A Rolling Bearing Fault Diagnosis Method and System under Different Working Conditions Based on Federated Feature Transfer Learning

Abstract

The present invention provides a rolling bearing fault diagnosis method and system under different working conditions based on federated feature transfer learning, which relates to the technical field of rolling bearing fault diagnosis. The technical key points of the present invention include: performing wavelet transform on the time-domain vibration data of the rolling bearing to obtain a time-frequency spectrogram, using the prior labeled common data as the source domain, and the multi-user unlabeled island privacy data as the target domain; introducing a multi-representation feature extraction structure to improve the original residual network, extracting the multi-representation features of the source domain and the target domain, and respectively constructing multi-user local models; using the model compression idea of the deep neural network to improve the parameter transfer strategy in the federated transfer learning framework, enhancing the security of the federated framework and reducing the communication overhead; constructing a federated global model on the server side that can be used for rolling bearing fault diagnosis under different working conditions. The present invention can integrate the knowledge of island data without multi-user data sharing, and has high accuracy and strong generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault diagnosis, and particularly relates to a rolling bearing fault diagnosis method and system under different working conditions based on federated feature transfer learning. Background Technique

[0002] Rolling bearings are one of the most critical components in rotating machinery and are widely used in many fields. In practice, rolling bearings usually work under complex working conditions, which leads to the vibration characteristics of the bearings changing with the change of working conditions. [1] Therefore, the effective diagnosis of rolling bearing faults under different working conditions can effectively prevent accidents from occurring, which is of great significance for ensuring the efficient and healthy operation of rotating machinery. [2]

[0003] In recent years, deep learning has been increasingly widely used in the field of intelligent fault diagnosis of rotating machinery. It has the ability to adaptively extract data features and can provide end-to-end diagnosis. Reference [3] proposed a multi-scale convolutional neural network to achieve fault diagnosis of rolling bearings in a large-noise environment. Reference [4] directly input the collected rolling bearing vibration signals into a bidirectional long short-term memory network model to automatically extract the fault features of rolling bearings and perform diagnosis. Reference [5] proposed a method combining a deep convolutional neural network and an improved random forest algorithm to effectively diagnose bearing faults. Reference [6] proposed a deep convolutional neural network with a first-layer multi-scale convolutional kernel, which uses one-dimensional convolutional kernels of different sizes to extract multi-scale features from the original vibration signals of bearings to achieve intelligent diagnosis of the bearing health state. Reference [7] proposed an overlapping group sparse model and extracted the composite fault features of rolling bearings, and input them into a new deep complex convolutional neural network model to obtain a high diagnostic accuracy.

[0004] Supported by a large amount of labeled training data, traditional deep learning methods have good fault diagnosis accuracy. However, due to reasons such as economic costs and relevant legal regulations, it is usually difficult for each industrial user to collect a sufficient amount of high-quality data, and they usually hope to keep their data locally, thus resulting in data islands and data privacy problems of rolling bearings. Facing these problems, the intelligent diagnosis method of rolling bearings based on deep learning has limitations.

[0005] ​Reference [8] proposed to use the federated learning framework to solve the data privacy and data silo problems that cannot be solved by traditional artificial intelligence methods. Reference [9] proposed a distributed training framework based on federated learning, which uses multi-source power metering data to "train" the global model, realizing the application of federated learning in the power system. Reference

[10] proposed a fault diagnosis method based on federated learning, which aggregates multi-user data knowledge to establish a diagnosis model and verifies the effectiveness of the method through the same working condition rolling bearing data. Reference

[11] added a dynamic verification scheme to the federated learning algorithm, weakened the impact of low-quality data on the performance of the global model, and obtained a higher bearing fault diagnosis accuracy under the same working conditions.

[0006] Applying the federated learning framework to the field of rolling bearing fault diagnosis can effectively solve the data silo and data privacy problems of bearings, but it still requires the premise that the feature distributions of multi-user bearing data are the same and the training data has labels. This setting obviously does not meet the actual situation of complex bearing working conditions and few labeled data in industrial production.

[0007] Transfer learning has the ability to use known knowledge to solve problems in other similar fields, and has received the attention of many relevant scholars. Reference

[12] proposed to combine the deep belief network with the hybrid kernel joint distribution adaptation method to realize the multi-state recognition of bearings under different working conditions. Reference

[13] added an adaptation layer to the residual network model and completed the transfer of fault diagnosis knowledge in the laboratory environment to actual engineering equipment through the domain adaptation method. Reference

[14] proposed a new intelligent diagnosis framework based on classifier differences, which achieved a high accuracy rate on planetary gear data and rolling bearing data in the case of no labels in the target domain. Reference

[15] proposed a deep feature transfer learning method based on sub-domain adaptation and verified the effectiveness of the method on two publicly available rolling bearing fault data sets with different working conditions.

[0008] The above-mentioned feature transfer methods of deep learning can better solve the rolling bearing fault diagnosis problem with distribution differences. Therefore, combining the federated learning framework with the transfer learning idea to realize federated knowledge transfer learning under the differential distribution of multi-user island privacy data is a problem worthy of in-depth study at present.

[0009] Reference

[16] combines domain adaptation technology with federated learning to complete the classification tasks of electroencephalogram and electrocardiogram while ensuring data privacy. Reference

[17] proposes a federated feature transfer learning method for wearable healthcare, constructs a source domain dataset using easily collected public data, and then constructs a user personalized model through domain adaptation methods, improving the accuracy of wearable activity recognition. Reference

[18] constructs a local model using a feature extractor, a reconstructor, and a classifier, obtains shared features of different clients in an adversarial learning manner, and realizes the construction of a global fault diagnosis model for rolling bearings.

[0010] Regarding the problem of fault diagnosis of multi-user privacy data, most existing methods are based on the conditions of constant working conditions and data with the same distribution. For the problem of distribution differences in multi-user data, existing methods cannot ensure data privacy. Summary of the Invention

[0011] Aiming at the problems of large distribution differences in vibration data of rolling bearings under different working conditions, difficulty in obtaining a large number of labeled data, and inability to centrally use isolated privacy data, the present invention proposes a method and system for fault diagnosis of rolling bearings under different working conditions based on federated feature transfer learning to realize fault diagnosis of different fault positions and different fault degrees of rolling bearings under different working conditions.

[0012] According to one aspect of the present invention, there is provided a method for fault diagnosis of rolling bearings under different working conditions based on federated feature transfer learning, the method comprising the following steps:

[0013] (1) Construction of the dataset

[0014] Select the publicly available and known-labeled rolling bearing data under a certain working condition to construct a source domain public dataset, and multiple users respectively construct target domain datasets using their respective local unlabeled rolling bearing data; wherein, the working conditions of the rolling bearings in the source domain public dataset and the multiple target domain datasets are different from each other;

[0015] (2) Construction of the local model

[0016] Each user among the multiple users uses an improved residual network to extract multiple features of the same sample in different representations from the target domain dataset and the source domain public dataset, and performs feature alignment in multiple feature spaces at the same time; the objective function of each user's local model is set as the sum of the cross-entropy classification loss and the conditional maximum mean discrepancy loss in different representations; after a certain number of iterative trainings, the objective function converges, and the construction of each user's local model is completed;

[0017] (3) Construction of the federated global fault diagnosis model

[0018] After the local models of each user among multiple users are built, the processed and compressed model parameters are uploaded to the central server. The central server aggregates the model parameters and distributes them to each user. Each user uses the aggregated model parameters to update the local model. The above process is called one round of federated iteration. After a predetermined number of federated iterations, a federated global model that can finally be used for rolling bearing fault diagnosis under different working conditions is obtained on the server side;

[0019] (4) Model testing

[0020] The test datasets of multiple users are respectively input into the federated global model, and the label prediction results of the model for the test data are compared with the true labels to obtain their respective classification accuracies;

[0021] (5) Fault diagnosis

[0022] The data of the rolling bearing to be diagnosed is input into the trained federated global model to obtain the diagnosis result.

[0023] Furthermore, the improvement of the improved residual network in (2) local model construction lies in: using a multi-representation domain feature extraction structure to replace the global pooling layer of the residual network; the improved residual network uses a multi-representation domain feature extraction structure with multiple sub-structures h1(.)…h n (.) to extract the features under multiple representations of the data:

[0024] (h1 og)(X),..., (h n og)(X)

[0025] Among them, each sub-structure is composed of convolutional layer pooling layers with different numbers and different structures; o represents a composite function; g represents the process of the network converting the input image from a high pixel to a low pixel; X represents the set of input data of the residual network; n represents the number of sub-structures.

[0026] Furthermore, in (2) local model construction, the feature alignment of the network is achieved by minimizing the domain adaptation loss. The domain adaptation loss based on multi-representation features is expressed as:

[0027]

[0028] Among them, represents an estimator of the difference between the source domain and target domain feature distributions, h i represents a sub-structure of the multi-representation domain feature extraction structure, X s and X t respectively represent the sets of source domain input data and target domain input data, and f represents the objective function.

[0029] Furthermore, in (2) local model construction, the formula for calculating the conditional maximum mean discrepancy loss is as follows:

[0030]

[0031] Among them, given a source domain sample set containing n s labeled samples and a target domain containing n unlabeled samples t is the label corresponding to the i-th source domain sample respectively represent the number of samples belonging to the c-th class in the source domain and the target domain,

[0032]

[0033]

[0034]

[0035]

[0036] where ω t+1 , are respectively the parameters of the global model in the (t + 1)-th iteration and the local model parameters of the k-th user, and n represents the number of users.

[0035] According to another aspect of the present invention, there is provided a rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning, and the system includes:

[0036] A data set construction module, configured to select publicly available and known-labeled rolling bearing data under a certain working condition to construct a source domain common data set, and multiple users respectively construct target domain data sets by using their respective local unlabeled rolling bearing data; among them, the rolling bearing working conditions of the source domain common data set and the multiple target domain data sets are different from each other;

[0037] A local model construction module, configured to enable each user among multiple users to extract multiple features of the same sample under different representations from the target domain data set and the source domain common data set, and perform feature alignment in multiple feature spaces simultaneously; the objective function of the local model of each user is set as the sum of the cross-entropy classification loss and the conditional maximum mean difference loss under different representations; after a certain number of iterative trainings, the objective function converges, and the construction of the local model of each user is completed;

[0038] The Federal Global Fault Diagnosis Model Construction Module is used to upload the compressed model parameters to the central server after the local model construction of each user among multiple users is completed. The central server aggregates the model parameters and distributes them to each user, and each user updates the local model using the aggregated model parameters. The above process is called one federal iteration. After a predetermined number of federal iterations, a federal global model that can finally be used for the fault diagnosis of rolling bearings under different working conditions is obtained on the server side;

[0039] The Model Testing Module is used to input the test data sets of multiple users into the federal global model respectively, compare the label prediction results of the model for the test data with the true labels, and obtain their respective classification accuracies;

[0040] The Fault Diagnosis Module is used to input the data of the rolling bearing to be diagnosed into the trained federal global model to obtain the diagnosis result.

[0041] Further, the improvement of the improved residual network in the local model construction module lies in: using a multi-representation domain feature extraction structure to replace the global pooling layer of the residual network; the improved residual network uses a multi-representation domain feature extraction structure with multiple sub-structures h1(.)…h n (.) to extract the features under the multi-representations of the data:

[0042] (h1 og)(X),..., (h n og)(X)

[0043] where each sub-structure is composed of convolutional layer pooling layers with different numbers and different structures; o represents a composite function; g represents the process of the network converting the input image from a high pixel to a low pixel; X represents the set of input data of the residual network; n represents the number of sub-structures.

[0044] Further, the feature alignment of the network in the local model construction module is achieved by minimizing the domain adaptation loss. The domain adaptation loss based on multi-representation features is expressed as:

[0045]

[0046] where, represents an estimator of the difference between the source domain and target domain feature distributions, h i represents a sub-structure of the multi-representation domain feature extraction structure, X s and X t represent the sets of source domain input data and target domain input data respectively, and f represents the objective function.

[0047] Further, the calculation formula of the conditional maximum mean discrepancy loss in the local model construction module is as follows:

[0048]

[0049] Among them, given a source domain sample set containing n s labeled samples and a target domain containing n t unlabeled samples is the label corresponding to the i-th source domain sample respectively represent the number of samples belonging to the c-th class in the source domain and the target domain, respectively represent the number of samples belonging to the c-th class in the source domain and the target domain, indicates that the corresponding sample belongs to the c-th class of the source domain, indicates that the j-th unlabeled target domain sample belongs to the c-th class of the target domain, H represents a reproducing Hilbert space with a characteristic kernel; represents a feature mapping that maps the original sample data to H.

[0050] Furthermore, in each federated iteration process of the federated global fault diagnosis model construction module, the aggregation method for the central server to aggregate model parameters is the federated averaging algorithm, which is expressed as:

[0051]

[0052] where ω t+1 , are the parameters of the global model and the local model parameters of the k-th user in the (t + 1)-th iteration respectively, and n represents the number of users.

[0053] The beneficial technical effects of the present invention are:

[0054] Based on the idea of "data does not move and the model moves, data is available but not visible" in federated learning, the present invention proposes a federated feature transfer learning framework and a method for rolling bearing faults under different working conditions based on federated multi-representation domain adaptation. This method uses a multi-representation feature extraction structure to improve the residual network, constructs a user local model, and improves the parameter transfer strategy in the federated transfer learning framework to compress the local model, and finally realizes the fault diagnosis of different fault positions and different fault degrees of rolling bearings under different working conditions. The present invention can establish an effective fault diagnosis model in the case of privacy distribution of multi-user bearing data islands and scarce labeled data under some working conditions. Experimental verification shows that the average fault diagnosis accuracy of the method proposed by the present invention can reach 97.6%, and it has strong generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The present invention can be better understood by referring to the description given below in conjunction with the accompanying drawings. The accompanying drawings, together with the following detailed description, are included in this specification and form a part of this specification, and are used to further illustrate the preferred embodiments of the present invention and explain the principles and advantages of the present invention.

[0056] Figure 1 is a diagram of the federated learning framework;

[0057] Figure 2 is a structure diagram of a convolutional neural network;

[0058] Figure 3 is a schematic diagram of the structure of the original residual block;

[0059] Figure 4 is a structure diagram of the multi-representation domain adaptation network in the embodiment of the present invention;

[0060] Figure 5 is a flowchart of a rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning in the embodiment of the present invention;

[0061] Figure 6 is the rolling bearing test bench in the embodiment of the present invention;

[0062] Figure 7 is a schematic diagram of the wavelet spectrum of the experimental data in the embodiment of the present invention;

[0063] Figure 8 is a visualization diagram of the features of the single-representation domain adaptation network in the embodiment of the present invention;

[0064] Figure 9 is a visualization diagram of the features of the multi-representation domain adaptation network in the embodiment of the present invention;

[0065] Figure 10 is a diagram of the relationship between the rate value and the average accuracy in the embodiment of the present invention;

[0066] Figure 11 is a diagram of the experimental results of the comparison between the method of the present invention and single-user modeling;

[0067] Figure 12 is a schematic diagram of the confusion matrix of user 1 in the embodiment of the present invention;

[0068] Figure 13 is a schematic diagram of the confusion matrix of user 2 in the embodiment of the present invention;

[0069] Figure 14 is a diagram of the experimental results of the comparison of working condition generalization in the embodiment of the present invention;

[0070] Figure 15 is a diagram of the experimental results of the comparison with NICE bearings in the embodiment of the present invention. Detailed implementation manners

[0071] To enable those skilled in the art to better understand the solution of the present invention, the exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are only a part of the embodiments or examples of the present invention, rather than all of them. All other embodiments or examples obtained by those of ordinary skill in the art based on the embodiments or examples in the present invention without creative work shall fall within the scope of protection of the present invention.

[0072] An embodiment of the present invention provides a rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning. The method will be described in detail below.

[0073] 1. Related Theories of Federated Learning

[0074] 1.1 Federated Learning

[0075] Federated learning is a distributed machine learning framework with privacy protection and security encryption technologies. It aims to train a high-performance federated learning model through the isolated data of multiple parties, reduce the disadvantages of data privacy leakage and excessive data volume requirements during the training of traditional artificial intelligence models, and provide a new idea for the training of artificial intelligence models. The federated learning framework is generally as Figure 1 shown. As can be seen from Figure 1 , different from traditional machine learning, federated learning mainly has three characteristics: the data of each party is retained locally, without privacy leakage and without violating regulations; multiple users jointly use the data to establish a virtual federated global model, getting "rich" together and benefiting together; under the system of federated learning, the identities and statuses of each user are the same.

[0076] In the federated learning framework, the most commonly used parameter aggregation method on the server side is the federated averaging algorithm (FederatedAveraging Algorithm, FedAvg). Assuming that a total of n users participate in the federated learning framework, FedAvg can be expressed as:

[0077]

[0078] where ω t+1 , are the parameters of the global model and the local model parameters of the k-th user in the (t + 1)-th iteration respectively.

[0079] 1.2 Federated Transfer Learning

[0080] Although federated learning theoretically breaks the deadlocks such as data silos and data privacy and security in traditional machine learning, there are still some problems that cannot be ignored in practical applications. For example, it requires that the data of multiple users are in the same distribution. When the data distributions of multiple users vary greatly, directly using the federated learning framework for modeling may produce a worse model than individual training. Transfer learning can apply the knowledge learned in a certain domain or task to different but related domains or problems. Therefore, federated transfer learning combines the idea of transfer learning with the federated learning framework and provides a solution for the scenario of Equation (2):

[0081]

[0082] where S i 、S j and I i 、I j represent the sample space and feature space of user i and user j respectively, and D i 、D j are the isolated privacy data sets owned by user i and user j.

[0083] The federated transfer learning framework can jointly establish a global model based on the knowledge of isolated privacy data of multiple users under different distributions, complete the "transfer from the big data domain to the small data domain", and finally perform label prediction (regression prediction) on the samples in the target domain.

[0084] 1.3 Model Compression

[0085] Although the federated learning framework does not require direct uploading of users' privacy data, the model parameters exchanged during the training process are also sensitive information and require certain technologies to enhance their privacy. Otherwise, once attacked, attackers can use the model parameters to infer valuable privacy information

[19] 。In addition, when the number of participating users is too large or the number of model parameters to be uploaded is too large, the network communication overhead generated during the transmission process may far exceed expectations

[20] 。Therefore, security and communication efficiency are also one of the criteria for judging the overall performance of the algorithm.

[0086] Since most deep neural network models have the problem of redundant weight parameters during model training, that is, among all the parameters, only a part play an important role in the model performance. Therefore, using the idea of model compression algorithms, only part of the parameters of the user's local model are transmitted during the transmission process, that is, the user only uploads all the parameters of the layers with relatively large intra-layer parameter changes during each model update. On the one hand, as the amount of transmitted data decreases, the bandwidth consumption of network transmission can be effectively reduced; on the other hand, since only part of the parameter data is transmitted, even if an attacker obtains this part of the data, it is difficult to use model inversion attacks to reverse the original data due to the lack of global information, thus effectively improving the security of the system. The degree of model compression is determined by the model layer compression rate (rate). For example, rate = 0.95 means that only the parameters of the first 95% of the layers with relatively large changes in this iteration update are transmitted.

[0087] 2. Improved Residual Network

[0088] Domain adaptation problem is one of the research contents of transfer learning, which can solve the problem of inconsistent feature distributions in the feature space and class space under unsupervised conditions. Traditional domain adaptation methods generally align the feature distributions under a single representation, which may only contain partial information. For example, it only contains partial saturation, brightness, and hue information of the picture. Therefore, the feature extraction of a single representation may provide false or distorted real images.

[21] In order to make the feature distributions of the transformed source domain and target domain as similar as possible and extract more globally applicable domain-invariant features, the present invention introduces a multi-representation feature extraction structure to improve the original residual network, extracts the features of the source domain and target domain under multiple representations, and constructs a user local model.

[0089] 2.1 Residual Network and Multi-Representation Domain Adaptation Network

[0090] Convolution Neural Network (CNN) is an artificial deep neural network with multiple hidden layers, having strong feature extraction ability and pattern recognition ability.

[22] Traditional CNN consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and its structure is as Figure 2 shown.

[0091] Residual Network (ResNet) is developed from the traditional CNN network. While retaining the backbone structure of the traditional CNN network, ResNet adds deeper network layers and introduces the gating mechanism in the long short-term memory network, adding an identity mapping unit between each network layer of the CNN, and performing a superposition operation on the identity mapping of the network layer input and its non-linear transformation. The residual block is the basic unit of the residual network, and its structure is as Figure 3 shown.

[0092] The input of the residual block is z, and the output is H(z). The residual refers to the difference between the output value H(z) and the input value z, that is:

[0093] f(z) = H(z) - z (3)

[0094] Due to the addition of the residual block structure, during the training process of the residual network, only the difference f(z) between the inputs and outputs of multiple residual blocks needs to be learned, thus solving the problem of network degradation that occurs as the number of layers of the CNN model increases. During the backpropagation process of the model, the input passes through the identity mapping, directly transmitting information from the input end to the output end of the residual block, ensuring the integrity of information during transmission.

[0095] The output of ResNet can be simply represented by three parts: g(·), h(·), and s(·). The first part g(·) can convert the input image with high pixels into an image with low pixels; the second part h(·) is used to extract features from the low pixels; the third part s(·) is used to map the feature representation to the sample label space as a classifier. Therefore, the ResNet output y = f(x) can be expressed as:

[0096] y = (sohog)(X) (4)

[0097] Among them, o represents the composite function, for example, ((hog)(X) = h(g(X))), and X is the set of input data x of the network.

[0098] The multi-representation domain adaptation network uses a multi-representation domain feature extraction structure to replace the global pooling layer of ResNet to extract the multi-representation features of data. The structure of the multi-representation domain adaptation network is as Figure 4 . From Figure 4 It can be seen that the multi-representation domain adaptation network uses a multi-representation feature extraction structure with multiple sub-structures h1(.)…h n (.) to extract the features under the multi-representation of data:

[0099] (h1og)(X),..., (h n og)(X) (5)

[0100] Among them, each sub-structure is composed of convolutional layers and pooling layers with different numbers and different structures; o represents the composite function; g represents the process of the network converting the input image from high pixels to low pixels; X represents the set of input data of the residual network; n represents the number of sub-structures.

[0101] At this time, the ResNet output y = f(x) can be re-expressed as:

[0102] y = s([(h1og)(X);...; (h nog)(X)]) (6)

[0103] Compared with the features in a single representation, the features extracted by the multi-representation domain adaptation network can cover more information. Aligning the features in the multi-representation of the source domain and the target domain can theoretically obtain better classification results. The feature alignment task of the network can be achieved by minimizing the domain adaptation loss. The domain adaptation loss based on multi-representation features is as follows:

[0104]

[0105] where is an estimator of the difference between the source domain and target domain feature distributions, X s and X t represent the sets of source domain input data x s and target domain input data x t respectively, and f represents the objective function. Based on Equation (6), the overall loss function of the multi-representation domain adaptation network can be expressed as the sum of the classification loss and the domain adaptation loss. The multi-representation domain adaptation network loss function is:

[0106]

[0107] where n s represents the number of source domain samples, J(·, ·) represents the cross-entropy loss function (classification loss), and λ > 0 is the domain adaptation decay term.

[0108] 2.2 Conditional Maximum Mean Discrepancy

[0109] The main idea of conditional distribution adaptation is to reduce the difference in conditional probability densities between the data in two domains, that is:

[0110] minP s (y s |x s ),P t (y t |x t ) (9)

[0111] The present invention uses Conditional Maximum Mean Discrepancy (CMMD) as the domain adaptation loss term to measure the distribution distance between the source domain and the target domain. The calculation formula of CMMD is shown in Equation (10):

[0112]

[0113] where, given a source domain sample set s containing n labeled samples and a target domain t containing n unlabeled samples is the one - hot label corresponding to the \(i\) - th source - domain sample and represents the number of samples belonging to the \(c\) - th class in the source domain and target domain. If it is in the source domain, it means the corresponding sample belongs to the \(c\) - th class of the source domain. If it is in the target domain, it represents that the \(j\) - th unlabeled target - domain sample belongs to the \(c\) - th class of the target domain. \(H\) is a reproducing Hilbert space with a feature kernel; \(\varPhi\) represents a feature mapping that can map the original sample data to \(H\). \(\varPhi\) represents a feature mapping that can map the original sample data to \(H\).

[0114] Since the target domain does not contain labeled samples and the labels (hard labels) predicted by the model for the target - domain data may be incorrect, Equation (7) cannot be directly calculated. However, the output of the deep neural network can be equivalent to a probability distribution:

[0115]

[0116] where, \(p(y = c|x)\) represents the probability that the network input \(x\) i belongs to each class.

[0117] In order to calculate CMMD in the target domain while reducing the error caused by hard - label prediction, probability predictions (soft labels) are used for target - domain samples during model training to calculate Equation (10). By minimizing Equation (10), the difference between the source - domain features and target - domain features under multiple representations is reduced.

[0118] 3. Rolling Bearing Fault Diagnosis Method under Different Working Conditions

[0119] Under different working conditions, the overall flow chart of the rolling bearing fault diagnosis method based on federated multi - representation domain adaptation is as Figure 5 shown. The specific steps are as follows:

[0120] (1) Construction of the dataset

[0121] Select the publicly available and labeled rolling - bearing data under a certain working condition to construct the source - domain common dataset. User 1 and User 2 use their respective local unlabeled rolling - bearing data to construct the target - domain 1 and target - domain 2 datasets respectively. If there is a new user 3, its data is also used as target - domain data. The working conditions of the rolling bearings that make up the source - domain and target - domain datasets are different from each other.

[0122] (2) Local model construction

[0123] Taking User 1 as an example, User 1 uses the multi-representation domain adaptation network to extract the features of the target domain 1 and the source domain data under multiple representations, and performs feature alignment in multiple feature spaces. Among them, the multi-representation features refer to the features of the same sample under different representations extracted by the multi-representation domain adaptation network mentioned above, such as the hue feature, contrast feature, etc. of a picture. Finally, the objective function of the local model 1 is set as the sum of the cross-entropy classification loss and the CMMD loss under different representations. After a certain number of iterative trainings, the objective function converges, and the construction of the local model 1 is completed. The construction process of other user models is the same.

[0124] (3) Construction of the federated global fault diagnosis model

[0125] After the construction of the user local model is completed, the compressed model parameters are uploaded to the central server. The server aggregates the model parameters and distributes them to each user, and the user updates the local model using the aggregated model parameters. The above process is called a single federated iteration. After a predetermined number of federated iterations, the final federated global model that can be used for the fault diagnosis of rolling bearings under different working conditions is obtained on the server side.

[0126] (4) Model testing

[0127] The test data sets of User 1 and User 2 are respectively input into the federated global model, and the label prediction results of the model for the test data are compared with the true labels to obtain their respective classification accuracies.

[0128] (5) Fault diagnosis

[0129] The rolling bearing data to be diagnosed under Working Condition I or Working Condition II of the user can be directly input into the existing federated global fault diagnosis model to obtain the diagnosis result; for the rolling bearing data to be diagnosed under other working conditions of the user, the preliminary diagnosis result can be directly obtained using the existing federated global fault diagnosis model. If there is new data to be diagnosed under other working conditions, it can participate in the federated migration process again, and a new global fault diagnosis model can be established before diagnosis.

[0130] 4. Application and analysis

[0131] 4.1 Experimental data set

[0132] The present invention takes the 6205-2RS deep groove ball bearing under different working conditions as the research object, and the vibration signals of its various states are collected from the bearing test bench

[23] . The rolling bearing test bench is as Figure 6 shown, and mainly includes three parts: a motor, a load, and a control circuit. Among them, an acceleration sensor installed at the driving end of the motor collects the multi-state vibration signals of the rolling bearing at a sampling frequency of 12 kHz.

[0133] In the experiment, the fault feature of the bearing is pitting fault, which is generated by using an electric spark machine to process at three positions: the inner ring, the rolling elements, and the outer ring of the bearing. The pitting diameters at the three fault positions are 0.1778 mm, 0.3556 mm, and 0.5334 mm. Together with the normal state, there are ten types of states in total. For the convenience of subsequent description of the experimental data, the ten types of states of the bearing are now set. The normal state of the bearing is set as N, and the representation methods of the nine fault states in the experimental data are shown in Table 1.

[0134] Table 1 Representation methods of experimental data

[0135]

[0136] The experimental data is collected under a total of 4 different loads of 0 hp, 1 hp, 2 hp, and 3 hp (1 hp ≈ 0.75 kW). According to the different loads, the rotational speed of the motor varies between 1730 rpm and 1797 rpm.

[0137] The corresponding relationship between the working conditions of the bearing operation, the load, and the rotational speed is shown in Table 2.

[0138] Table 2 Corresponding relationship between bearing working conditions, load, and rotational speed

[0139]

[0140] During the data preprocessing process, taking 1024 points as the length of a sample, wavelet transform is performed to obtain the dataset used in the experiment. Taking an OR07 sample at 0 hp as an example, its wavelet spectrum is as Figure 7 shown.

[0141] In the subsequent experiment, a total of 8 transfer tasks are set, and the corresponding working conditions of the datasets used in each task are shown in Table 3. Regarding the problem of isolated island privacy of rolling bearing data, a public dataset of 1000 samples is specifically set. The datasets of User 1 and User 2 are both 100 samples, which is only 1 / 10 of the number of samples in the public dataset. The test set is 1000 samples. Among them, all datasets contain the ten types of states of the rolling bearing.

[0142] Table 3 Composition of datasets used in each task

[0143]

[0144] Taking Task 1 as an example, the source domain data (public dataset) is the bearing data working under Condition A, the dataset of Target Domain 1 (User 1) is the bearing data working under Condition B, and the dataset of Target Domain 2 (User 2) is the bearing data working under Condition C.

[0145] The method proposed in the present invention uses the ReLU function as the activation function, sets the number of local model iterations local_epochs to 100, the number of federated iterations epochs to 50, the learning rate lr to 0.01, and is implemented based on the following assumptions:

[0146] (1) There are two client sides in the federated feature transfer framework, and the classification accuracy of the global fault diagnosis model is the average of the accuracies obtained by the two users inputting their respective test sets into the global fault diagnosis model.

[0147] (2) All client sides have the same fault diagnosis task, that is, different client sides have the same label space.

[0148] (3) Different clients share common data, but the unique local privacy data of each user is not shared.

[0149] The hardware environment used in the experiment: The CPU model is Intel Xeon W-2123; the memory is 32GB; the GPU model is NVIDIA GeForce GTX1080Ti.

[0150] 4.2 Comparative experiment between multi-representation domain adaptation and single-representation domain adaptation

[0151] In order to verify that the multi-representation domain adaptation network can better extract the deep features of data, the bearing data is input into the single-representation domain adaptation network and the multi-representation domain adaptation network respectively, and the t-distribution stochastic neighbor embedding (t-SNE) algorithm is used to perform dimensionality reduction visualization on the deep features extracted by the network.

[24] . Taking migration task 1 as an example, the feature visualization effect diagram is as Figure 8 、 Figure 9 shown.

[0152] Comparative observation Figure 8 、 Figure 9 It can be known that the features extracted by the multi-representation domain adaptation network and the single-representation domain adaptation network both show misclassification phenomena after visualization, but there are still significant differences. After visualization, misclassified samples exist in the L1 to L10 regions of Figure 8 for the deep features of the samples extracted by the single-representation domain adaptation network, and the class boundaries are very blurred. After visualization of the deep features extracted by the multi-representation domain adaptation network, misclassification phenomena only occur in the L1 and L2 regions of Figure 9 , and the class boundaries are obvious. Therefore, it is preliminarily judged that using the multi-representation domain adaptation network to construct the user local model can better extract the deep features of rolling bearings, which is beneficial to improving the performance of the federated global model.

[0153] To further verify that constructing a local model using a multi-representation domain adaptation network can improve the performance of the federated global model, according to the set migration tasks, a rolling bearing fault diagnosis experiment was conducted using the multi-representation domain adaptation network and the single-representation domain adaptation network to construct the federated local model respectively. The results are shown in Table 4.

[0154] Table 4 Fault diagnosis accuracy (%)

[0155]

[0156] Comparing the fault diagnosis accuracies in Table 4, it can be seen that in most migration tasks, the federated global model obtained by constructing the local model using multi-representation domain adaptation has a higher fault diagnosis accuracy, and only migration task 6 is slightly lower. Calculating the average fault diagnosis accuracy of the 8 migration tasks, it can be seen that multi-representation domain adaptation is about 2.3% higher than single-representation domain adaptation. This result shows that constructing the local model using the multi-representation domain adaptation network can better learn the "common" features between user data and public data, and improve the accuracy of rolling bearing fault diagnosis under different working conditions.

[0157] 4.3 Comparative experiment on different model compression degrees

[0158] To explore the influence of model compression processing on the fault diagnosis accuracy of the federated global model, experiments were conducted with different model compression degrees. rate is the layer compression rate, and the smaller the value, the higher the model compression degree and the stronger the security. When rate = 1.00, no model compression processing is performed. The experimental results are shown in Table 5.

[0159] Table 5 Fault diagnosis accuracy (%) for different rate values

[0160]

[0161] As can be seen from Table 5, the fault diagnosis accuracy when rate = 1.00 is not higher than the fault diagnosis accuracies corresponding to the other 4 rate values in all migration tasks. To more intuitively show the influence of the rate value on the performance of the federated global diagnosis model, a dot-line graph of different rate values and the average fault diagnosis accuracy of 8 migration tasks is drawn, as Figure 10 shown.

[0162] From Figure 10It can be clearly seen that as the rate value increases, the average fault diagnosis accuracy shows an upward trend. When rate = 0.80, compared with rate = 1.00, the average accuracy decreases significantly; when rate = 0.85, the average accuracy is only 0.1% lower than that when rate = 1.00. That is to say, selecting an inappropriate layer compression rate will have an adverse effect on the accuracy; selecting an appropriate layer compression rate can improve the security of the federated learning framework and reduce the communication overhead while ensuring the fault diagnosis accuracy. Therefore, considering comprehensively, the rate value is set to 0.85 in the subsequent experiments.

[0163] 4.4 Comparative Experiments between the Proposed Method and Single-User Modeling

[0164] To verify that compared with the situation where a single user directly uses local isolated privacy data and public datasets to construct a deep transfer model, the method proposed in the present invention can effectively improve the fault diagnosis accuracy of rolling bearings under different working conditions, in this section, the proposed method is respectively compared with five deep feature transfer methods, namely Deep Subdomain Adaptation Network for Image Classification (DSAN), Deep Adaptation Network (DAN), Domain adaptive Neural Network (DaNN), Dynamic Adversarial Adaptation Network (DAAN), and Multi-representation Adaptation Network (MRAN). Since the classical deep feature transfer methods cannot directly complete the transfer tasks set in the present invention. Therefore, to simulate the scenario where a user directly uses local isolated data and public data to construct a deep transfer model, with the public dataset as the source domain and the datasets of User 1 and User 2 as the target domains respectively, the average of the diagnosis accuracies of the two experiments is taken as the accuracy of this task. The comparison results are as Figure 11 shown.

[0165] From Figure 11 the experimental results, it can be seen that in the 8 transfer tasks, the fault diagnosis accuracy of the method proposed in the present invention is higher than that of a single user modeling using the other 5 deep feature transfer methods, and when the user only has a small amount of data, the average fault diagnosis accuracy can reach 97.6%. Compared with single-user modeling, the fault diagnosis accuracy is improved by at least 3.2%.

[0166] To more intuitively observe the effectiveness of the method proposed in the present invention in fault diagnosis problems under different working conditions, a multi-class confusion matrix is introduced to analyze the diagnostic results. Due to space limitations, only the fault diagnosis results in Task 1 are taken as an example, and the confusion matrix is drawn as Figure 12 , Figure 13 shown. As can be seen from Figure 12 , among the 1000 target domain test set samples of User 1, misclassification phenomena only occur in 3 types of faults, namely IR07, B07, and OR14, with a total of 10 misclassified samples. As can be seen from Figure 13 , among the 1000 target domain test samples of User 2, misclassification phenomena only occur in 4 types of faults, namely IR21, B07, B21, and OR14, with a total of 8 misclassified samples. Generally speaking, the diagnostic accuracy of most fault types has reached 100%, and the lowest diagnostic accuracy of the fault types with misclassification phenomena is also 95%. The experiment shows that compared with the deep feature transfer method without combining the federated framework, the federated multi-representation domain adaptation method proposed in the present invention can protect user data privacy, aggregate the knowledge of multi-user island data, and effectively solve the fault diagnosis problem of rolling bearings under different working conditions.

[0167] 4.5 Generalization Experiment

[0168] 4.5.1 Working Condition Generalization Experiment

[0169] To further verify that the method proposed in the present invention can cope with different working environments of rolling bearings, the trained federated global diagnostic model can solve the bearing fault diagnosis problems of other working conditions that did not participate in the model training. The method of the present invention is compared with the 5 deep feature transfer methods in Section 4.4 in the working condition generalization experiment. The data sets corresponding to the working conditions of the working condition generalization experiment are shown in Table 6. The generalization data set only participates in the model test and does not participate in the model training, with a total of 1000 samples.

[0170] Table 6 Composition of the Data Set for the Working Condition Generalization Experiment

[0171]

[0172] The experiment is carried out according to the experimental settings in Table 6, and the results are as Figure 14 shown. As can be seen from Figure 14 , the method proposed in the present invention achieved an average fault diagnosis accuracy of 95.3% in the 4 different working condition generalization transfer tasks from Task 9 to Task 12. Compared with the other 5 deep transfer learning methods, the proposed method achieved better fault diagnosis effects in all tasks. Therefore, the method proposed in the present invention can better extract the domain-invariant features of bearings under different working conditions, and can ensure a high diagnostic accuracy even if the bearing data to be diagnosed is data under other working conditions that did not participate in the model training. While solving the problems of rolling bearing data islands and data privacy, it has strong generalization ability.

[0173] 4.5.2 Generalization Experiment on NICE Bearing Dataset

[0174] To verify that the method of the present invention has strong generalization on different datasets, data under different working conditions of NICE bearings are selected as the experimental dataset for the generalization experiment of rolling bearing fault diagnosis.

[0175] The NICE bearing dataset contains four working conditions: E, F, G, and H. Each working condition contains three state vibration data of the bearing: inner race fault (IR), outer race fault (OR), and normal (N). The corresponding relationship between the migration tasks and the data working conditions is set as shown in Table 7.

[0176] Table 7 Composition of NICE Bearing Experimental Dataset

[0177]

[0178] The experiment is carried out according to the migration tasks set in Table 7, and the results are as Figure 15 shown. It can be seen from Figure 15 that in all migration tasks, the fault diagnosis accuracy of the method proposed in the present invention is higher than that of a single user using the 5 deep feature migration methods in Section 4.4 for modeling. The diagnosis accuracies of Task 13 and Task 17 reach 100%, and the average fault diagnosis accuracy is also 98.9%, with at least a 3.4% improvement compared to the user's individual modeling. Through experimental verification, the method proposed in the present invention can effectively solve the fault diagnosis problem of NICE bearings under different working conditions, further proving that the method proposed in the present invention has strong generalization on different datasets.

[0179] The present invention proposes to integrate feature transfer learning into the federated learning framework and introduce a multi-representation feature extraction structure to improve the original residual network, extract more comprehensive common features between the source domain and the target domain, and then construct a federated local feature transfer learning model. Through feature visualization and fault diagnosis accuracy, it is proved that constructing a federated local model using a multi-representation domain adaptation network can obtain a federated global diagnosis model with better performance compared to a single-representation domain adaptation network.

[0180] Based on the idea of deep learning model compression, the present invention improves the parameter transfer strategy of the traditional federated learning framework. Through experimental verification, selecting an appropriate layer compression rate will not seriously affect the performance of the global model. At the same time, only partial local model parameters are transmitted after model compression, which can reduce the communication overhead of the federated learning framework and further ensure the privacy of user data.

[0181] The present invention proposes a rolling bearing fault diagnosis method based on federated multi-representation domain adaptation, which can establish an effective fault diagnosis model in the case of privacy distribution of bearing data islands of multiple users and scarcity of labeled data under some working conditions. Experimental verification shows that the average fault diagnosis accuracy of the proposed method can reach 97.6%. Compared with the case where a single user uses other methods to build a model in the present invention, the fault diagnosis accuracy is improved by at least 3.2%, which proves the effectiveness of the method. The generalization experiment proves that the proposed method also has strong generalization ability.

[0182] Another embodiment of the present invention also proposes a rolling bearing fault diagnosis system based on federated feature transfer learning, which includes:

[0183] A dataset construction module, which is used to select the rolling bearing data with public and known labels under a certain working condition to construct a source domain public dataset, and multiple users use their respective local unlabeled rolling bearing data to construct target domain datasets respectively; wherein, the working conditions of the rolling bearings in the source domain public dataset and multiple target domain datasets are different from each other;

[0184] A local model construction module, which is used for each user among multiple users to extract multiple features of the same sample under different representations from the target domain dataset and the source domain public dataset, and perform feature alignment in multiple feature spaces at the same time; the objective function of each user's local model is set as the sum of the cross-entropy classification loss and the conditional maximum mean difference loss under different representations; after a certain number of iterative trainings, the objective function converges, and the construction of each user's local model is completed;

[0185] A federated global fault diagnosis model construction module, which is used for each user among multiple users to upload the compressed model parameters to the central server after the construction of the local model is completed. The central server aggregates the model parameters and distributes them to each user. Each user updates the local model using the aggregated model parameters. The above process is called a federated iteration. After a predetermined number of federated iterations, a federated global model that can be finally used for rolling bearing fault diagnosis under different working conditions is obtained on the server side;

[0186] A model testing module, which is used to input the test datasets of multiple users into the federated global model respectively, and compare the label prediction results of the model for the test data with the true labels to obtain their respective classification accuracies;

[0187] A fault diagnosis module, which is used to input the rolling bearing data to be diagnosed into the trained federated global model to obtain the diagnosis result.

[0188] In this embodiment, preferably, the improvement of the improved residual network in the local model construction module lies in: using a multi-representation domain feature extraction structure to replace the global pooling layer of the residual network; the improved residual network uses a multi-representation domain feature extraction structure with multiple sub-structures h1(.)…h n Extract features under multi-representations of data using the multi-representation domain feature extraction structure with sub-structures (.)

[0189] (h1 og)(X),..., (h n og)(X)

[0190] where each sub-structure is composed of convolutional layer pooling layers with different numbers and different structures; o represents a composite function; g represents the process of the network converting the input image from high pixels to low pixels; X represents the set of input data of the residual network; n represents the number of sub-structures.

[0191] In this embodiment, preferably, the feature alignment of the network in the local model construction module is achieved by minimizing the domain adaptation loss, and the domain adaptation loss based on multi-representation features is expressed as:

[0192]

[0193] where represents an estimator of the difference between the source domain and target domain feature distributions, h i represents a sub-structure of the multi-representation domain feature extraction structure, X s and X t represent the set of source domain input data and target domain input data respectively, and f represents the objective function.

[0194] In this embodiment, preferably, the calculation formula of the conditional maximum mean discrepancy loss in the local model construction module is as follows:

[0195]

[0196] where, given a source domain sample set s containing n labeled samples and a target domain t containing n unlabeled samples is the label corresponding to the i-th source domain sample , represent the number of samples belonging to the c-th class in the source domain and target domain respectively, indicates that the corresponding sample belongs to the c-th class of the source domain, indicates that the j-th unlabeled target domain sample belongs to the c-th class of the target domain, H represents a reproducing kernel Hilbert space with a feature kernel; represents a feature mapping that maps the original sample data to H.

[0197] In this embodiment, preferably, in each federated iteration process of the federated global fault diagnosis model construction module, the aggregation method of aggregating model parameters on the central server side is the federated average algorithm, which is expressed as:

[0198]

[0199] where ω t+1 and are the parameters of the global model and the local model parameters of the k-th user in the (t + 1)-th iteration respectively, and n represents the number of users.

[0200] The functions of the rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning in this embodiment can be illustrated by the aforementioned rolling bearing fault diagnosis method based on federated feature transfer learning under different working conditions. Therefore, for the parts not detailed in this embodiment, reference can be made to the above method embodiments and will not be elaborated here.

[0201] Although the present invention has been described based on a limited number of embodiments, those skilled in the art in this technical field understand that other embodiments can be conceived within the scope of the present invention thus described. For the scope of the present invention, the disclosure of the present invention is illustrative rather than restrictive, and the scope of the present invention is defined by the appended claims.

[0202] The literature cited in the present invention is as follows:

[0203] [1] CHEN Zhiqiang, DENG Shengcai, CHEN Xudong, et al. Deep neuralnetworks-based rolling bearing fault diagnosis[J]. MicroelectronicsReliability, 2017, 75, 327 - 333.

[0204] [2] Kang Shouqiang, Zou Jiayue, Wang Yujing, et al. Rolling bearing fault diagnosis method under variable load based on unsupervised feature alignment[J]. Proceedings of the CSEE, 2020, 40(1): 274 - 281.

[0205] [3] Xu Zifei, Jin Jiangtao, Li Chun. Rolling bearing fault diagnosis method based on multi-scale convolutional neural network[J]. Journal of Vibration and Shock, 2021, 40(18): 212 - 220.

[0206] [4] Zhao Zhihong, Zhao Jingjiao, Wei Ziyang, et al. Research on rolling bearing fault diagnosis based on BiLSTM[J]. Journal of Vibration and Shock, 2021, 40(1): 95 - 101.

[0207] [5] XU Gaowei, LIU Min, JIANG Zhuofu, et al. Bearing fault diagnosis method based on deep convolutional neural network and random forest ensemble learning[J]. Sensors, 2019, 19(5): 1 - 21.

[0208] [6] FU Liangcheng, LI Zhang, TAO Junyong. An improved deep convolutional neural network with multiscale convolution kernels for fault diagnosis of rolling bearing[C]. IOP Publishing, 2021, 1043(5): 052021.

[0209] [7] AN Fengping, WANG Jianrong. Rolling bearing fault diagnosis algorithm using overlapping group sparse - deep complex convolutional neural network[J]. Nonlinear Dynamics, 2022, 108: 2353 - 2368.

[0210] [8] Yang Qiang. Federated Learning: The Last Mile of Artificial Intelligence[J]. CAAI Transactions on Intelligent Systems, 2020, 15(01): 183 - 186.

[0211] [9] Zheng Kaihong, Xiao Yong, Wang Xin, et al. A Federated Learning Framework for Power Metering Systems[J]. Proceedings of the CSEE, 2020, 40(S1): 122 - 133.

[0212]

[10] LI Zhinong, LI Zedong, LI Yunlong, et al. An intelligent diagnosis method for machine fault based on federated learning[J]. Applied Sciences, 2021, 11(24): 12117.

[0213]

[11] ZHANG Wei, LI Xiang, MA Hui, et al. Federated learning for machinery fault diagnosis with dynamic validation and self-supervision[J]. Knowledge-Based Systems, 2021, 213: 106679.

[0214]

[12] KANG Shouqiang, CHEN Weiwei, WANG Yujing, et al. Method of state identification of rolling bearings based on deep domain adaptation under varying loads[J]. Measurement & Technology, 2020, 14(3): 303 - 313.

[0215]

[13] Lei Yaguo, Yang Bin, Du Zhaojun, et al. Deep transfer diagnosis method for mechanical equipment faults under big data[J]. Chinese Journal of Mechanical Engineering, 2019, 55(07): 1 - 8.

[0216]

[14] JIAO Jinyang, ZHAO Ming, LIN Jing. Unsupervised adversarial adaptation network for intelligent fault diagnosis[J]. IEEE Transactions on Industrial Electronics, 2019, 67(11): 9904 - 9913.

[0217]

[15] Dong Shaojiang, Zhu Peng, Pei Xuewu, et al. Fault diagnosis of rolling bearings under varying working conditions based on sub - domain adaptation[J]. Journal of Jilin University (Engineering and Technology Edition), 2022, 52(02): 288 - 295.

[0218]

[16] JU Ce, GAO Dashan, MANE R, et al. Federated transfer learning for EEG signal classification[C]. EMBC2020.

[0219]

[17] CHEN Yiqing, QIN Xin, WANG Jindong, et al. Fedhealth: A federated transfer learning framework for wearable healthcare[J]. IEEE Intelligent Systems, 2020, 35(4): 83-93.

[0220]

[18] ZHANG Wei, LI Xiang. Federated transfer learning for intelligent fault diagnostics using deep adversarial networks with data privacy[J]. IEEE / ASME Transactions on Mechatronics, 2021, 27(1): 430-439.

[0221]

[19] SHOKRI R, STRONATI M, SONG Congzheng, et al. Membership inference attacks against machine learning models[C]. SP 2017.

[0222]

[20] LI Tian, SAHU A K, TALWALKAr A, et al. Federated learning: challenges, methods, and future directions[J]. IEEE Signal Processing Magazine, 2020, 37(3): 50-60.

[0223]

[21] ZHU Yongchun, ZHUANG Fuzhen, WANG Jindong, et al. Multi-representation adaptation network for cross-domain image classification[J]. Neural Networks, 2019, 119: 214-221.

[0224]

[22] Y, Bottou L, Bengio Y, et al. Gradient-Based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.

[0225]

[23] KENNETH A Loparo. Bearing data center[M / OL]. Cleveland: Case Western Reserve University, 2013. http: / / csegroups.case.edu / bearingdatacenter / home.

[0226]

[24] VAN Der Maaten L, HINTON G. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(11): 2579-2625.

Claims

1. A rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning, characterized in that It includes the following steps: (1) Construction of the dataset Select the rolling bearing data with public and known labels under a certain working condition to construct the source domain common dataset, and multiple users respectively construct the target domain datasets by using their own local unlabeled rolling bearing data; among them, the working conditions of the rolling bearings in the source domain common dataset and multiple target domain datasets are different from each other; (2) Construction of the local model Each user among multiple users uses the improved residual network to extract multiple features of the same sample in different representations from the target domain dataset and the source domain common dataset, and performs feature alignment in multiple feature spaces simultaneously; the objective function of each user's local model is set as the sum of the cross-entropy classification loss and the conditional maximum mean discrepancy loss under different representations; after a certain number of iterative trainings, the objective function converges, and the construction of each user's local model is completed; (3) Construction of the federated global fault diagnosis model After the construction of each user's local model among multiple users is completed, the compressed model parameters are uploaded to the central server, the central server aggregates the model parameters and distributes them to each user, and each user updates the local model by using the aggregated model parameters. The above process is called a single federated iteration. After a predetermined number of federated iterations, the final federated global model for rolling bearing fault diagnosis under different working conditions is obtained on the server side; (4) Model testing Input the test datasets of multiple users into the federated global model respectively, compare the label prediction results of the model for the test data with the true labels, and obtain their respective classification accuracies; (5) Fault diagnosis Input the rolling bearing data to be diagnosed into the trained federated global model to obtain the diagnosis result.

2. A rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning according to claim 1, characterized in that The improvement of the improved residual network in (2) local model construction lies in: using a multi-representation domain feature extraction structure to replace the global pooling layer of the residual network; the improved residual network uses a multi-representation domain feature extraction structure with multiple sub-structures h1(.)…h n (.) to extract features under multiple representations of data: Among them, each sub-structure h i is composed of convolutional layers and pooling layers with different numbers and different structures; represents a composite function; g represents the process of the network converting the input image from high pixels to low pixels; X represents the set of input data of the residual network; n represents the number of sub-structures.

3. A rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning according to claim 2, characterized in that, In (2) the construction of the local model, the feature alignment of the network is realized by minimizing the domain adaptation loss, and the domain adaptation loss based on multi-representation features is expressed as: Among them, represents an estimator of the difference between the source domain and target domain feature distributions, h i represents a sub-structure of the multi-representation domain feature extraction structure, X s and X t respectively represent the sets of source domain input data and target domain input data, and f represents the objective function.

4. A rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning according to claim 3, characterized in that, In (2) the construction of the local model, the calculation formula of the conditional maximum mean discrepancy loss is as follows: Among them, given a source domain sample set containing n s labeled samples and a target domain containing n t unlabeled samples is the label corresponding to the i-th source domain sample , respectively represent the number of samples belonging to the c-th class in the source domain and the target domain, indicates that the corresponding sample belongs to the c-th class of the source domain, indicates that the j-th unlabeled target domain sample belongs to the c-th class of the target domain, and H represents a reproducing Hilbert space with a characteristic kernel; represents a feature mapping that maps the original sample data to H.

5. A rolling bearing fault diagnosis method under different working conditions based on federated feature transfer learning according to claim 4, characterized in that In (3) the construction of the federated global fault diagnosis model, in each federated iteration process, the aggregation method of the central server to aggregate the model parameters is the federated average algorithm, which is expressed as: where ω t+1 and are the parameters of the global model and the local model parameters of the k-th user in the (t + 1)-th iteration, respectively, and n represents the number of users.

6. A rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning, characterized in that, It includes: The dataset construction module is used to select the rolling bearing data with public and known labels under a certain working condition to construct the source domain common dataset, and multiple users respectively construct the target domain datasets by using their own local unlabeled rolling bearing data; among them, the working conditions of the rolling bearings in the source domain common dataset and multiple target domain datasets are different from each other; The local model construction module is used for each user among multiple users to use the improved residual network to extract multiple features of the same sample in different representations from the target domain dataset and the source domain common dataset, and perform feature alignment in multiple feature spaces simultaneously; the objective function of each user's local model is set as the sum of the cross-entropy classification loss and the conditional maximum mean discrepancy loss under different representations; after a certain number of iterative trainings, the objective function converges, and the construction of each user's local model is completed; The Federal Global Fault Diagnosis Model Construction Module is used to upload the compressed model parameters to the central server after the local model construction of each user among multiple users is completed. The central server aggregates the model parameters and distributes them to each user, and each user updates the local model using the aggregated model parameters. The above process is called one round of federated iteration. After a predetermined number of rounds of federated iteration, the final federal global model for rolling bearing fault diagnosis under different working conditions is obtained on the server side; The Model Testing Module is used to input the test data sets of multiple users into the federal global model respectively, compare the label prediction results of the model for the test data with the true labels, and obtain their respective classification accuracies; The Fault Diagnosis Module is used to input the data to be diagnosed of the rolling bearing into the trained federal global model to obtain the diagnosis result.

7. A rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning according to claim 6, characterized in that, The improvement of the improved residual network in the local model construction module lies in: using a multi-representation domain feature extraction structure to replace the global pooling layer of the residual network; the improved residual network uses a multi-representation domain feature extraction structure with multiple sub-structures h1(.)…h n (.) to extract features under multiple representations of data: Among them, each sub-structure h i is composed of convolutional layers and pooling layers with different numbers and different structures; represents a composite function; g represents the process of the network converting the input image from a high pixel count to a low pixel count; X represents the set of input data of the residual network; n represents the number of sub-structures.

8. A rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning according to claim 7, characterized in that The feature alignment of the network in the Local Model Construction Module is achieved by minimizing the domain adaptation loss. The domain adaptation loss based on multi-representation features is expressed as: Among them, denotes an estimator of the difference between the source domain and target domain feature distributions, h i denotes a sub-structure of the multi-representation domain feature extraction structure, X s and X t respectively denote the sets of source domain input data and target domain input data, and f denotes the objective function.

9. A rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning according to claim 8, characterized in that, The calculation formula of the conditional maximum mean discrepancy loss in the Local Model Construction Module is as follows: Among them, given a source domain sample set containing n s labeled samples and a target domain containing n t unlabeled samples is the label corresponding to the i-th source domain sample , respectively represent the number of samples belonging to the c-th class in the source domain and the target domain, indicates that the corresponding sample belongs to the c-th class of the source domain, indicates that the j-th unlabeled target domain sample belongs to the c-th class of the target domain, and H represents a reproducing Hilbert space with a characteristic kernel; represents a feature mapping that maps the original sample data to H.

10. A rolling bearing fault diagnosis system under different working conditions based on federated feature transfer learning according to claim 9, characterized in that, In each round of federated iteration in the Federal Global Fault Diagnosis Model Construction Module, the aggregation method for the central server to aggregate the model parameters is the federated averaging algorithm, which is expressed as: Among them, ω t+1 and are the parameters of the global model and the local model parameters of the k-th user in the (t + 1)-th iteration respectively, and n represents the number of users.

Images