CN116226784A - Federal domain adaptive fault diagnosis method based on statistical feature fusion - Google Patents

Abstract

The invention discloses a federal domain adaptive fault diagnosis method based on statistical feature fusion, wherein a source domain client receives statistical features of a target domain client and statistical features of active domain data and respectively performs data standardization on a source domain data set; the central server sends the fault diagnosis model to all source domain clients, and then the source domain clients train the received global model in the source domain clients by using two standardized source domain data as input; the trained models are returned to a central server, and the central server carries out global model aggregation on the received models of all source domain clients; when the training round number reaches a set value, the training is finished to obtain a final fault diagnosis model, the final fault diagnosis model is sent to a target domain client, the target domain client obtains mechanical fault data to be diagnosed, and a fault diagnosis result is obtained through the fault diagnosis model. The invention utilizes the statistical distribution difference of the data between the target domain and the source domain to improve the generalization capability of the fault diagnosis model in the target domain.

Description

Federated Domain Adaptive Fault Diagnosis Method Based on Statistical Feature Fusion

technical field

The invention relates to fault diagnosis technology, in particular to a federal domain adaptive fault diagnosis method based on statistical feature fusion.

Background technique

Rotating machinery data usually comes from equipment with different models and working conditions, and different operating and use environments. The fault diagnosis model trained with these data has low prediction accuracy and poor generalization ability for new data.

The document "Toward Secure Data Fusion in Industrial IoT Using Transfer Learning" proposes that domain generalization and domain adaptation methods in transfer learning can solve the domain drift problem by aligning the data feature space. The document "Deep Domain Generalization Combining A Priori Diagnosis Knowledge Toward Cross-Domain Fault Diagnosis of Rolling Bearing" proposes a rolling bearing fault diagnosis method based on domain generalization. Under the condition that the target domain has only healthy samples, the method eliminates the potential differences among multiple domains and realizes efficient fault diagnosis. The document "Conditional Adversarial Domain Generalization With a Single Discriminator for Bearing Fault Diagnosis" proposes a conditional adversarial domain generalization method with a discriminator, the purpose of which is to extract domain invariant features from data with different working conditions and extend these features to new in the fault data. To achieve conditional adversarial training, a new conditional adversarial strategy is designed, that is, the feature extractor can allow the discriminator to distinguish fault categories, but not domains, to better confuse the discriminator and generalize features. The document "Intelligent FaultIdentification Based on Multisource Domain Generalization Towards ActualDiagnosis Scenario" proposes a new intelligent fault identification method based on multisource domains. The method uses local Fisher discriminant analysis, describing the discriminative structure of each source domain as a point on a Grassmann manifold. By preserving the intra-class local structure, local Fisher discriminant analysis can learn effective discriminators from multimodal fault data. The document "ANewMultiple Source Domain Adaptation Fault Diagnosis Method Between DifferentRotating Machines" proposes a migration learning method based on multi-source domain adaptation. The method uses a multi-adversarial learning strategy to obtain domain-aligned feature representations while being discriminative to the target domain. The literature "Deep AdversarialDomain Adaptation Model for Bearing Fault Diagnosis" proposes a deep adversarial domain adaptation model for rolling bearing fault diagnosis. This model constructs an adversarial domain adaptation network to solve the problem of inconsistent feature distribution of source and target domains.

Data is the basis of deep learning fault diagnosis algorithms. In order to ensure the effectiveness of deep learning, it is necessary to aggregate and use as much data as possible. In order to safely and effectively aggregate data from different clients and solve the problem of "data islands" in the process of deep learning, federated learning emerged as the times require. In federated learning, the learning task is solved in the form of loose federation by multiple participating devices (ie clients) under the coordination of the central server. Federated learning uses data from different clients to collaboratively train models, but due to different working conditions or models, these data usually have domain drift problems, so federated transfer learning has attracted more and more attention from researchers.

The document "Federated Transfer Learning for Intelligent Fault Diagnostics Using Deep Adversarial Networks With Data Privacy" proposes a federated transfer learning method for fault diagnosis, which designs different network model structures for different edges, and uses deep adversarial learning to achieve federation communication. The document "Data privacy preserving federated transfer learning inmachinery fault diagnostics using prior distributions. Structural Health Monitoring" proposes a federated transfer learning method for mechanical fault diagnosis. This method proposes to use prior distributions to indirectly address the domain drift problem and perform fault diagnosis by extracting domain-invariant features of different users.

Although federated transfer learning has achieved comparable performance to transfer learning in fault diagnosis, existing federated domain adaptation methods usually need to transmit feature data from all clients to the central server to achieve feature space alignment, which increases the distance between the client and the model. communication cost between them. In addition, existing federated domain adaptation methods lack research on feature fusion methods for client-side multi-sensor input signals.

Contents of the invention

In order to solve the deficiencies in the prior art, the object of the present invention is to provide a federal domain adaptive fault diagnosis method based on statistical feature fusion.

For realizing the purpose of the present invention, the technical scheme adopted in the present invention is:

A federal domain adaptive fault diagnosis method based on statistical feature fusion, including steps:

(1) The target domain client sends the statistical features of unmarked fault data to the central server, and the central server sends these statistical features to all source domain clients; the source domain client uses the received statistical features and its own source domain The statistical characteristics of the data perform data standardization on the source domain data set respectively;

(2) The fault diagnosis model of the central server includes a feature extraction network and a classification network. The central server sends the global model to all source domain clients, and then the source domain clients use two standardized source domain data as input, based on correlation The alignment method trains the received global model in the source domain client;

(3) After a round of model training for all source domain clients, the trained models will be sent to the central server, and the central server will perform global model aggregation on the received models of all source domain clients; and then send them to all source domain clients The next round of training is carried out in the terminal. When the number of training rounds reaches the set value, the training task ends and the final fault diagnosis model is obtained;

(4) In the test phase, the central server sends the trained fault diagnosis model to the target domain client, and the target domain client obtains the mechanical fault data to be diagnosed, and obtains the fault diagnosis result through the fault diagnosis model.

Further, in step (1), the statistical features are mean and standard deviation.

Further, in step (2), in the feature extraction network, the statistical feature exchange of different feature extraction channels is used to standardize the output features of this channel, and the exchanged standardized features are used as the input of the classification network.

Further, in step (2), the global model received by the source domain client includes a feature extraction network and a classification network.

Further, the received global model is trained in the source domain client based on the correlation alignment method. Specifically, the source domain client also includes a feature space metric, and the feature space metric uses the correlation alignment method to constrain the relationship between different features of the same sensor data. distance between.

Furthermore, the input of the feature space metric is the standardized feature achieved by the statistical feature exchange of different feature extraction channels.

Further, in step (3), the global model on the central server is aggregated as,

和N个源域客户端上的N个全局模型的分类器/>

平均化后，更新中心服务器上的全局模型的特征提取器/>

和分类器/>

其中，/>

为第i个源域客户端上的全局模型的特征提取器，/>

为第i个源域客户端上的全局模型的分类器。The feature extractors of N global models on N source domain clients

and classifiers of N global models on N source domain clients />

After averaging, update the feature extractor of the global model on the central server />

and classifier />

where, />

is the feature extractor of the global model on the i-th source domain client, />

is the classifier of the global model on the i-th source domain client.

Further, in step (4), the model received by the target domain client includes a feature extraction network and a classification network.

Furthermore, the structure of the two channels of the feature extraction network is the same, and the feature extraction network consists of three groups of sequentially connected convolutional layers, regularization layers, linear correction units, and maximum pooling layers; the input of the feature extraction network first enters the first group The convolutional layers of the first set of max pooling layers are connected to the second set of convolutional layers, and the second set of max pooling layers are connected to the third set of convolutional layers.

Further, the classification network consists of sequentially connected flattening layer, first fully connected layer, regularization layer, modified linear unit layer, second fully connected layer and softmax function layer.

The beneficial effect of the present invention is that, compared with the prior art, the federal domain adaptive fault diagnosis method based on statistical feature fusion of the present invention utilizes the difference in statistical distribution of data between the target domain and the source domain, and the statistics of multi-sensor signals of source domain data The difference in distribution improves the generalization ability of the fault diagnosis model in the target domain. In the present invention, only the statistical features of the target domain data are transmitted to each source domain client, which reduces the communication burden. In addition, at the source domain client, the statistical features of multi-sensor data features will be exchanged with each other to achieve feature domain space enhancement.

Description of drawings

Fig. 1 is the flowchart of the federal domain adaptive fault diagnosis method based on statistical feature fusion according to the present invention;

Fig. 2 is a schematic diagram of feature extraction network;

Fig. 3 is a schematic diagram of a client training model.

Detailed ways

The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The following examples are only used to illustrate the technical solutions of the present invention more clearly, but not to limit the protection scope of the present application.

As shown in Figure 1, the federal domain adaptive fault diagnosis method based on statistical feature fusion described in the present invention includes the following steps:

(1) First, standardize the source domain data. Specifically, the target domain client sends the statistical characteristics (mean and standard deviation) of unmarked fault data to the central server, and the central server sends these statistical characteristics to all source domain clients end; the source domain client uses the received statistical characteristics and the statistical characteristics of its own source domain data to standardize the source domain data set;

源域客户端k中的数据集为

其中，m_k为源域客户端k中包含的数据样本数，/>

为第k个源域客户端中的第j个数据样本，/>

为第k个源域客户端中的的第j个数据样本的标签。Let the N data sets in the N source domain clients be

The data set in source domain client k is

Among them, m _k is the number of data samples contained in the source domain client k, />

is the jth data sample in the kth source domain client, />

is the label of the jth data sample in the kth source domain client.

在第N+1个客户端中，n_t为目标域样本数，/>

为目标域客户端中的第i个样本。设目标域数据集的均值为/>

标准差/>

Datasets for target domain clients

In the N+1th client, n _t is the number of samples in the target domain, />

is the i-th sample in the target domain client. Let the mean of the target domain dataset be />

standard deviation />

The normalized target domain data is:

从μ_t和δ_t随机选取一组μ_i,δ_i，对/>

进行标准化，即：The target domain client sends μ _t and δ _t to the central server, and the central server sends them to each source domain client. In the kth source domain client,

Randomly select a set of μ _i , δ _i from μ _t and δ _t , for />

to standardize, i.e.:

为第k个源域客户端使用目标域的统计特征标准化的数据。

Data normalized for the kth source domain client using the statistical characteristics of the target domain.

的均值和标准差分别为/>

使用/>

和/>

标准化/>

的结果为：In addition, set

The mean and standard deviation of

use />

and />

Standardization />

The result is:

为第k个源域客户端使用自有源域数据的统计特征标准化的数据。

Data normalized using the statistical characteristics of the own source domain data for the kth source domain client.

和/>

用于在源域客户端中训练故障诊断模型。Normalized source domain data

and />

It is used to train the fault diagnosis model in the source domain client.

(2) The source domain client uses two standardized source domain data as input, and trains the fault diagnosis model in the source domain client based on the Correlation Alignment (CORAL) method;

The fault diagnosis model of the central server includes a feature extraction network (feature extractor) and a classification network (classifier).

As shown in Figure 3, the model received by the source domain client includes the feature extraction network and the classification network; the source domain client also includes the feature distance measurement method CORAL, which is used to limit the distance between different features of the same sensor data.

和/>

在训练模型时，中心服务器首先将/>

和/>

发送给所有源域客户端，然后源域客户端k使用其本地训练数据集训练接收到的全局模型。The feature extraction network and classification network on the central server are denoted as

and />

When training the model, the central server first sends the />

and />

Sent to all source domain clients, then source domain client k uses its local training dataset to train the received global model.

和/>

将分别作为/>

和/>

的输入。

由电流数据/>

和振动数据/>

组成，同样地，/>

由电流数据/>

和振动数据/>

组成。In source domain client k,

and />

will be respectively as />

and />

input of.

By current data />

and vibration data />

Composition, likewise, />

By current data />

and vibration data />

composition.

有两个特征提取通道，通道1的输入为电流数据，通道2的输入为振动数据。通道1和通道2输出的特征将使用两个通道交换的统计特征实现标准化。通道1和通道2的输出的特征交换标准化后将作为分类网络/>

和CORAL的输入。As shown in Figure 2, in the feature extraction network, the feature domain adaptability is enhanced by exchanging the statistical features of different sensor data. Feature Extraction Network

There are two feature extraction channels, the input of channel 1 is current data, and the input of channel 2 is vibration data. The features output by channel 1 and channel 2 will be normalized using the statistical features exchanged by the two channels. The feature exchange of the output of channel 1 and channel 2 after normalization will be used as a classification network />

and CORAL input.

经过通道1处理后的输出为/>

经过通道2处理后的输出为/>

经过通道1处理后的输出为/>

经过通道2处理后的输出为/>

的均值和标准差用于标准化/>

的均值和标准差用于标准化/>

同理，/>

的均值和标准差用于标准化/>

的均值和标准差用于标准化/>

标准化后的/>

将作为/>

和CORAL的输入。

The output after processing by channel 1 is />

The output after channel 2 processing is />

The output after processing by channel 1 is />

The output after channel 2 processing is />

The mean and standard deviation of are used for standardization />

The mean and standard deviation of are used for standardization />

In the same way, />

The mean and standard deviation of are used for standardization />

The mean and standard deviation of are used for standardization />

Normalized />

will be used as />

and CORAL input.

是源域客户端k中分类网络的分类损失值，其公式为：

is the classification loss value of the classification network in the source domain client k, and its formula is:

为数据样本j的真实标签，/>

为源域客户端k中分类网络对数据样本j的预测结果。in,

is the real label of data sample j, />

is the prediction result of the classification network on the data sample j in the source domain client k.

In addition, in order to constrain the distance between different features of the same sensor data, the feature distance measurement loss based on Correlation Alignment (CORAL) is used, loss _{1, CORAL} , loss _{2, CORAL} calculation process is as follows:

和C_1,g分别为标准化后的/>

的特征协方差矩阵，/>

和C_2,g分别为标准化后的/>

的特征协方差矩阵；d代表特征的维度，||·||_F代表Frobenius范数。in,

and C _{1, g} are the normalized />

The characteristic covariance matrix of , />

and C _2,g are the normalized />

The feature covariance matrix of ; d represents the dimension of the feature, ||·|| _F represents the Frobenius norm.

Therefore, loss _k contains the weighted sum of classification loss and feature distance metric loss in source domain client k of the source domain dataset, namely:

表示源域客户端k中的分类损失，loss_1,CORAL表示电流数据的特征距离度量损失，loss_2,CORAL表示振动数据的特征距离度量损失，β＝0.01。Among them, loss _k represents the sum of losses,

Indicates the classification loss in source domain client k, loss _{1, CORAL} represents the characteristic distance metric loss of current data, loss _{2, CORAL} represents the characteristic distance metric loss of vibration data, β = 0.01.

(3) After one round of model training for all source domain clients, all the global models trained on the source domain clients will be sent to the central server; the central server averages all received source domain client fault diagnosis models, Get the final fault diagnosis model;

和N个源域客户端上的N个全局模型的分类器/>

平均化后，更新中心服务器上的全局模型的特征提取器/>

和分类器/>

其中，/>

为第i个源域客户端上的全局模型的特征提取器，/>

为第i个源域客户端上的全局模型的分类器。Global model aggregation on the central server, feature extractors of N global models on N source domain clients

and classifiers of N global models on N source domain clients />

After averaging, update the feature extractor of the global model on the central server />

and classifier />

where, />

is the feature extractor of the global model on the i-th source domain client, />

is the classifier of the global model on the i-th source domain client.

In this embodiment, the global model is not further trained on the central server, but is sent to all source domain clients for the next round of training.

(4) In the testing phase, the central server sends the trained fault diagnosis model to the target domain client for fault diagnosis.

When the number of training rounds reaches the set value, the training task ends, and the central server sends the global model to the client in the target domain for fault diagnosis. The client in the target domain obtains the mechanical fault data to be diagnosed, and through the global model sent by the central server, obtains Fault diagnosis results.

The model received by the target domain client includes a feature extraction network and a classification network.

The structure of the two channels of the feature extraction network is the same, and the feature extraction network consists of three groups of sequentially connected convolutional layers, regularization layers, linear correction units, and maximum pooling layers; the input of the feature extraction network first enters the convolutional layer of the first group The first set of max pooling layers is connected to the second set of convolutional layers, and the second set of max pooling layers is connected to the third set of convolutional layers. The number of convolution kernels of the three convolutional layers is 128, and the convolution kernel sizes of the convolutional layers in the first group, the second group and the third group are 17, 17 and 3 respectively; the first group, the second group and the parameters of the max pooling layer in the third group are 16, 16 and 2 in order.

The classification network is composed of a flattening layer, the first fully connected layer, a regularization layer, a modified linear unit layer, a second fully connected layer and a softmax function layer connected in sequence; the parameters of the first fully connected layer are 512, and the parameters of the second The parameter of the fully connected layer is the number of fault types.

The present invention will be described in detail below with the Paderborn University bearing failure case.

The dataset used in the experiments is the Paderborn dataset. The bearing code used is shown in Table 1. Sensor data for each bearing includes vibration data and current data.

This dataset contains bearings in three different states: Inner Race Failure (IR), Outer Race Failure (OR) and Health Status (H). The data sets are from bearings operating at different speeds, radial forces and load torques. The working conditions of the bearings used in the present invention are shown in Table 2.

The present invention assumes that A, B, C, and D are distributed in four clients, and uses two or three of them as source domain clients, based on the proposed federated domain adaptive fault diagnosis method based on statistical feature fusion (FDG ), to co-train the model without generating data aggregation. The trained model will be tested on the target domain client.

Table 1 Paderborn data set experimental bearing code

Bearing Code Bearing status KA15, KA16, KA30, KA04, KA22 OR KI14, KI17, KI21, KI16, KI18 IR K001, K005, K002, K004, K003 h

Table 2 Paderborn dataset under different working conditions

Data set code speed(rpm) Radial force (N) Load torque (Nm) A 1500 1000 0.7 B 1500 1000 0.1 C 1500 400 0.7 D. 900 1000 0.7

On the Paderborn data set, the results of the present invention and the comparison with other methods are shown in Table 3.

Table 3 Experimental results of Paderborn dataset

source domain dataset test data set FedAvg FDG A,B C 0.8842 0.9072 B,D C 0.9117 0.9373 A,B,D C 0.9137 0.9422 B,C,D A 0.9978 1.0 A,C,D B 0.9942 1.0

Compared with the federated average (FedAvg), the proposed method achieves better results. This shows that compared with other methods, the method proposed in the present invention has better generalization ability, which means that the model trained on the source domain client by the method proposed in the present invention can be adapted to other fields.

The beneficial effect of the present invention is that, compared with the prior art, the federal domain adaptive fault diagnosis method based on statistical feature fusion of the present invention utilizes the difference in statistical distribution of data between the target domain and the source domain, and the statistics of multi-sensor signals of source domain data The difference in distribution improves the generalization ability of the fault diagnosis model in the target domain. In the present invention, only the statistical features of the target domain data are transmitted to each source domain client, which reduces the communication burden. In addition, at the source domain client, the statistical features of multi-sensor data features will be exchanged with each other to achieve feature domain space enhancement.

The applicant of the present invention has made a detailed description and description of the implementation examples of the present invention in conjunction with the accompanying drawings, but those skilled in the art should understand that the above implementation examples are only preferred implementations of the present invention, and the detailed description is only to help readers better To understand the spirit of the present invention rather than limit the protection scope of the present invention, on the contrary, any improvement or modification made based on the spirit of the present invention shall fall within the protection scope of the present invention.

Claims (10)

1. A federal domain adaptive fault diagnosis method based on statistical feature fusion is characterized by comprising the following steps:

(1) The target domain client transmits the statistical features of the unlabeled fault data to the central server, and the central server transmits the statistical features to all source domain clients; the source domain client uses the received statistical characteristics and the statistical characteristics of the active domain data to respectively carry out data standardization on the source domain data set;

(2) The fault diagnosis model of the central server comprises a feature extraction network and a classification network, the central server sends a global model to all source domain clients, and then the source domain clients use two standardized source domain data as input and train the received global model in the source domain clients based on a correlation alignment method;

(3) After the models of all source domain clients are trained for one round, the trained models are sent to a central server, and the central server carries out global model aggregation on the received models of all source domain clients; then the training data are sent to all source domain clients to carry out the next training, and when the training round number reaches a set value, the training task is ended, so that a final fault diagnosis model is obtained;

(4) In the test stage, the central server sends the trained fault diagnosis model to the target domain client, the target domain client acquires mechanical fault data to be diagnosed, and a fault diagnosis result is obtained through the fault diagnosis model.

2. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 1, wherein in step (1), the statistical features are mean and standard deviation.

3. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 1, wherein in step (2), in the feature extraction network, the standardization of the output features of the present channel is achieved by adopting the statistical feature exchange of different feature extraction channels, and the features of the exchange standardization are adopted as the input of the classification network.

4. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 1, wherein in step (2), the global model received by the source domain client includes a feature extraction network and a classification network.

5. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 4, wherein the received global model is trained in the source domain client based on a correlation alignment method, in particular, the source domain client further comprises feature space metrics, which constrain the distance between different features of the same sensor data using the correlation alignment method.

6. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 5, wherein the input of feature space metrics is to implement standardized features using statistical feature exchange of different feature extraction channels.

7. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 1, wherein in step (3), the global model on the central server is aggregated as,

feature extractor for N global models on N source domain clients

Classifier of N global models on N source domain clients +.>

After averaging, the feature extractor of the global model on the central server is updated +.>

And classifier->

Wherein (1)>

Feature extractor for global model on ith source domain client,/i>

A classifier that is a global model on the ith source domain client.

8. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 1, wherein in step (4), the model received by the target domain client includes a feature extraction network and a classification network.

9. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 8, wherein the two channels of the feature extraction network have the same structure, and the feature extraction network is composed of three groups of convolution layers, regularization layers, linear correction units and a maximum pooling layer which are sequentially connected; the input of the feature extraction network first enters the first set of convolution layers, the first set of largest pooling layers is connected to the second set of convolution layers, and the second set of largest pooling layers is connected to the third set of convolution layers.

10. The federal domain adaptive fault diagnosis method based on statistical feature fusion according to claim 8, wherein the classification network is composed of a flattening layer, a first fully-connected layer, a regularization layer, a modified linear unit layer, a second fully-connected layer and a softmax function layer, which are sequentially connected.

Images