CN116878885A - Bearing fault diagnosis method based on self-adaptive joint domain adaptive network - Google Patents

Abstract

The invention discloses a bearing fault diagnosis method based on an adaptive joint domain adaptation network and relates to the technical field of fault diagnosis. The method includes obtaining unlabeled target domain data; inputting the target domain data into pre-trained bearing fault diagnosis training The model performs detection and diagnosis to obtain diagnostic evaluation data; the fault type is determined based on the diagnostic evaluation data; among them, the bearing fault diagnosis training model includes a feature extractor, a label classifier and a domain adaptation module. This invention can simultaneously reduce the edge distribution difference and conditional distribution difference between the source domain and the target domain, and adjust the attention of the two distribution differences in the training process in real time through adaptive weighting factors without manual experience to adjust and constrain the feature extractor. training direction, so as to better narrow the joint distribution and achieve unsupervised domain adaptation tasks.

Description

A bearing fault diagnosis method based on adaptive joint domain adaptation network

Technical Field

The present invention relates to the technical field of fault diagnosis, and in particular to a bearing fault diagnosis method based on an adaptive joint domain adaptation network.

Background Art

In the unsupervised domain adaptation problem in the field of bearing faults, bringing the marginal distribution of the source domain and the target domain closer can effectively reduce the distance between domains, and bringing the conditional distribution closer can effectively reduce the distance within the domain. However, bringing only one distribution closer can hardly achieve good results in some complex situations. By bringing the joint distribution of the two domains closer, the distance between and within the domains can be simultaneously reduced from different levels, achieving better domain adaptation effects. At present, most joint distribution adaptation works face two key problems: on the one hand, when reducing the difference in conditional distribution, the reliability of the pseudo-labels of target domain samples is not reasonably considered, and the continuous accumulation of erroneous pseudo-label samples will affect the direction of the training process; on the other hand, the importance of marginal distribution and conditional distribution in different domain adaptation tasks is not well considered, and the balance coefficient between the two distributions is set by relying on manual experience, which has great uncertainty.

Summary of the invention

The technical problem to be solved by the present invention is to provide a bearing fault diagnosis method based on an adaptive joint domain adaptation network, which does not need to rely on manual experience to adjust the balance coefficient of the two distributions, can reduce the impact of the accumulation of erroneous pseudo-labels in the target domain on the training direction, bring the joint distribution of the two domains closer, constrain the training direction of the feature extractor, reduce inter-domain differences and intra-domain differences, and thus better realize the unsupervised domain adaptation task.

In order to solve the above technical problems, the present invention provides a bearing fault diagnosis method based on an adaptive joint domain adaptation network, including: obtaining unlabeled target domain data; inputting the target domain data into a pre-trained bearing fault diagnosis training model for detection and diagnosis to obtain diagnostic evaluation data; determining the fault type according to the diagnostic evaluation data; wherein the bearing fault diagnosis training model includes a feature extractor, a label classifier and a domain adaptation module; the domain adaptation module includes: using an AJMMD-based adaptive joint domain adaptation difference to measure the marginal distribution difference and conditional distribution difference between the source domain and the target domain and adaptively adjusting the importance of the distribution difference between the two; the step of adaptively adjusting the importance of the distribution difference between the two includes: calculating an adaptive weighting factor through the marginal distribution difference and the conditional distribution difference, and using the adaptive weighting factor to adaptively adjust the importance of the distribution difference between the two.

As an improvement of the above solution, the training steps of the bearing fault diagnosis training model include:

S1. Obtain n _s labeled source domain data and n _t unlabeled target domain data; S2. Input the source domain data and the target domain data into the feature extractor for feature extraction to obtain source domain feature data and target domain feature data; S3. Input the source domain feature data and the target domain feature data into the label classifier for label classification to obtain the predicted pseudo-label of the target domain; S4. Input the source domain feature data, the target domain feature data, the predicted pseudo-label of the target domain and the true label of the source domain data into the domain adaptation module for domain adaptation loss calculation to obtain the domain adaptation loss _LAJMMD ; S5. Calculate the classification loss _LC according to the true label of the source domain data and the cross entropy loss function; S6. Back-propagate the gradient according to the classification loss _LC and the domain adaptation loss _LAJMMD to optimize and update the parameters of the feature extractor and the parameters of the label classifier; S7. When the current number of iterative calculations is greater than or equal to the preset number of iterations, output the trained bearing fault diagnosis training model, otherwise return to step S2 to continue training.

As an improvement of the above scheme, the feature extractor is a one-dimensional residual feature extractor, which includes four Block blocks. The first Block block includes a convolution layer, the convolution layer adopts a one-dimensional convolution kernel, and the remaining Block blocks include two Bottleneck blocks; the label classifier includes a fully connected output layer.

As an improvement of the above scheme, the step of back-propagating gradients according to the classification loss _LC and the domain adaptation loss _LAJMMD to optimize and update the parameters of the feature extractor and the parameters of the label classifier includes: constructing an overall loss function according to the classification loss _LC , the domain adaptation loss _LAJMMD and the overall domain adaptation factor and calculating the overall loss L; performing model parameter optimization calculation according to the gradient descent algorithm and the overall loss L to optimize and update the parameters of the feature extractor and the parameters of the label classifier.

As an improvement of the above scheme, the domain adaptation loss _LAJMMD is calculated by the AJMMD adaptive joint domain adaptation loss function in the domain adaptation module, and its function calculation formula is:

Among them, _LAJMMD represents the domain adaptation loss, _LMMD represents the marginal distribution loss, _LMMD represents the conditional distribution loss, and w represents the adaptive weighting factor of each batch data in training. , MMD (·) is the maximum mean difference loss function, LMMD (·) is the local maximum mean difference loss function, F(x _s ) represents the source domain feature data, F(x _t ) represents the target domain feature data, y _s represents the true label of the source domain data, Denoted as the predicted pseudo-label of the target domain.

As an improvement of the above scheme, the calculation formula of the cross entropy loss function is:

in, Denoted as classification loss, is represented as the prediction result of the label classifier, y ^s is represented as the true category label of the source domain data, represents the true label of the i-th sample in the source domain, M represents the total number of fault types in the source domain, I(·) is the marker function, m represents the current fault category, and a _i,m represents the The predicted probability that a sample belongs to m categories.

As an improvement of the above scheme, the calculation formula of the overall loss function is:

in, Expressed as the overall loss, Denote as the ideal parameters of the feature extractor, The ideal parameters for the label classifier, Represented as the parameter to be optimized for the feature extractor, Represented as the parameter to be optimized for the label classifier, is the overall domain adaptation factor.

As an improvement of the above solution, the one-dimensional convolution kernel is a 32x1 convolution kernel.

The present invention also provides a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, the above-mentioned bearing fault diagnosis method can be implemented.

The present invention also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the above-mentioned bearing fault diagnosis method can be implemented.

The implementation of the present invention has the following beneficial effects:

The present invention is based on the bearing fault diagnosis method, computer equipment and computer readable storage medium of the adaptive joint domain adaptation network. The marginal distribution difference and conditional distribution difference between the source domain and the target domain can be measured through the adaptive joint domain adaptation difference of AJMMD, and the importance of the distribution difference between the two can be adaptively adjusted. There is no need to rely on artificial experience to adjust the balance coefficient of the two distributions, and the influence of the accumulation of erroneous pseudo-labels in the target domain on the training trend can be reduced, the joint distribution of the two domains is brought closer, the training trend of the constraint feature extractor is reduced, and the inter-domain difference and the intra-domain difference are reduced, so as to better realize the unsupervised domain adaptation task. The model parameter optimization and update processing is carried out through the gradient descent method and the overall loss function to output the bearing fault diagnosis training model with the trained model parameters.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a bearing fault diagnosis method based on an adaptive joint domain adaptation network according to the present invention;

FIG2 is a training flow chart of the bearing fault diagnosis training model of the present invention;

FIG3 is a schematic diagram of accuracy data of convolution kernels of different widths of the present invention;

FIG4 is a schematic diagram of data of the present invention and other weighting methods;

FIG5 is a data schematic diagram of a migration task performed by the method of the present invention and other model methods based on a data set from Western Reserve University;

FIG6 is a data schematic diagram of the migration task performed by the present invention and other model methods based on the Jiangnan University dataset.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. It is hereby stated that the directional terms such as up, down, left, right, front, back, inside, outside, etc. that appear or will appear in the text of the present invention are only based on the accompanying drawings of the present invention, and are not specific limitations of the present invention.

As shown in FIG1 , a specific embodiment of the present invention provides a bearing fault diagnosis method based on an adaptive joint domain adaptation network, comprising:

S1, obtain unlabeled target domain data;

S2, inputting the target domain data into a pre-trained bearing fault diagnosis training model for detection and diagnosis to obtain diagnostic evaluation data;

S3, determining the fault type based on the diagnostic evaluation data;

It should be noted that the bearing fault diagnosis training model includes a feature extractor, a label classifier, and a domain adaptation module. Among them, the domain adaptation module includes: using the AJMMD adaptive joint domain adaptation difference to measure the marginal distribution difference and conditional distribution difference between the source domain and the target domain and adaptively adjusting the importance of the distribution difference between the two; the step of adaptively adjusting the importance of the distribution difference between the two includes: calculating the adaptive weighting factor through the marginal distribution difference and the conditional distribution difference, and using the adaptive weighting factor to adaptively adjust the importance of the distribution difference between the two. In this way, there is no need to rely on manual experience to adjust the balance coefficient of the two distributions, which can reduce the impact of the accumulation of erroneous pseudo-labels in the target domain on the training direction, close the joint distribution of the two domains, constrain the training direction of the feature extractor, reduce the difference between domains and the difference within the domain, so as to better realize the unsupervised domain adaptation task.

In the case of bearing fault diagnosis, unlabeled target domain data is obtained as test sample data, and the test sample data is input into the bearing fault diagnosis training model for detection and diagnosis, so that the corresponding bearing fault category diagnostic evaluation data can be obtained. Through the diagnostic evaluation data, the bearing fault type can be quickly and accurately determined to meet the needs of users. Among them, AJMMD is a custom word, which stands for Adaptive Joint Domain Adaptive Difference.

As shown in Figure 2, the training steps of the bearing fault diagnosis training model include:

S10, obtaining n _s labeled source domain data and n _t unlabeled target domain data;

It should be noted that, in the fault diagnosis of motor bearings, the bearing source vibration data includes unlabeled target data and labeled source domain data.

Get n _s labeled source domain data and n _t unlabeled target domain data .in, is the i-th sample data in the source domain data, is the label data corresponding to the i-th sample data in the source domain data, is the i-th sample data in the target domain data. They have the same feature space and the same category space, but their marginal distribution and conditional distribution are different.

S20, inputting the source domain data and the target domain data into a feature extractor for feature extraction to obtain source domain feature data and target domain feature data;

S30, inputting the source domain feature data and the target domain feature data into a label classifier for label classification to obtain a predicted pseudo label of the target domain;

Specifically, the feature extractor is a one-dimensional residual feature extractor, which includes four Block blocks. The first Block block includes a convolution layer, which uses a one-dimensional convolution kernel. The remaining Block blocks include two Bottleneck blocks, which are used to implement residual connections. The label classifier includes a fully connected output layer, which is used to classify the features extracted from the feature extractor and output the predicted pseudo-labels of the target domain. A global average pooling layer is also provided before the fully connected output layer. A one-dimensional residual network model (1D-ResNet) is constructed through the above structural layers. The main structural parameters of the model are shown in Table 1 below:

Table 1 Main structural parameters of the model

As shown in Table 1, the convolution layer of the first Block of this feature extractor uses a convolution kernel of size 32x1, and the remaining Blocks all use 2 Bottleneck blocks, which can effectively reduce the number of parameters and training time of the model, while extracting deep features of the vibration signal. Among them, Bottleneck1 represents a regular residual block, Bottleneck2 represents a downsampled residual block, C represents Conv1d, B represents BatchNorm1d batch normalization, A represents activation, and all use Relu activation function, P represents Maxpool1d, and S represents ShortCut residual connection part. This feature extractor uses a 32x1 convolution kernel to increase the width of the convolution kernel of the first layer, so that the receptive field is increased, more data information can be considered, the influence of noise data is reduced, and the training process is more stable.

S40, inputting the source domain feature data, the target domain feature data, the predicted pseudo-label of the target domain, and the real label of the source domain data into the domain adaptation module to calculate the domain adaptation loss, so as to obtain the domain adaptation loss _LAJMMD ;

It should be noted that the present invention embeds the AJMMD adaptive joint domain adaptation part behind the global average pooling layer, which uses the maximum mean difference MMD to calculate the marginal distribution difference, and measures the conditional distribution loss between the two domains by the local maximum mean difference LMMD loss. The importance of the marginal distribution difference and the conditional distribution difference in the training process is adjusted in real time by the adaptive weighting factor, and the conditional distribution difference and the marginal distribution difference are weightedly combined to make them approximate to the joint distribution difference.

MMD is one of the most widely used marginal distribution difference metrics in domain adaptation. It is a kernel method that maps data points from the input space to the Hilbert feature space and compares the mean difference between the source and target domains in the feature space.

LMMD (Local Maximum Mean Difference) is a conditional distribution adaptation method based on the existing CMMD. It not only uses the characteristics of the samples, but also uses the label information of the samples to weight the samples, and achieves conditional distribution alignment by obtaining pseudo labels in the target domain. One difference from CMMD is that LMMD does not take absolute weights, but uses the category probabilities of the model's output in the target domain sample data as the weighting matrix, which reduces the impact of the accumulation of incorrect pseudo labels in the target domain on the training direction.

The source domain features F(x _t ) and target domain features F(x _t ) extracted by the feature extractor are sent to the domain adaptation space of AJMMD, and the true label y _s of the source domain data and the predicted pseudo label of the target domain are converted into As an input parameter, the domain adaptation loss L _AJMMD is calculated through the AJMMD adaptive joint domain adaptation loss function in the domain adaptation module. The function calculation formula is:

Among them, _LAJMMD represents domain adaptation loss, _LMMD represents marginal distribution loss, _LMMD represents conditional distribution loss, w represents the adaptive weighting factor of each batch data in training; MMD (·) is the maximum mean difference loss function, LMMD (·) is the local maximum mean difference loss function, F( _xs ) represents the source domain feature data, F( _xt ) represents the target domain feature data, _ys represents the true label of the source domain data, Denotes the predicted pseudo-label of the target domain, H represents the Hilbert space, Φ(·) is the mapping function, _Ds is the source domain data, and _Dt is the target domain data.

Where w is the adaptive weighting factor for each batch of data in training, and its calculation formula is:

in, is an exponential function with e as base and L _MMD as exponent, is an exponential function with e as the base and _LLMMD as the exponent. When the difference in marginal distribution is greater, it means that the overall distribution difference in the domain is large. At this time, w will increase adaptively so that the marginal distribution gets relatively more attention during the optimization process, otherwise it will decrease. w allows the importance of the two distributions to be adaptively adjusted according to the actual situation during the training process.

The adaptive joint domain adaptation part of AJMMD does not need to rely on manual experience to adjust the balance coefficient of the two distributions. It can reduce the impact of the accumulation of erroneous pseudo-labels in the target domain on the training direction, bring the joint distribution of the two domains closer, constrain the training direction of the feature extractor, reduce the differences between and within domains, and thus better realize the unsupervised domain adaptation task.

Specifically, the expanded calculation formula of the maximum mean difference loss function MMD(·) is:

Among them, H represents the Hilbert space, and the features of the source domain and the target domain are mapped to the space through the mapping function Φ (·) for difference measurement. Represents the inner product of the kernel function between the source domain features and the target domain sample features.

The expanded calculation formula of the local maximum mean difference loss function LMMD (·) is

Wherein, M is the total number of categories in the category space. In fault diagnosis, M is preferably 10 fault categories, but is not limited thereto. and Respectively represent the weights of source domain and target domain samples belonging to category m, Used to calculate the weighted sum of m categories to balance the possible imbalance in each batch category in the mini-batch stochastic gradient descent method. For the sample _xi weight The calculation of is as follows:

_yim is the i-th sample label vector, and D represents the source domain or the target domain. In domain adaptation, the entire LMMD module needs to input 4 parameters: source domain feature data, target domain feature data, the true label of the source domain data, and the output probability distribution of the target domain (i.e., the predicted pseudo label of the target domain).

S50, calculating the classification loss _LC according to the real label of the source domain data and the cross entropy loss function;

It should be noted that after the label classifier outputs, the Softmax function is used to output the probability distribution of the corresponding category, and the index corresponding to the maximum probability is used as the predicted label. The cross entropy loss function is used as the classification loss function, and its calculation formula is:

in, Denoted as classification loss, is represented as the prediction result of the label classifier, y ^s is represented as the true category label of the source domain data, represents the true label of the i-th sample in the source domain, M represents the total number of fault types in the source domain, m represents the current fault category, and a _i,m represents the The predicted probability that a sample belongs to m categories, I(·) is the label function, if , then I=1, otherwise I=0.

S60, back-propagating the gradient according to the classification loss _LC and the domain adaptation loss _LAJMMD to optimize and update the parameters of the feature extractor and the parameters of the label classifier;

Specifically, the steps of back-propagating gradients according to the classification loss _LC and the domain adaptation loss _LAJMMD to optimize and update the parameters of the feature extractor and the parameters of the label classifier include:

Step 1: construct an overall loss function according to the classification loss _LC , the domain adaptation loss _LAJMMD and the overall domain adaptation factor and calculate the overall loss L; specifically, the calculation formula of the overall loss function is:

in, Expressed as the overall loss, Denote as the ideal parameters of the feature extractor, The ideal parameters for the label classifier, Represented as the parameter to be optimized for the feature extractor, Represented as the parameter to be optimized for the label classifier, is the overall domain adaptation factor.

Step 2: Perform model parameter optimization calculation based on the gradient descent algorithm and the overall loss L to optimize and update the parameters of the feature extractor and the parameters of the label classifier.

Based on the current overall loss and the parameters to be optimized of the feature extractor, the parameters of the corresponding feature extractor are updated according to the gradient descent algorithm; accordingly, the parameters of the label classifier can also be updated according to the above principle, thereby updating the parameters of the model in reverse.

S70. When the current iterative calculation number is greater than or equal to the preset iterative number, output the trained bearing fault diagnosis training model; otherwise, return to step S20 to continue training.

It should be noted that when the number of iterative calculations is greater than or equal to the preset number of iterations, the trained bearing fault diagnosis training model is output, otherwise the parameters of the currently updated feature extractor and the parameters of the label classifier are input into step S20 to continue training until the preset number of iterations is met, so as to obtain a bearing fault diagnosis training model with output trained model parameters. The preset number of iterations can be set according to actual conditions and is not specifically limited here.

In order to verify the effectiveness, feasibility and superiority of the method of the present invention, the method of the present invention is described and verified by specific examples below.

Example 1

The data set of this embodiment uses the Western Reserve University data set (CWRU) as an example. This data set uses the original drive end vibration signal of 12Khz to collect vibration signals at four different load speeds (1730 rpm, 1750rpm, 1772rpm, 1797rpm), and regards them as data sets of four different working conditions (A, B, C and D data sets). The fault type is a single-point processing operation of the electrospark machining technology. There are four states in total, which are divided into normal data (N), ball fault (BF), inner ring fault (IF), and outer ring fault (OF). Each fault type has a fault diameter of different sizes (0.007 inches, 0.014 inches, 0.021 inches), and each working condition can be divided into 10 categories in total, including 1 group of normal data and 9 groups of fault data. The acquisition window size of each fault category is 784 data points, and the sliding overlapping sampling method is used to maintain the correlation between continuous samples. The sliding step size is 80, and 150 samples are sampled for each fault category, with a total of 10*150, which is 1500 samples. The specific information of the data set is shown in Table 2 below:

Table 2 Information of four data sets under CWRU variable operating conditions

The influence of the width of the convolution kernel of the first convolution layer of 1D-ResNet on the training process is verified by using two different working condition (D→A) data sets. The optimizer uses small batch SGD, the learning rate is 0.01, and the convolution kernel size of the first convolution layer is selected from {7, 14, 32, 64, 96}. The accuracy of convolution kernels with different widths is shown in Figure 3.

As shown in Figure 3, when the convolution kernel width is set to 7, relatively few effective features are obtained, the training process is relatively unstable, and its highest accuracy converges to 97.40%; when the convolution kernel is increased to 14, due to the increase in the receptive field and model parameters, more key information can be taken into account, and the accuracy is greatly improved, reaching 100% for the first time in 42 epochs, and after a small oscillation, it basically converges to 100% in 80 epochs; when the convolution kernel is set to 32, it reaches 100% for the first time in the 36th epoch, and then it converges to this optimal state very stably; and as the convolution kernel continues to increase to 64, it is limited to a small oscillation between 98% and 99% before the 100th epoch, and it reaches 100% basically after the 120th epoch. When the convolution kernel is further increased to 96, the accuracy drops to 98.47%. Therefore, the invention adopts a 32X2 convolution kernel to make the stability, convergence speed, and accuracy of model training better than the bandwidth of other convolution kernels and achieve a good balance.

As shown in Figure 4, Figure 4 shows the coefficients of commonly used conditional distribution and marginal distribution, and its coefficient combination is selected from {[0.2, 0.2], [0.5, 0.5], [0.8, 0.8], [0.1, 0.9], [0.3, 0.7], [0.7, 0.3], [0.9, 0.1]}. The value corresponding to the first index of each element is the importance weight of LMMD, and the second index corresponds to the importance weight of MMD.

When the combination coefficient is set to [0.2, 0.2], [0.5, 0.5], [0.8, 0.8], it is equivalent to placing the conditional distribution and the marginal distribution at the same level of importance. Especially when it is set to [0.2, 0.2], the classification performance of the model does not reach 80% accuracy. The combination coefficient [0.5, 0.5] can be regarded as a special case of the marginal distribution. In the marginal distribution coefficient combination [0.1, 0.9], the classification performance of the model is also relatively poor. The method of the present invention can realize adaptive weighted allocation, which has high accuracy and stability. Therefore, setting the balance coefficient through artificial experience has very large uncertainty, and the weights of the two distributions cannot be adjusted adaptively according to the actual situation. Moreover, the decision-making ability of the model often cannot achieve the effect of the adaptive weighting method of the present invention.

Furthermore, the method of the present invention and other model methods are used to perform the same migration task to verify the effectiveness of the method of the present invention. Among them, mutual migration is performed between 4 different working conditions ABCD, with a total of 12 migration tasks. The model parameter combination is set so that the overall domain adaptation factor is fixed to 1.2, the optimizer uses small batch SGD, the learning rate is 0.01, the batch size is 128, and the total number of iterations is 150 epochs.

The accuracy information of the above migration tasks performed by the method of the present invention and other model methods is shown in Figure 5. As can be seen from Figure 5, the average accuracy of the backbone network 1D-ResNet of the working model of the present invention on 12 migration tasks reached 90.75%, and even reached 98.62% when the source domain and the target domain were not much different in the two working conditions, indicating that the use of the improved residual network can extract some key features of the same fault category. However, when the difference increases, the performance of the model without domain adaptation will drop significantly; and the average accuracy of other methods with domain adaptation on 12 migration tasks is generally higher than that of the method without domain adaptation. Among them, the CMMD method does not consider the confidence of pseudo labels, so its classification accuracy is not outstanding compared with DDC and D-CORAL of edge distribution network; DSAN has a relatively stable classification ability in each migration task, and its average accuracy reaches 98.61%, making the 1D-ResNet+LMMD combination even higher than JAN and BDA of joint distribution method, indicating that the reliability of considering pseudo labels is very critical; and the joint distribution network method JAN and BDA are generally better than the edge distribution network and conditional distribution network except DSAN; SC-1DCNN with adaptive weighting factor is higher than JAN and BDA with manually set parameters, and the overall migration effect is close to the present invention; the method of the present invention has achieved an accuracy of more than 99% in all migration tasks of the Western Reserve University data set and is very stable, and the average accuracy of 12 migration tasks has reached 99.90%, which is higher than 98.61%, 98.46%, and 99.76% of DSAN, BDA, and SC-1DCNN. Therefore, the accuracy and stability of migration tasks using the method of the present invention are the highest.

Example 2

The data set of this embodiment uses the Jiangnan University data set (JNU) as an example. Among them, the data acquisition frequency is 50KHz, the sampling time is 20s, and the data are collected at three speeds of 600rpm, 800rpm, and 1000rpm. Compared with the Western Reserve University data set, the difference in speed is greater, so it is regarded as data of three different working conditions. Each working condition contains 1 healthy state and 3 fault states. The three faults are ball fault, inner ring fault and outer ring fault. There are 500 samples in each category, a total of (4*500) 2000 samples, 2048 data points for each sample, and a sliding step of 240. By recording the working conditions at the three speeds of 600rpm, 800rpm and 1000rpm as E, F, and G respectively, the data set is specifically shown in Table 3 below:

Table 3 Information of three data sets under JNU variable operating conditions

There are 6 migration tasks in total when the three working conditions E, F and G are mutually migrated. The data results of the 6 migration tasks performed by the method of the present invention and other model methods are shown in Table 4 and FIG6 .

Table 4 Test results of Jiangnan University dataset (%)

As can be seen from Table 4 and Figure 6, the average accuracy of almost all models on the six migration tasks of Jiangnan University is lower than the test results of Western Reserve University. The average accuracy of the CMMD model is not outstanding compared with 1D-ResNet without any domain adaptation method, with a difference of only 1.34%. This is because CMMD is a conditional distribution domain adaptation method that requires label information to align subdomains. When the difference between the two domains increases, the output target domain pseudo-label is already very unreliable. Strong alignment will mistakenly bring different categories closer. The DSAN model using the LMMD method can consider the confidence of the pseudo-label by weighting, and can achieve a relatively ideal effect in the migration task with a small difference, but the accuracy in the migration task with a large difference (G→E) is only 83.91%; because even if the pseudo-label confidence is taken into account, the probability of obtaining an incorrect pseudo-label will increase when the difference is greater, which will seriously affect the method of conditional distribution alignment through label information. It can be seen from Table 3 and Figure 6 that the average accuracy of the method of the present invention in these six migration tasks is still better than that of other compared models, with an accuracy of 95.26%, which is higher than 94.98%, 93.07% and 93.25% of SC-ADCNN, DSAN and BDA.

The results of the above embodiments fully demonstrate the effectiveness, feasibility and superiority of the bearing fault diagnosis method based on the adaptive joint domain adaptation network provided by the present invention.

The present invention also provides a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, the above-mentioned bearing fault diagnosis method can be implemented.

The present invention also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the above-mentioned bearing fault diagnosis method can be implemented.

In summary, the present invention can measure the marginal distribution difference and conditional distribution difference between the source domain and the target domain through the AJMMD adaptive joint domain adaptation difference and can adaptively adjust the importance of the distribution difference between the two, without relying on manual experience to adjust the balance coefficient of the two distributions, and can reduce the impact of the accumulation of erroneous pseudo-labels in the target domain on the training trend, close the joint distribution of the two domains, constrain the training trend of the feature extractor, reduce the inter-domain difference and intra-domain difference, so as to better realize the unsupervised domain adaptation task. The model parameters are optimized and updated by the gradient descent method and the overall loss function to output the bearing fault diagnosis training model with trained model parameters.

By improving the width of the convolution kernel of the first layer in the one-dimensional residual feature extractor, the receptive field is increased, more data information can be taken into account, the impact of noise data is reduced, and the training process is made more stable.

The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims (10)

1. A bearing fault diagnosis method based on an adaptive joint domain adaptation network, characterized by comprising: Obtain unlabeled target domain data; Inputting the target domain data into a pre-trained bearing fault diagnosis training model for detection and diagnosis to obtain diagnostic evaluation data; determining a fault type based on the diagnostic evaluation data; Wherein, the bearing fault diagnosis training model includes a feature extractor, a label classifier and a domain adaptation module; The domain adaptation module includes: using the AJMMD-based adaptive joint domain adaptation difference to measure the marginal distribution difference and conditional distribution difference between the source domain and the target domain and adaptively adjusting the importance of the distribution difference between the two; The step of adaptively adjusting the importance of the difference between the two distributions includes: An adaptive weighting factor is calculated by using the marginal distribution difference and the conditional distribution difference, and the importance of the distribution difference between the two is adaptively adjusted by using the adaptive weighting factor. 2. The bearing fault diagnosis method according to claim 1, characterized in that the training step of the bearing fault diagnosis training model comprises: S1, obtain n _s labeled source domain data and n _t unlabeled target domain data; S2, inputting the source domain data and the target domain data into the feature extractor for feature extraction to obtain source domain feature data and target domain feature data; S3, inputting the source domain feature data and the target domain feature data into the label classifier for label classification to obtain a predicted pseudo label of the target domain; S4, inputting the source domain feature data, the target domain feature data, the predicted pseudo label of the target domain and the real label of the source domain data into the domain adaptation module to calculate the domain adaptation loss to obtain the domain adaptation loss _LAJMMD ; S5, calculating the classification loss _LC according to the real label of the source domain data and the cross entropy loss function; S6. Back-propagate gradients according to the classification loss _LC and the domain adaptation loss _LAJMMD to optimize and update parameters of the feature extractor and the label classifier; S7. When the current iterative calculation number is greater than or equal to the preset iterative number, output the trained bearing fault diagnosis training model, otherwise return to step S2 to continue training. 3. The bearing fault diagnosis method according to claim 2 is characterized in that the feature extractor is a one-dimensional residual feature extractor, which includes four Block blocks, the first Block block includes a convolution layer, the convolution layer uses a one-dimensional convolution kernel, and the remaining Block blocks each include two Bottleneck blocks; the label classifier includes a fully connected output layer. 4. The bearing fault diagnosis method according to claim 2, characterized in that the step of back-propagating gradients according to the classification loss _LC and the domain adaptation loss _LAJMMD to optimize and update the parameters of the feature extractor and the parameters of the label classifier comprises: Constructing an overall loss function according to the classification loss _LC , the domain adaptation loss _LAJMMD and the overall domain adaptation factor and calculating the overall loss L; The model parameter optimization calculation is performed according to the gradient descent algorithm and the overall loss L to optimize and update the parameters of the feature extractor and the parameters of the label classifier. 5. The bearing fault diagnosis method according to claim 4, characterized in that the domain adaptation loss _LAJMMD is calculated by the AJMMD adaptive joint domain adaptation loss function in the domain adaptation module, and its function calculation formula is: Among them, _LAJMMD represents the domain adaptation loss, _LMMD represents the marginal distribution loss, _LMMD represents the conditional distribution loss, and w represents the adaptive weighting factor of each batch data in training. , MMD (·) is the maximum mean difference loss function, LMMD (·) is the local maximum mean difference loss function, F(x _s ) represents the source domain feature data, F(x _t ) represents the target domain feature data, y _s represents the true label of the source domain data, Denoted as the predicted pseudo-label of the target domain. 6. The bearing fault diagnosis method according to claim 5, characterized in that the calculation formula of the cross entropy loss function is: in, Denoted as classification loss, is represented as the prediction result of the label classifier, y ^s is represented as the true category label of the source domain data, represents the true label of the i-th sample in the source domain, M represents the total number of fault types in the source domain, I(·) is the marker function, m represents the current fault category, and a _i,m represents the The predicted probability that a sample belongs to m categories. 7. The bearing fault diagnosis method according to claim 6, characterized in that the calculation formula of the overall loss function is: in, Expressed as the overall loss, Denote as the ideal parameters of the feature extractor, The ideal parameters for the label classifier, Represented as the parameter to be optimized for the feature extractor, Represented as the parameter to be optimized for the label classifier, is the overall domain adaptation factor. 8. The bearing fault diagnosis method according to claim 3, characterized in that the one-dimensional convolution kernel is a 32x1 convolution kernel. 9. A computer device, characterized in that it comprises a processor and a memory, wherein the memory is used to store a computer executable program, and the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, it can implement the bearing fault diagnosis method described in any one of claims 1 to 8. 10. A computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the bearing fault diagnosis method according to any one of claims 1 to 8 can be implemented.