Rolling bearing data gravity center bridging intelligent diagnosis method for mechanical equipment cluster

Abstract

A rolling bearing data center-of-gravity bridging intelligent diagnosis method for a mechanical equipment cluster comprises the following steps: firstly, collecting distribution parameters of rolling bearing data of mechanical equipment in each source domain, and calculating a dynamic center-of-gravity as intermediate bridging distribution of rolling bearing data of multi-source domain mechanical equipment and rolling bearing data of target domain mechanical equipment; then, broadcasting gravity center distribution to each node for adjusting a diagnosis model of the rolling bearing data of the mechanical equipment in each source domain, designing an adaptive path through an optimal transmission theory, aligning a label-free sample of the rolling bearing data of the mechanical equipment in a target domain to a dynamic distribution gravity center, and keeping unification of edge distribution and condition distribution at the same time; and finally, through cooperative training between the source domain mechanical equipment rolling bearing data and the target domain mechanical equipment rolling bearing data, obtaining a diagnosis model with high cross-domain generalization ability. According to the invention, the data transmission demand is reduced, the privacy protection is enhanced, and the accuracy and robustness of the fault diagnosis model on the target domain are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault diagnosis, and in particular relates to an intelligent diagnosis method for rolling bearing data center bridge of a mechanical equipment cluster. Background Art

[0002] Mechanical equipment clusters play a key role in modern industries such as wind farms and flexible production plants. They rely on multiple machine nodes to work together to complete complex tasks. Ensuring the safe operation of these machine nodes is essential to maintaining high efficiency and product quality. Intelligent fault diagnosis achieves early identification, precise positioning and predictive maintenance of faults through data collection, feature extraction, machine learning and real-time monitoring, which is an important means to ensure the healthy operation of machine nodes.

[0003] With the rapid development of deep transfer learning, diagnostic knowledge can be transferred from the rolling bearing node of one device in a mechanical equipment cluster (source domain) to the rolling bearing nodes of other similar devices (target domain), thereby reducing the strong dependence of intelligent diagnostic models on high-quality monitoring data in the target domain in cross-device applications. Multi-source transfer learning ([1] B. Yang, S. Xu, Y. Lei, C.-G. Lee, E. Stewart, and C. Roberts, "Multisource transfer learning network to complement knowledge for intelligent diagnosis of machines with unseen faults," Mechanical Systems and Signal Processing, vol. 162, p. 108095, 2022.) learns domain-invariant features from data of multiple source domain mechanical nodes and fuses the multi-source domain diagnostic knowledge to apply it to rolling bearing fault identification of mechanical equipment in a single target domain. However, the current rolling bearing data of multiple source domain mechanical nodes should be centralized, which is difficult to achieve in practice. The reasons are: (1) Data privacy: The monitoring data of a machine contains sensitive information such as production quality and manufacturing process, which is a trade secret, especially for machines of different users. Therefore, most companies are reluctant to centralize these data to prevent the leakage of negative information. (2) Cost: The concentration and analysis of data consumes huge costs. Fast data transmission, especially long-distance data transmission, requires a large-bandwidth industrial Internet.

[0004] Federated learning ([2]T.Li,AKSahu,A.Talwalkar,and V.Smith, "Federated learning: Challenges, methods, and future directions,"IEEE Signal Processing Magazine,vol.37,no.3,pp.50-60,2020.) uses the local data of multiple clients to train a global model, that is, building a server to aggregate the rolling bearing data of each source domain, such as the model parameters and high-level features of the source domain, and broadcast the aggregated model back to each source domain for further training; however, existing methods have the following disadvantages: (1) Most methods only consider the marginal distribution in the process of migrating feature distribution adaptation, but do not consider the conditional distribution, resulting in poor feature adaptation and low model migration diagnosis accuracy. (2) Existing methods use intermediate distribution to bridge the source domain and the target domain, but the generation method of the intermediate distribution is only generated through a random Gaussian distribution model, which is difficult to control and lacks theoretical guidance.

[0005] In summary, the existing multi-source domain migration diagnosis methods are difficult to fully adapt to the feature distribution of multi-source domain and target domain data while ensuring data privacy, which reduces the generalization performance of fault diagnosis of intelligent diagnosis methods in mechanical equipment clusters. Summary of the invention

[0006] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a rolling bearing data center of gravity bridging intelligent diagnosis method for a mechanical equipment cluster. Through the collaborative training of the server and multi-domain terminals, the trained general model can complete the diagnosis task on the rolling bearing data of the target domain mechanical equipment, thereby protecting the data privacy of each node, reducing the transmission cost, and improving the generalization diagnosis performance of the diagnosis model on the mechanical equipment cluster.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A rolling bearing data center bridge intelligent diagnosis method for a mechanical equipment cluster includes the following steps:

[0009] Step 1: Construct a feature extractor to perform distribution modeling on the extracted high-level features; the feature extractor consists of stacked residual blocks, denoted as Where s represents the source domain mechanical equipment rolling bearing data set, i represents the i-th source domain, θ represents the parameters of the feature extractor, and f represents the network structure used for feature extraction; the extracted high-level features v s:i =g(x s:i ) is input to a fully connected layer for classification, where x s:iRepresents the feature vector of the sample in the i-th source domain mechanical equipment rolling bearing data; then the deep residual network on the source domain mechanical equipment rolling bearing data side is trained by minimizing the cross entropy loss, that is:

[0010]

[0011] Where L represents the classification cross entropy loss, C is the cost function associated with the classification task, n is the number of samples, y is the true label of the sample, j is the index of the health status, k is the number of samples, is the indicator function, that is, when the true label of the kth sample in the i-th source domain mechanical equipment rolling bearing data is equal to j, its value is 1, otherwise it is 0; It represents the probability that the kth sample in the i-th source domain mechanical equipment rolling bearing data belongs to the j-th health state predicted by the classifier, that is, the output of the classifier; after the classification is completed, the Gaussian mixture model is used to fit the distribution information of the high-level features in each source domain mechanical equipment rolling bearing data, that is:

[0012]

[0013] Among them G s:i represents the Gaussian mixture model of the rolling bearing data of the source domain mechanical equipment, Θ represents the parameter set of the Gaussian mixture model, α represents the weight of the Gaussian mixture model, represents the probability density function of the t-th Gaussian mixture model in the i-th source domain mechanical equipment rolling bearing data, where is the mean vector, is the covariance matrix; finally, the network server is used to collect and aggregate the model parameters and characteristic distribution parameters of the rolling bearing data of mechanical equipment in each source domain to obtain the multi-source Gaussian mixture model parameters;

[0014] Step 2: Solve the dynamic distribution center of gravity through the multi-source Gaussian mixture model parameters obtained in step 1; define a distribution center of gravity Φ * The set includes the Gaussian mixture model of the rolling bearing data of mechanical equipment in each source domain to the final candidate center of gravity distribution All possible transmission plans are:

[0015]

[0016] Where inf means finding a transmission plan that minimizes the weighted Wasserstein distance. represents the candidate centroid Gaussian distribution at time or iteration step t, Represents the Gaussian distribution of rolling bearing data from various source domain mechanical equipment To the candidate centroid distribution The set of all possible transmission plans, λi represents the Gaussian distribution weight of the rolling bearing data of each source domain mechanical equipment and is non-negative, and its sum is equal to 1. The square of the Wasserstein distance between the Gaussian distribution of the rolling bearing data of each source domain mechanical equipment and the candidate center of gravity distribution is represented; the parameters of the center of gravity are calculated, namely:

[0017]

[0018] in represents the mean and covariance matrix of the center of gravity of each source domain mechanical equipment rolling bearing data, t represents the number of iteration steps, N represents the total number of source domain mechanical equipment rolling bearing data, n and n+1 represent the iteration steps respectively, and the center of gravity and covariance matrix of the current iteration are used to calculate the mean and covariance matrix of the center of gravity in the next iteration step until the parameters converge to a stable value;

[0019] Step 3: Perform targeted adaptation to the distribution center; first randomly generate a data set of intermediate samples based on the center distribution parameters Each sample By the feature vector and tags In the target adaptation process, the corresponding OT divergence is solved to evaluate the distribution difference, that is, in the rolling bearing data of mechanical equipment in each source domain, the relationship matrix A is first constructed, and the matrix elements are 0 or 1, indicating inconsistent or consistent labels respectively; then, according to the relationship matrix A, the cost function C1 is designed, that is:

[0020]

[0021] where a j,k represents the label relationship between the source domain mechanical equipment rolling bearing data sample and the intermediate sample, bc represents the intermediate sample, ε represents a very small positive number, j, k represents the index of the position; using the cost function C1, the optimal transmission divergence from the source domain mechanical equipment rolling bearing data to the intermediate sample set is calculated, that is:

[0022]

[0023] Where L da represents the OT divergence value during the transmission process, P represents the data distribution of rolling bearings of mechanical equipment in the source domain, and G bc represents the intermediate distribution, Γ represents the transmission strategy from the source domain mechanical equipment rolling bearing data to the center of gravity of the intermediate distribution, which is optimized through this objective function, namely:

[0024]

[0025] Where Π(P s:i ,Gbc ) represents the joint distribution space of the rolling bearing data distribution and the intermediate distribution of the source domain mechanical equipment, and must meet the following constraints: where u k and v j They are the marginal distributions of the joint distribution, and the initialization values ​​are and Solving Γ by the Sinkhorn-Knopp fixed-point iteration algorithm s:i→bc ; Finally, the adapted samples and corresponding labels are used to train the local diagnosis model of the rolling bearing data of mechanical equipment in each source domain, namely:

[0026]

[0027] Where γ is a trade-off parameter for balancing the classification loss function and the target distribution adaptation; then the Adam optimizer is used to update the model parameters of the rolling bearing data of mechanical equipment in each source domain, and the updated model parameters are uploaded to the network server;

[0028] Step 4: Perform target adaptation on the target domain mechanical equipment rolling bearing data side; according to the aggregation algorithm of FedAvg, aggregate the model parameters of the mechanical equipment rolling bearing data in each source domain in step 3 to obtain the average parameter 1 / N∑ i θ s:i , initialize the diagnostic model parameters θ of the target domain mechanical equipment rolling bearing data t ; Define a small number of labeled samples of rolling bearing data of mechanical equipment in the target domain as anchor samples, and define the label relationship b between the anchor samples and the intermediate samples j,i for:

[0029]

[0030] in represents the label of the jth anchor sample, Represents the label of the i-th intermediate sample; defines the loss function of the anchor sample for:

[0031]

[0032] Where X represents the anchor sample set, t represents the rolling bearing data of mechanical equipment in the target domain, and A represents the anchor sample set. represents the label of the i-th anchor sample, is an indicator function, that is, when the label of the i-th sample is equal to category j, it returns 1, otherwise it returns 0; represents the model's predicted probability that the i-th anchor point sample in the target domain mechanical equipment rolling bearing data belongs to category j, R represents the total number of categories, and n bcrepresents the number of intermediate samples; Gaussian mixture model is fitted for the rolling bearing data samples of mechanical equipment in the target domain, and the number of Gaussian mixture models is equal to the number of healthy states of anchor samples; the label relationship e between unlabeled samples and anchor samples is defined k,j for:

[0033]

[0034] in is the pseudo label of the kth unlabeled sample, is the pseudo label of the jth anchor sample, U is the set of unlabeled samples; the cost function C2 is defined according to the label relationship as:

[0035]

[0036] where e k,j represents the label relationship between the unlabeled samples and the anchor samples, b j,i Represents the label relationship between the anchor sample and the intermediate sample; based on the cost function, calculates the OT divergence from the target domain mechanical equipment rolling bearing data to the intermediate domain:

[0037]

[0038] in represents the domain adaptation loss of the target domain mechanical equipment rolling bearing data, P t represents the distribution of rolling bearing data samples of mechanical equipment in the target domain, G bc represents the distribution of samples in the middle domain, It represents the transmission strategy from the target domain mechanical equipment rolling bearing data to the intermediate domain. Its calculation method is the same as step 3. The gradient of the target domain mechanical equipment rolling bearing data is calculated, that is:

[0039]

[0040] Where β is the balance parameter; the Adam optimizer is used to update the diagnostic model parameter θ of the rolling bearing data of the target domain mechanical equipment based on the gradient. t ;

[0041] Step 5: Repeat steps 1 to 4 to iteratively update the parameters of the target domain mechanical equipment rolling bearing data diagnosis model until the model parameters converge;

[0042] Step 6: Save the diagnostic model g(·|θ) of the rolling bearing data of the target domain mechanical equipment t ) and output the diagnosis results.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention proposes a rolling bearing data center of gravity bridging intelligent diagnosis method for a mechanical equipment cluster. By calculating the dynamic center of gravity of rolling bearing data of mechanical equipment from multiple source domains, the intermediate distribution is continuously adjusted, and the difference between the source domain mechanical equipment rolling bearing data and the target domain mechanical equipment rolling bearing data is reduced. The controllable dynamic center of gravity is generated to perform targeted distribution adjustment with the source domain data. The model can be trained in a data decentralized environment, which protects the data privacy of each node and reduces the transmission cost. Not only the edge distribution of multi-domain data is aligned, but also the conditional distribution difference is processed through the adaptive alignment of the target domain mechanical equipment rolling bearing data and the intermediate domain, thereby improving the generalization diagnosis performance of the diagnosis model on the mechanical equipment cluster. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with the embodiments and drawings.

[0047] like Figure 1 As shown, a rolling bearing data center bridge intelligent diagnosis method for a mechanical equipment cluster includes the following steps:

[0048] Step 1: Construct a feature extractor to extract high-level features from the source domain mechanical equipment rolling bearing data and perform distribution modeling on the extracted high-level features. For each source domain mechanical equipment rolling bearing data, a deep residual network is constructed to extract high-level features from the local data of the domain. The feature extractor consists of stacked residual blocks, denoted as Where s represents the source domain mechanical equipment rolling bearing data set, i represents the i-th source domain, θ represents the parameters of the feature extractor, and f represents the network structure used for feature extraction; the extracted high-level features v s:i =g(x s:i ) is input to a fully connected layer for classification, where x s:i Represents the feature vector of the sample in the i-th source domain mechanical equipment rolling bearing data; then the deep residual network on the source domain mechanical equipment rolling bearing data side is trained by minimizing the cross entropy loss, that is:

[0049]

[0050] Where L represents the classification cross entropy loss, C is the cost function associated with the classification task, n is the number of samples, y is the true label of the sample, j is the index of the health status, k is the number of samples, is the indicator function, that is, when the true label of the kth sample in the i-th source domain mechanical equipment rolling bearing data is equal to j, its value is 1, otherwise it is 0; It represents the probability that the kth sample in the i-th source domain mechanical equipment rolling bearing data belongs to the j-th health state predicted by the classifier, that is, the output of the classifier; after the classification is completed, the Gaussian mixture model is used to fit the distribution information of the high-level features in each source domain mechanical equipment rolling bearing data, that is:

[0051]

[0052] Among them G s:i represents the Gaussian mixture model of the rolling bearing data of the source domain mechanical equipment, Θ represents the parameter set of the Gaussian mixture model, α represents the weight of the Gaussian mixture model, represents the probability density function of the t-th Gaussian mixture model in the i-th source domain mechanical equipment rolling bearing data, where is the mean vector, is the covariance matrix; finally, the network server is used to collect and aggregate the model parameters and characteristic distribution parameters of the rolling bearing data of mechanical equipment in each source domain to obtain the multi-source Gaussian mixture model parameters;

[0053] Step 2: Solve the dynamic distribution center of gravity through the multi-source Gaussian mixture model parameters obtained in step 1; first establish a sub-category Gaussian model for the health status of the rolling bearing of each source domain mechanical equipment, and then define a distribution center of gravity Φ * The set includes the Gaussian mixture model of the rolling bearing data of mechanical equipment in each source domain to the final candidate center of gravity distribution All possible transmission plans of the transmission plan represent the distribution of rolling bearing data from the source domain mechanical equipment to the center of gravity distribution, that is:

[0054]

[0055] Where inf means finding a transmission plan that minimizes the weighted Wasserstein distance. represents the candidate centroid Gaussian distribution at time or iteration step t, Represents the Gaussian distribution of rolling bearing data from various source domain mechanical equipment To the candidate centroid distribution The set of all possible transmission plans, λ i represents the Gaussian distribution weight of the rolling bearing data of each source domain mechanical equipment and is non-negative, and its sum is equal to 1. Represents the square of the Wasserstein distance between the Gaussian distribution of the rolling bearing data of each source domain mechanical equipment and the candidate center of gravity distribution; the parameters of the center of gravity are calculated according to the iterative equation of the mean value and covariance matrix of the center of gravity, that is:

[0056]

[0057] in represents the mean and covariance matrix of the center of gravity of each source domain mechanical equipment rolling bearing data, t represents the number of iteration steps, N represents the total number of source domain mechanical equipment rolling bearing data, n and n+1 represent the iteration steps respectively, and the center of gravity and covariance matrix of the current iteration are used to calculate the mean and covariance matrix of the center of gravity in the next iteration step until the parameters converge to a stable value;

[0058] Step 3: Broadcast the center of gravity parameters obtained in step 2, i.e., the mean and covariance matrix, back to the rolling bearing data of each source domain mechanical equipment to perform targeted adaptation of the distribution center of gravity; first, randomly generate intermediate samples based on the center of gravity distribution parameters, and the intermediate samples constitute the data set Each sample By the feature vector and tags The samples are drawn from a Gaussian mixture distribution, and then the marginal and conditional distribution differences between multiple data domains are reduced through target adaptation. In the target adaptation process, the adaptation trajectory is pre-planned according to the label so that samples from the same health state are adapted together, and according to the optimal transmission theory, the corresponding OT divergence is solved to evaluate the distribution difference. The specific method is as follows: In each source domain mechanical equipment rolling bearing data, a relationship matrix A is first constructed to describe the label relationship between the source domain mechanical equipment rolling bearing data samples and the intermediate samples. The matrix elements are 0 or 1, indicating that the labels are inconsistent or consistent, respectively. Then, according to the relationship matrix A, a cost function C1 is designed to calculate the transmission cost of the source domain mechanical equipment rolling bearing data samples to the intermediate samples, that is:

[0059]

[0060] where a j,k represents the label relationship between the source domain mechanical equipment rolling bearing data sample and the intermediate sample, bc represents the intermediate sample, ε represents a very small positive number, and j, k represents the index of the position; using the cost function C1, the optimal transmission divergence from the source domain mechanical equipment rolling bearing data to the intermediate sample set is calculated to evaluate the difficulty of sample migration, that is:

[0061]

[0062] Where L da represents the OT divergence value during the transmission process, P represents the data distribution of rolling bearings of mechanical equipment in the source domain, and G bc represents the intermediate distribution, Γ represents the transmission strategy from the source domain mechanical equipment rolling bearing data to the center of gravity of the intermediate distribution, which is optimized through this objective function, namely:

[0063]

[0064] Where Π(Ps:i ,G bc ) represents the joint distribution space of the rolling bearing data distribution and the intermediate distribution of the source domain mechanical equipment, and must meet the following constraints: where u k and v j They are the marginal distributions of the joint distribution, and the initialization values ​​are and Solving Γ by the Sinkhorn-Knopp fixed-point iteration algorithm s:i→bc ; Finally, the adapted samples and corresponding labels are used to train the local diagnosis model of the rolling bearing data of mechanical equipment in each source domain, namely:

[0065]

[0066] Where γ is a trade-off parameter for balancing the classification loss function and the target distribution adaptation; then the Adam optimizer is used to update the model parameters of the rolling bearing data of mechanical equipment in each source domain, and the updated model parameters are uploaded to the network server;

[0067] Step 4: Perform target adaptation on the target domain mechanical equipment rolling bearing data side; according to the aggregation algorithm of FedAvg, aggregate the model parameters of the mechanical equipment rolling bearing data in each source domain in step 3 to obtain the average parameter 1 / N∑ i θ s:i , to initialize the diagnostic model parameters θ of the target domain mechanical equipment rolling bearing data t ; A small number of labeled samples of rolling bearing data of mechanical equipment in the target domain are defined as anchor samples. The label relationship between the anchor samples and the intermediate samples is transferred to other unlabeled samples through cluster analysis to assist in the pseudo-label assignment of unlabeled samples. The label relationship between the anchor samples and the intermediate samples is defined as b j,i for:

[0068]

[0069] in represents the label of the jth anchor sample, Represents the label of the i-th intermediate sample; defines the loss function of the anchor sample for:

[0070]

[0071] Where X represents the anchor sample set, t represents the rolling bearing data of mechanical equipment in the target domain, and A represents the anchor sample set. represents the label of the i-th anchor sample, is an indicator function, that is, when the label of the i-th sample is equal to category j, it returns 1, otherwise it returns 0; represents the model's predicted probability that the i-th anchor point sample in the target domain mechanical equipment rolling bearing data belongs to category j, R represents the total number of categories, which is used to calculate the loss on each category, and n bc represents the number of intermediate samples; Gaussian mixture model is fitted for the rolling bearing data samples of mechanical equipment in the target domain, and the number of Gaussian mixture models is equal to the number of healthy states of anchor samples; the label relationship e between unlabeled samples and anchor samples is defined k,j for:

[0072]

[0073] in is the pseudo label of the kth unlabeled sample, is the pseudo label of the jth anchor sample, U is the set of unlabeled samples; the cost function C2 is defined according to the label relationship as:

[0074]

[0075] where e k,j represents the label relationship between the unlabeled samples and the anchor samples, b j,i Represents the label relationship between the anchor sample and the intermediate sample; based on the cost function, calculates the OT divergence from the target domain mechanical equipment rolling bearing data to the intermediate domain:

[0076]

[0077] in represents the domain adaptation loss of the target domain mechanical equipment rolling bearing data, P t represents the distribution of rolling bearing data samples of mechanical equipment in the target domain, G bc represents the distribution of samples in the middle domain, It represents the transmission strategy from the target domain mechanical equipment rolling bearing data to the intermediate domain. Its calculation method is the same as step 3. The gradient of the target domain mechanical equipment rolling bearing data is calculated, that is:

[0078]

[0079] Where β is the balance parameter; the Adam optimizer is used to update the diagnostic model parameter θ of the rolling bearing data of the target domain mechanical equipment based on the gradient. t ;

[0080] Step 5: Repeat steps 1 to 4 to iteratively update the parameters of the target domain mechanical equipment rolling bearing data diagnosis model until the model parameters converge;

[0081] Step 6: Save the diagnostic model g(·|θ) of the rolling bearing data of the target domain mechanical equipment t ) and output the diagnosis results.

[0082] The embodiment takes four bearing data sets collected from different machines as examples to demonstrate the performance of the proposed dynamic center of gravity bridging network and verify the feasibility of the present invention. Detailed information on the bearing data sets used by multiple machines in the embodiment is shown in Table 1.

[0083] Data set A in Table 1 comes from the rotor bearing test bench. The data includes four states of the bearing: normal state (N), inner ring fault (IF), outer ring fault (OF) and roller fault (RF). The bearing model is ER-8K and is installed at the left end of the rotor bearing system. The motor speed is set to 2100r / min. At a sampling frequency of 12.8kHz, a total of 2048 samples are collected and balanced for each health state.

[0084] Dataset B is collected from the motor drive end bearing test bench. The data is collected under the conditions of motor speed of 1500r / min and output torque of 10N·m (provided by eddy current brake). The bearing model is NSK / NFC 6205. The data includes four states of the bearing: normal state (N), inner ring fault (IF), outer ring fault (OF) and roller fault (RF), and each state has 546 samples.

[0085] Dataset C comes from the water pump test bench. The tested bearing is installed between the motor and the water pump. When the motor speed is 1800r / min and the sampling frequency is 12.8kHz, the accelerometer installed on the bearing seat collects vibration data. The bearing model is ER-16K. The data includes four states of the bearing: normal state (N), inner ring fault (IF), outer ring fault (OF) and roller fault (RF). The data set contains 2160 samples.

[0086] Dataset D is collected from a locomotive bearing test bench, which is driven by a hydraulic motor. During the experiment, the inner ring of the bearing rotates at a speed of about 500 r / min. The bearing model is 552732QT. The monitoring system returns the vibration data of the bearing at a sampling frequency of 12.8 kHz. The data set contains a total of 1092 samples.

[0087] These datasets represent different machine nodes and working conditions, resulting in differences between the source domain mechanical equipment rolling bearing data and the target domain mechanical equipment rolling bearing data, as well as between the source domain mechanical equipment rolling bearing data themselves. These changes create the task of evaluating the performance of the proposed dynamic center of gravity bridging network in multi-source transmission diagnosis.

[0088] Table 1 Detailed information of the bearing dataset used by multiple machines in the embodiment

[0089]

[0090] Based on the dataset shown in Table 1, a multi-source transmission diagnosis task is created to verify the effectiveness of the proposed dynamic center of gravity bridging network, i.e., (B, C, D) → A. 25% of the target domain samples are randomly selected for training, and the remaining 75% are used for testing. In the training set, only one labeled sample is used for each health state to design the adaptation trajectory, ensuring that the labeled data available in the target domain mechanical equipment rolling bearing data is minimal. The proposed dynamic center of gravity bridging network is repeated 10 times. Under the optimal hyperparameter setting, the statistical results (mean and standard deviation) of the test dataset are shown in Table 2. On the given task, the accuracy of the proposed dynamic center of gravity bridging network is 99.28%, which verifies the effectiveness and feasibility of the dynamic center of gravity bridging network in multi-source transmission diagnosis in the federated setting.

[0091] The proposed method is further compared with other methods applied to data concentration or data decentralization, as shown in Table 2. Among them, method one is FedAvg, which has the lowest accuracy among the given methods because it has no ability to reduce cross-domain differences; method two is Finetune, which uses a small number of labeled target domain equipment rolling bearing data samples to retrain the global model, but the sample size is too small and the efficiency is not high; method three is FADA, which reduces cross-domain differences by introducing adversarial learning, aligns the distribution of target domain mechanical equipment rolling bearing data samples with the local distribution of each source domain mechanical equipment rolling bearing data, and obtains an accuracy of 71.07%, which is lower than this method; method four is FTL-PD, which randomly generates a multivariate normal distribution as an intermediate sample. This distribution may be too different from the source domain mechanical equipment rolling bearing data and the target domain mechanical equipment rolling bearing data, thereby increasing the difficulty of simultaneous adaptation between domains. In comparison, the method proposed in the present invention has a higher accuracy rate, solves the problem of the center of gravity of dynamic distribution of rolling bearing data of mechanical equipment in multiple source domains, and reduces the difficulty of synchronous adaptation between rolling bearing data of mechanical equipment in multiple source domains; in addition, the adaptation trajectory is designed using the optimal transmission theory to achieve targeted adaptation, so that the diagnosis of a given task has a higher accuracy.

[0092] Table 2 Diagnostic accuracy of different methods for multi-source migration diagnosis tasks (%) [mean ± SD]

[0093]

Claims

1. A rolling bearing data center bridge intelligent diagnosis method for a mechanical equipment cluster, characterized in that: The following steps are involved: Step 1: Construct a feature extractor to extract high-level features from the source domain mechanical equipment rolling bearing data and perform distribution modeling on the extracted high-level features. For each source domain mechanical equipment rolling bearing data, a deep residual network is constructed separately to extract high-level features from the local data of the domain. The feature extractor consists of stacked residual blocks, denoted as Where s represents the source domain mechanical equipment rolling bearing data set, i represents the i-th source domain, θ represents the parameters of the feature extractor, and f represents the network structure used for feature extraction; the extracted high-level features v s:i =g(x s:i ) is input to a fully connected layer for classification, where x s:i Represents the feature vector of the sample in the i-th source domain mechanical equipment rolling bearing data; then the deep residual network on the source domain mechanical equipment rolling bearing data side is trained by minimizing the cross entropy loss, that is: Where L represents the classification cross entropy loss, C is the cost function associated with the classification task, n is the number of samples, y is the true label of the sample, j is the index of the health status, k is the number of samples, is the indicator function, that is, when the true label of the kth sample in the i-th source domain mechanical equipment rolling bearing data is equal to j, its value is 1, otherwise it is 0; It represents the probability that the kth sample in the i-th source domain mechanical equipment rolling bearing data belongs to the j-th health state predicted by the classifier, that is, the output of the classifier; after the classification is completed, the Gaussian mixture model is used to fit the distribution information of the high-level features in each source domain mechanical equipment rolling bearing data, that is: Among them G s:i represents the Gaussian mixture model of the rolling bearing data of the source domain mechanical equipment, Θ represents the parameter set of the Gaussian mixture model, α represents the weight of the Gaussian mixture model, represents the probability density function of the t-th Gaussian mixture model in the i-th source domain mechanical equipment rolling bearing data, where is the mean vector, is the covariance matrix; finally, the network server is used to collect and aggregate the model parameters and characteristic distribution parameters of the rolling bearing data of mechanical equipment in each source domain to obtain the multi-source Gaussian mixture model parameters; Step 2: Solve the dynamic distribution center of gravity through the multi-source Gaussian mixture model parameters obtained in step 1; first establish a sub-category Gaussian model for the health status of the rolling bearing of each source domain mechanical equipment, and then define a distribution center of gravity Φ * The set includes the Gaussian mixture model of the rolling bearing data of mechanical equipment in each source domain to the final candidate center of gravity distribution All possible transmission plans of the transmission plan represent the distribution of rolling bearing data from the source domain mechanical equipment to the center of gravity distribution, that is: Where inf means finding a transmission plan that minimizes the weighted Wasserstein distance. represents the candidate centroid Gaussian distribution at time or iteration step t, Represents the Gaussian distribution of rolling bearing data from various source domain mechanical equipment To the candidate centroid distribution The set of all possible transmission plans, λ i represents the Gaussian distribution weight of the rolling bearing data of each source domain mechanical equipment and is non-negative, and its sum is equal to 1. Represents the square of the Wasserstein distance between the Gaussian distribution of the rolling bearing data of each source domain mechanical equipment and the candidate center of gravity distribution; the parameters of the center of gravity are calculated according to the iterative equation of the mean value and covariance matrix of the center of gravity, that is: in represents the mean and covariance matrix of the center of gravity of each source domain mechanical equipment rolling bearing data, t represents the number of iteration steps, N represents the total number of source domain mechanical equipment rolling bearing data, n and n+1 represent the iteration steps respectively, and the center of gravity and covariance matrix of the current iteration are used to calculate the mean and covariance matrix of the center of gravity in the next iteration step until the parameters converge to a stable value; Step 3: Broadcast the center of gravity parameters obtained in step 2, i.e., the mean and covariance matrix, back to the rolling bearing data of each source domain mechanical equipment to perform targeted adaptation of the distribution center of gravity; first, randomly generate intermediate samples based on the center of gravity distribution parameters, and the intermediate samples constitute the data set Each sample By the feature vector and tags The samples are drawn from a Gaussian mixture distribution, and then the marginal and conditional distribution differences between multiple data domains are reduced through target adaptation. In the target adaptation process, the adaptation trajectory is pre-planned according to the label so that samples from the same health state are adapted together, and the corresponding OT divergence is solved according to the optimal transmission theory to evaluate the distribution difference. Step 4: Perform target adaptation on the target domain mechanical equipment rolling bearing data side; according to the aggregation algorithm of FedAvg, aggregate the model parameters of the mechanical equipment rolling bearing data in each source domain in step 3 to obtain the average parameter 1 / N∑ i θ s:i , to initialize the diagnostic model parameters θ of the target domain mechanical equipment rolling bearing data t ; A small number of labeled samples of rolling bearing data of mechanical equipment in the target domain are defined as anchor samples. The label relationship between the anchor samples and the intermediate samples is transferred to other unlabeled samples through cluster analysis to assist in the pseudo-label assignment of unlabeled samples. The label relationship between the anchor samples and the intermediate samples is defined as b j,i for: in represents the label of the jth anchor sample, Represents the label of the i-th intermediate sample; defines the loss function of the anchor sample for: Where X represents the anchor sample set, t represents the rolling bearing data of mechanical equipment in the target domain, and A represents the anchor sample set. represents the label of the i-th anchor sample, is an indicator function, that is, when the label of the i-th sample is equal to category j, it returns 1, otherwise it returns 0; represents the model's predicted probability that the i-th anchor point sample in the target domain mechanical equipment rolling bearing data belongs to category j, R represents the total number of categories, which is used to calculate the loss on each category, and n bc represents the number of intermediate samples; Gaussian mixture model is fitted for the rolling bearing data samples of mechanical equipment in the target domain, and the number of Gaussian mixture models is equal to the number of healthy states of anchor samples; the label relationship e between unlabeled samples and anchor samples is defined k,j for: in is the pseudo label of the kth unlabeled sample, is the pseudo label of the jth anchor sample, U is the set of unlabeled samples; the cost function C2 is defined according to the label relationship as: where e k,j represents the label relationship between the unlabeled samples and the anchor samples, b j,i Represents the label relationship between the anchor sample and the intermediate sample; based on the cost function, calculates the OT divergence from the target domain mechanical equipment rolling bearing data to the intermediate domain: in represents the domain adaptation loss of the target domain mechanical equipment rolling bearing data, P t represents the distribution of rolling bearing data samples of mechanical equipment in the target domain, G bc represents the distribution of samples in the middle domain, It represents the transmission strategy from the target domain mechanical equipment rolling bearing data to the intermediate domain, and its calculation method is the same as step 3; Calculate the gradient of the rolling bearing data of the target domain mechanical equipment, that is: Where β is the balance parameter; the Adam optimizer is used to update the diagnostic model parameter θ of the rolling bearing data of the target domain mechanical equipment based on the gradient. t ; Step 5: Repeat steps 1 to 4 to iteratively update the parameters of the target domain mechanical equipment rolling bearing data diagnosis model until the model parameters converge; Step 6: Save the diagnostic model g(·|θ) of the rolling bearing data of the target domain mechanical equipment t ) and output the diagnosis results.

2. The method according to claim 1, characterized in that Step 3 is as follows: broadcast the center of gravity parameters obtained in step 2, namely the mean and covariance matrix, back to the rolling bearing data of each source domain mechanical equipment to perform targeted adaptation of the distribution center of gravity; first randomly generate intermediate samples according to the center of gravity distribution parameters, and the intermediate samples constitute the data set Each sample By the feature vector and tags The samples are drawn from a Gaussian mixture distribution, and then the marginal and conditional distribution differences between multiple data domains are reduced through target adaptation. In the target adaptation process, the adaptation trajectory is pre-planned according to the label so that samples from the same health state are adapted together, and the corresponding OT divergence is solved according to the optimal transmission theory to evaluate the distribution difference. The specific method is as follows: In the rolling bearing data of mechanical equipment in each source domain, the relationship matrix A is first constructed to describe the label relationship between the source domain mechanical equipment rolling bearing data samples and the intermediate samples. The matrix elements are 0 or 1, indicating inconsistent or consistent labels respectively. Then, according to the relationship matrix A, the cost function C1 is designed to calculate the transmission cost of the rolling bearing data samples of the source domain mechanical equipment to the intermediate samples, that is: where a j,k represents the label relationship between the source domain mechanical equipment rolling bearing data sample and the intermediate sample, bc represents the intermediate sample, ε represents a very small positive number, and j, k represents the index of the position; using the cost function C1, the optimal transmission divergence from the source domain mechanical equipment rolling bearing data to the intermediate sample set is calculated to evaluate the difficulty of sample migration, that is: Where L da represents the OT divergence value during the transmission process, P represents the data distribution of rolling bearings of mechanical equipment in the source domain, and G bc represents the intermediate distribution, Γ represents the transmission strategy from the source domain mechanical equipment rolling bearing data to the center of gravity of the intermediate distribution, which is optimized through this objective function, namely: Where Π(P s:i ,G bc ) represents the joint distribution space of the rolling bearing data distribution and the intermediate distribution of the source domain mechanical equipment, and must meet the following constraints: where u k and v j They are the marginal distributions of the joint distribution, and the initialization values ​​are and Solving Γ by the Sinkhorn-Knopp fixed-point iteration algorithm s:i→bc ; Finally, the adapted samples and corresponding labels are used to train the local diagnosis model of the rolling bearing data of mechanical equipment in each source domain, namely: Among them, γ is a trade-off parameter for balancing the classification loss function and the target distribution adaptation; then the Adam optimizer is used to update the model parameters of the rolling bearing data of mechanical equipment in each source domain, and the updated model parameters are uploaded to the network server.

Images