construct a feature extraction module. Utilize dense blocks to densely link all preceding and subsequent convolutional layers, reusing low-dimensional features for feature extraction and obtaining hierarchical discrimination between different layers. Use two fully connected layers, FC1 and FC2, to construct the output layer, which represents the classifier. The fully connected layer uses Conv4 and Conv5 to extract high-level features as input. Embed multiple non-linear instance normalization (IN) and Dropout methods in the fully connected layer to alleviate overfitting. Use forward propagation to perform non-linear mapping layer by layer between each network layer. Use backpropagation to update the parameters of each layer and minimize the cross-entropy error.

[0008] S4. Construct the final optimized loss function L, which consists of parallel ensemble optimization, classification error, and trade-off function. Update the parameters through backpropagation to obtain the optimal CNN-based transfer learning network. Input the data preprocessed in step S2 into the optimal network to complete the intelligent health status identification and diagnosis of bearings.

[0009] Preferably, step S2 specifically includes the following: dividing the vibration signal into segments of length L using overlapping data segmentation technology. i Several parts of the signal are processed using an envelope spectrum-based signal demodulation method to extract low-frequency features.

[0010] The maximum value of the fault characteristic frequency f and the bandwidth B ω The relationship between them is expressed by the following formula:

[0011] B ω =t×f,

[0012] Where t∈(1,2,3,4) represents a multiple of the fault characteristic frequency;

[0013] The input length L of the vibration signal a The formula is expressed as follows:

[0014]

[0015] Where f is the maximum value of the bearing fault characteristic frequency, B ω For bandwidth, L i f is the signal length. s The sampling frequency.

[0016] Preferably, the CNN-based transfer learning network includes five convolutional modules (Conv1, Conv2, Conv3, Conv4, and Conv5), one deconvolutional module (Tconv), and two fully connected layers (FC1 and FC2); each convolutional module includes a grouped convolutional layer (GC), an instance normalization layer (IN), and an activation layer.

[0017] Preferably, in step S3, in order to ensure the consistency of data dimensions, a deconvolutional layer Tconv is used after Conv2, and dense blocks are used to combine the output of Conv1 with the output of the deconvolutional layer TConv, which together serve as the input of Conv3.

[0018] Preferably, the final optimization loss function L is expressed as follows:

[0019] L = L c +βL es +h(α);

[0020]

[0021]

[0022] Where β and α are respectively L c and L es The trade-off parameter, L c It is the classification loss function, L es It is the parallel ensemble optimization loss function, which is the sum of the loss functions of the five alignment methods.

[0023] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: a bearing intelligent fault diagnosis system based on parallel ensemble optimization and transfer learning, the system comprising the following:

[0024] The signal acquisition module acquires vibration signals from the source and target domains.

[0025] Adaptive length input preprocessing module: Construct an adaptive input length module to preprocess the vibration signals in the source and target domains to obtain the input length of the vibration signal;

[0026] The module for constructing a CNN-based transfer learning network involves building a CNN-based transfer learning network. Parameters are obtained through pre-training on source domain samples and shared with target domain samples. A feature extraction module is constructed using five convolutional networks and one deconvolutional network. Dense blocks are used to densely link all preceding and subsequent convolutional layers, reusing low-dimensional features for feature extraction and obtaining hierarchical discrimination across different layers. Two fully connected layers, FC1 and FC2, are used to construct the output layer, representing the classifier. The fully connected layer uses Conv4 and Conv5 to extract high-level features as input. Multiple non-linear instance normalization (IN) and Dropout methods are embedded in the fully connected layer to alleviate overfitting. Forward propagation is used to perform non-linear mapping layer by layer between network layers, and backpropagation is used to update the parameters of each layer and minimize cross-entropy error.

[0027] The fault diagnosis module constructs the final optimized loss function L, which consists of parallel integrated optimization, classification error, and a trade-off function. The parameters are updated through backpropagation to obtain the optimal CNN-based transfer learning network. The data preprocessed by the adaptive length input preprocessing module is then input into the optimal network to complete the intelligent health status identification and diagnosis of the bearing.

[0028] The beneficial effects of this invention are as follows: First, based on signal demodulation methods, an adaptive input module is constructed to automatically select the input length from the original vibration signal. Second, a TL network with low-dimensional feature reuse is constructed to achieve weight and bias sharing. Third, a parallel ensemble optimization loss function is proposed to align data with unknown distributions between the source and target domains. Finally, the performance of the proposed method is verified through two examples: multi-source, unsupervised, and cross-domain TL. The results show that this method not only overcomes the limitations of manual input length setting but also overcomes the limitations of the optimization function, making it more effective than existing intelligent fault diagnosis models. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0030] Figure 2 This is a schematic diagram of the system module structure of the present invention;

[0031] In the diagram, 110 is the signal acquisition module; 120 is the adaptive length input preprocessing module; 130 is the CNN-based transfer learning network construction module; and 140 is the fault diagnosis module. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1 This invention provides a technical solution: a bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning, comprising the following steps:

[0034] S1. Use an accelerometer to collect vibration signals from the source and target domains;

[0035] S2. Adaptive Length Input Preprocessing: An adaptive input length module is constructed to preprocess the vibration signals in the source and target domains. The input lengths of the source and target domains are selected respectively to reduce the loss of fault characteristics and obtain the input length of the vibration signal.

[0036] S3. Construct a CNN-based transfer learning network. Obtain parameters through pre-training with source domain samples and share training parameters with target domain samples. Use five convolutional networks and one deconvolutional network to construct a feature extraction module. Utilize dense blocks to densely link all preceding and subsequent convolutional layers, reusing low-dimensional features for feature extraction and obtaining hierarchical discrimination between different layers. Use two fully connected layers, FC1 and FC2, to construct the output layer, which represents the classifier. The fully connected layer uses Conv4 and Conv5 to extract high-level features as input. Embed multiple non-linear instance normalization (IN) and Dropout methods in the fully connected layer to alleviate overfitting. Use forward propagation to perform non-linear mapping layer by layer between each network layer. Use backpropagation to update the parameters of each layer and minimize the cross-entropy error.

[0037] CNN networks employ methods such as group convolution, instance normalization, dense blocks, ReLU, dropout, and low-dimensional feature reuse to achieve feature extraction and transfer learning. CNN networks are used to extract features and obtain hierarchical discrimination across different layers. This invention proposes an improved CNN-based method that combines an adaptive input length selection strategy and low-dimensional feature reuse to enhance the model's generalization performance.

[0038] S4. Construct the final optimized loss function L, which consists of parallel ensemble optimization, classification error, and trade-off function. Update the parameters through backpropagation to obtain the optimal CNN-based transfer learning network. Input the data preprocessed in step S2 into the optimal network to complete the intelligent health status identification and diagnosis of bearings.

[0039] Furthermore, step S2 specifically includes the following: dividing the vibration signal into segments of length L using overlapping data segmentation technology. i Several parts of the signal are processed using an envelope spectrum-based signal demodulation method to extract low-frequency features.

[0040] The maximum value of the fault characteristic frequency f and the bandwidth B ω The relationship between them is expressed by the following formula:

[0041] B ω =t×f,

[0042] Where t∈(1,2,3,4) represents a multiple of the fault characteristic frequency;

[0043] The input length L of the vibration signal a The formula is expressed as follows:

[0044]

[0045] Where f is the maximum value of the bearing fault characteristic frequency, B ω For bandwidth, L i f is the signal length. s Where is the sampling frequency. Based on the above formula, the input length of the vibration signal can be obtained.

[0046] Furthermore, the CNN-based transfer learning network includes five convolutional modules (Conv1, Conv2, Conv3, Conv4, and Conv5), one deconvolutional module (Tconv), and two fully connected layers (FC1 and FC2). Each convolutional module includes a grouped convolutional layer (GC), an instance normalization layer (IN), and an activation layer. The network structure is shown in Table 1.

[0047] Table 1 Network Structure Parameters

[0048]

[0049] To ensure data dimensionality consistency, a deconvolutional layer Tconv is used after Conv2. The dense blocks of a Dense Net combine the output of Conv1 with the output of the deconvolutional layer TConv and use them together as the input to Conv3, making the model more compact.

[0050] Grouped convolution (GC) is a type of CNN network that utilizes the sparsity of inter-layer filter dependencies to quickly reduce the number of parameters without compromising accuracy. The input feature map is grouped and convolved. Assume the input x... i The feature mapping is divided into layers G in layer i, named and This is the weight matrix of the grouped convolution. Grouped convolution can be represented as:

[0051]

[0052]

[0053] Where * indicates convolution calculation.

[0054] Instance normalization (IN) is a type of batch normalization.

[0055]

[0056]

[0057]

[0058] Where x i and y i These are the input and output of the normalization layer, respectively, and i is an exponent.

[0059] For a one-dimensional signal, i = (i N i C ) is a feature vector with index order (N, C), where N is the batch axis, C is the channel axis, and u i σ, ∈ represent the mean, standard deviation, and constant, respectively. i It is the set of pixels for calculating the mean and standard deviation, and t is the size of the set.

[0060] DenseNet, composed of dense blocks, offers two advantages: mitigating the vanishing gradient problem and encouraging feature reuse. th The layer receives features from the previous layer and combines them with the feature maps [x0, x1, ..., xn] generated in layers 0, ..., l-1. l-1 As input, that is:

[0061] x l =H l ([x0,x1,...,x l-1 ])

[0062] Where H l(·) is used as a composite function. As a direct result of the input connections, the feature maps are learned through the previous layer. This encourages the reuse of features in earlier layers, resulting in a more compact model.

[0063] High-level features are defined as high-level semantic information, which is information derived from other information and then used for classification.

[0064] Multiple non-linear instance normalization (IN) and Dropout are embedded in the fully connected layer (dropout can effectively alleviate overfitting).

[0065] Furthermore, a forward propagation method is used to perform nonlinear mapping layer by layer between each network layer, and a backpropagation method is used to update the parameters of each layer, thereby minimizing the cross-entropy error.

[0066] To address the unknown distributional differences between the source and target domains, a parallel ensemble optimization framework is employed.

[0067] When dealing with unknown distributional differences between the source and target domains, current popular methods each have their advantages and disadvantages. The construction of a parallel ensemble optimization framework involves assigning certain weights to these differences.

[0068] The final optimization loss function L is expressed as follows:

[0069] L = L c +βL es +h(α);

[0070]

[0071]

[0072] Where β and α are respectively L c and L es The trade-off parameter, L c It is the classification loss function, L es It is the parallel ensemble optimization loss function, which is the sum of the loss functions of the five alignment methods. β is a hyperparameter in [0.001, 0.01, 0.05, 0.5, 1, 5, 10], and α is a trainable parameter.

[0073] Furthermore, the principles of the methods included in the ensemble optimization framework are described as follows:

[0074] 1. The Wasserstein metric appears in optimal transportation problems. The dual representation of the first Wasserstein distance (bulldozing distance) can be written in the form of an integral probability metric:

[0075]

[0076] The Lipschitz semi-normal is defined as ||f|| L = sup|f(x)-f(y)| / ρ(x,y), W1(S,T) is the minimum expected transportation cost. From position x to position x, the marginal distributions x~P and y~T are satisfied, that is, the cost of transporting one unit of material is given by ρ(x,y).

[0077] 2. DAN: The invariance of deep features, the adaptive distribution of the domain, and the separation of low-density regions are all key factors affecting the domain adaptation performance. A unified DAN ensemble is as follows:

[0078]

[0079] in and These are the points that CNN will use. and The probability distribution assigned to all labels and categories. L(·,·) is the cross-entropy loss function. H(·) is the entropy loss. MK-MMD / ME evaluates the source and target data through a layer l representation. λ and γ These are the multilevel MK-MMD / ME penalty factor and the conditional entropy penalty trade-off parameter, respectively.

[0080] 3. MMD is used as the source domain X for test statistics. s and target domain X t The difference between them is based on the probability distribution of the embedding to the Hilbert space of the reproduction kernel.

[0081]

[0082] Where φ(*) represents the feature mapping function, n s n t These are the sampling sizes for the source and target domains, respectively.

[0083]

[0084] 4. JMD is composed of The empirical estimate is calculated as the squared distance of the empirical kernel mean embedding, i.e.:

[0085]

[0086] Where, sample points n s and n t The deep network will generate activation layers from the source domain and the target domain, respectively. L As and JMDD uses non-uniform weights to reflect other factors. The influence of layers.

[0087] 5. The purpose of MCD is to maximize the difference between the outputs of two classifiers to detect target samples far from the source support. Xs and Xt represent the K-dimensional probability outputs of the source and target domains for the input x, respectively. The absolute value of the difference between the probability outputs of the two domains is used as the difference loss:

[0088]

[0089] Where X sk and X tk X represents category K respectively s and X t For the output, this study chooses L1-distance based on this theorem.

[0090] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: a bearing intelligent fault diagnosis system based on parallel ensemble optimization and transfer learning, the system comprising the following:

[0091] Signal acquisition module 110 acquires vibration signals from the source domain and the target domain;

[0092] Adaptive length input preprocessing module 120: Constructs an adaptive input length module to preprocess the vibration signals in the source domain and the target domain to obtain the input length of the vibration signal;

[0093] The CNN-based transfer learning network construction module 130 constructs a CNN-based transfer learning network. Parameters are obtained through pre-training on source domain samples and shared with target domain samples. A feature extraction module is constructed using five convolutional networks and one deconvolutional network. Dense blocks are used to densely link all preceding and subsequent convolutional layers, reusing low-dimensional features for feature extraction and obtaining hierarchical discrimination between different layers. Two fully connected layers, FC1 and FC2, are used to construct the output layer, represented by the classifier. The fully connected layer uses Conv4 and Conv5 to extract high-level features as input. Multiple non-linear instance normalization (IN) and Dropout are embedded in the fully connected layer to alleviate overfitting. The forward propagation method is used to perform non-linear mapping layer by layer between each network layer, and the backpropagation method is used to update the parameters of each layer and minimize the cross-entropy error.

[0094] The fault diagnosis module (140) constructs the final optimized loss function L, which consists of parallel integrated optimization, classification error and trade-off function, and updates the parameters through backpropagation to obtain the optimal CNN-based transfer learning network. The data preprocessed by the adaptive length input preprocessing module is input into the optimal network to complete the intelligent health status identification and diagnosis of bearings.

[0095] Experimental verification

[0096] To verify the superiority of the ensemble optimization method, experiments were conducted with individual alignment methods and with ensemble optimization lacking one alignment method. The alignment methods included Ws, DAN, MMD, JMD, and MCD. Ensemble optimization lacking one alignment method included ensemble optimization without Ws (Es_wws), ensemble optimization without DAN (Es_wdan), ensemble optimization without MMD (Es_wmmd), ensemble optimization without JMD (Es_wjmd), and ensemble optimization without MCD (Es_wmcd). To further verify the superiority of the proposed method, DANN, TICNN, and MBSCNN networks were used to compare the state of the proposed network.

[0097] Comparison method:

[0098] DANN: DANN is an efficient method that trains on unlabeled data from the source domain and tests on unlabeled data from the target domain. It uses standard layers, gradient inversion layers, stochastic momentum gradient descent, and loss functions.

[0099] TICNN: TICNN processes the raw vibration signal directly without any preprocessing. It uses a wide convolutional kernel in the first convolutional layer and multiple layers of small convolutional kernels in deeper layers. Batch normalization is implemented after the convolutional and fully connected layers to accelerate the training process.

[0100] MBSCNN: MBSCNN combines traditional filtering methods with multi-scale learning concepts, integrating signal conversion, multi-feature learning, multi-feature fusion, and fault classification.

[0101] Verification 1: Case Western Reserve University Dataset

[0102] 0HP, 1HP, 2HP, and 3HP correspond to four different domains, each containing four health states. Health states include Normal (Nm), Inner Ring Failure (IF), Outer Ring Failure (OF), and Roller Failure (BF). A ij and A k (i,j,k=0,1,2,3,i≠j≠k) represent the source and target domain datasets, respectively. 0, 1, 2,3 represent 0HP, 1HP, 2HP, and 3HP, respectively. For example, the task A01→A2 represents the transfer of knowledge learned from the target multi-source domain A01 to the target domain A2. The accuracy of various methods on the CWRU dataset is shown in Table 2.

[0103] Table 2: Accuracy of various methods on the CWRU dataset

[0104]

[0105] Continued from Table 2

[0106]

[0107]

[0108] As shown in the table, the CWRU bearing dataset is high-quality data, and most methods achieved good results. However, the accuracy of a single alignment method varies depending on the task. JMD and Ws performed poorly on the TL task. This indicates that, using only the simplest addition principle, the ensemble optimization method compensates for the shortcomings of a single alignment method.

[0109] Verification 2: High-speed train dataset

[0110] Three health states were established in high-speed rail vehicles, and detailed dataset information was set. A fault with a diameter of 0.014 inches was selected for the study. Each task had three operating states: normal (Nm), outer ring fault (OF), and inner ring fault (IF). H_ij and H_k (i,j,k=0,1,2,i≠j≠k) represent the source and target domain datasets, respectively. 0, 1, and 2 represent 30km / h, 50km / h, and 100km / h, respectively. For example, the task H01→H2 represents the transfer of knowledge learned from the multi-source domain H01 to the target domain H2. The accuracy of various methods on the RHV dataset is shown in Table 3.

[0111] Table 3. Accuracy of each method on the RHV dataset

[0112]

[0113] Continued from Table 3

[0114]

[0115]

[0116] As can be observed from Table 3, the results show that Es performs significantly better when the data quality is poor. The high-speed train dataset is closer to actual engineering conditions, thus better validating the effectiveness of the proposed method. Therefore, the parallel ensemble optimization method is a very promising approach.

[0117] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning, characterized in that, The steps include the following: S1. Acquire vibration signals from the source and target domains; S2. Adaptive Length Input Preprocessing: An adaptive input length module is constructed to preprocess the vibration signals in the source and target domains to obtain the input length of the vibration signals; S3. Construct a CNN-based transfer learning network. Obtain parameters through pre-training with source domain samples and share training parameters with target domain samples. Use five convolutional networks and one deconvolutional network to construct a feature extraction module. Utilize dense blocks to densely link all preceding and subsequent convolutional layers, reusing low-dimensional features for feature extraction and obtaining hierarchical discrimination between different layers. Use two fully connected layers, FC1 and FC2, to construct the output layer, which represents the classifier. The fully connected layer uses Conv4 and Conv5 to extract high-level features as input. Embed multiple non-linear instance normalization (IN) and Dropout methods in the fully connected layer to alleviate overfitting. Use forward propagation to perform non-linear mapping layer by layer between each network layer. Use backpropagation to update the parameters of each layer and minimize the cross-entropy error. S4. Construct the final optimized loss function consisting of parallel ensemble optimization, classification error and trade-off function, and update the parameters through backpropagation to obtain the optimal CNN-based transfer learning network. Input the data preprocessed in step S2 into the optimal network to complete the intelligent health status identification and diagnosis of bearings. The final optimized loss function formula is expressed as follows: ； ； ； in and They are and The trade-off parameters, It is a classification loss function. It is the parallel ensemble optimization loss function.

2. The bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning according to claim 1, characterized in that: Step S2 specifically includes the following: dividing the vibration signal into segments of length [length missing] using overlapping data segmentation technology. Several parts of the signal are processed using an envelope spectrum-based signal demodulation method to extract low-frequency features. Maximum value of fault characteristic frequency With bandwidth The relationship between them is expressed by the following formula: ； in (1, 2, 3, 4) represent multiples of the fault characteristic frequency; input length of vibration signal The formula is expressed as follows: ； in, This represents the maximum value of the characteristic frequency of bearing failure. For bandwidth, For signal length, The sampling frequency.

3. The bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning according to claim 1, characterized in that: The CNN-based transfer learning network includes five convolutional modules (Conv1, Conv2, Conv3, Conv4, and Conv5), one deconvolutional module (Tconv), and two fully connected layers (FC1 and FC2). Each convolutional module includes a grouped convolutional layer (GC), an instance normalization layer (IN), and an activation layer.

4. The bearing intelligent fault diagnosis method based on parallel ensemble optimization and transfer learning according to claim 1, characterized in that: In step S3, in order to ensure the consistency of data dimensions, a deconvolutional layer Tconv is used after Conv2. Dense blocks are used to combine the output of Conv1 with the output of the deconvolutional layer TConv and use them together as the input of Conv3.

5. A system for intelligent bearing fault diagnosis based on parallel ensemble optimization and transfer learning according to any one of claims 1-4, characterized in that: The system includes the following: The signal acquisition module (110) acquires vibration signals from the source and target domains; Adaptive length input preprocessing module (120): Constructs an adaptive input length module to preprocess the vibration signals in the source domain and the target domain to obtain the input length of the vibration signal; The CNN-based transfer learning network construction module (130) constructs a CNN-based transfer learning network, obtains parameters through pre-training of source domain samples, and shares training parameters with target domain samples. It uses five convolutional networks and one deconvolutional network to construct a feature extraction module, and uses dense blocks to densely link all the preceding convolutional layers and the following convolutional layers, reuses low-dimensional features, and uses them to extract features and obtain hierarchical discrimination of different layers. It uses two fully connected layers FC1 and FC2 to construct the output layer represented by the classifier. The fully connected layer uses Conv4 and Conv5 to extract high-level features as input. Multiple non-linear instance normalization IN and Dropout are embedded in the fully connected layer to alleviate overfitting. The forward propagation method is used to perform non-linear mapping layer by layer between each network layer. The back propagation method is used to update the parameters of each layer and minimize the cross-entropy error. The fault diagnosis module (140) constructs the final optimized loss function consisting of parallel integrated optimization, classification error and trade-off function, and updates the parameters through backpropagation to obtain the optimal CNN-based transfer learning network. The data preprocessed by the adaptive length input preprocessing module is input into the optimal network to complete the intelligent health status identification and diagnosis of bearings.

Images