Multi-source-domain ball screw fault diagnosis method and system with partial source domain data missing

Abstract

The invention belongs to the technical field related to fault signal processing, and discloses a multi-source-domain ball screw fault diagnosis method and system for partial source domain data missing. The method comprises the following steps: carrying out data complementation on vibration signals with partial fault category missing under different working conditions in the operation process of the lead screw, and obtaining complete source domains corresponding to all the working conditions; the source domains and the target domains are input into a domain adaptive network for training, the domain adaptive network comprises a prototype learning module and a classifier, and in the prototype learning module, a minimum maximum mean value difference loss function and a feature norm loss function are adopted to close the feature distribution distance between each source domain and the target domain; the classifier takes the fault type as output; and inputting a to-be-diagnosed vibration signal into the classifier of the domain adaptive network to obtain a fault type corresponding to the to-be-diagnosed vibration signal. Through the method, the problem that an existing diagnosis method cannot diagnose unknown working condition faults under the condition that partial working condition training data are missing is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to fault signal processing, and more specifically, relates to a multi-source domain ball screw fault diagnosis method and system with partial source domain data missing. Background Art

[0002] With the rapid development of the equipment manufacturing industry, ball screws, as transmission elements, play an increasingly important role in industrial development. However, because they are often in a harsh environment, their reliability is difficult to guarantee. Therefore, more and more researchers have begun to pay attention to the problem of ball screw fault diagnosis. In recent years, many scholars at home and abroad have applied deep learning algorithms represented by convolutional neural networks to the fault diagnosis of ball screws. Through a comprehensive and in-depth analysis of relevant signals such as vibration, sound, and motor current during the operation of the ball screw, different network models have been built and trained to achieve a more accurate prediction of the health status of the ball screw. However, these models assume that the training data and the test data are consistent in data distribution and require a certain amount of labeled data with the same working condition, which is difficult to obtain or expensive in actual scenarios, making it difficult to promote and apply on a large scale.

[0003] Fault diagnosis methods based on transfer learning make up for the shortcomings of the above two methods to a certain extent. The domain adaptation algorithm solves the problem caused by the difficulty in obtaining labels by migrating source domain data with easy labels to target domain data with difficult labels. However, most traditional fault diagnosis methods based on transfer learning assume that the source domain label space is consistent with the target domain label space, which is often difficult to meet in actual industrial scenarios. In real scenarios, data from different working conditions or sources are usually required to obtain different health status information to meet all health states in the target domain. Summary of the invention

[0004] In response to the above defects or improvement needs of the prior art, the present invention provides a multi-source domain ball screw fault diagnosis method and system with partial source domain data missing, which solves the problem that the existing diagnosis method cannot perform fault diagnosis under different working conditions or data missing.

[0005] To achieve the above object, according to one aspect of the present invention, a multi-source domain ball screw fault diagnosis method with partial source domain data missing is provided, the method comprising the following steps:

[0006] Offline training:

[0007] For vibration signals with some fault categories missing under different working conditions during the operation of the screw, the vibration signals under any working condition are supplemented to obtain vibration signals of various fault types, and all vibration signals and the fault labels corresponding to each vibration signal are used as the source domain under this working condition; in this way, the source domain corresponding to each working condition is obtained;

[0008] The source domain and the target domain corresponding to each working condition are simultaneously input into the domain adaptive network to train the domain adaptive network, wherein the domain adaptive network includes a prototype learning module and a classifier. During the training process, the prototype learning module adopts the method of minimizing the maximum mean difference loss function and the feature norm loss function to shorten the feature distribution distance between each source domain and the target domain. After shortening the feature distribution distance between the source domain and the target domain, the classifier classifies the data in the source domain and the target domain and outputs the fault type with the highest probability corresponding to the target domain;

[0009] Online Application:

[0010] The vibration signal to be diagnosed is input into the classifier of the domain adaptive network to obtain the fault type corresponding to the vibration signal to be diagnosed.

[0011] Further preferably, the data completion module includes a generator and a domain identifier, the generator is used to generate source domain pseudo data from data in the source domain, the domain identifier is used to identify the similarity between the source domain pseudo data and the data in the domain source, and adjust the parameters in the source domain pseudo data according to the similarity result, so that the similarity between the final source domain pseudo data domain and the source domain data meets the preset requirements.

[0012] Further preferably, the data completion module adopts a recurrent adversarial network.

[0013] Further preferably, before the source domain and the target domain corresponding to each working condition are simultaneously input into the domain adaptive network, the data in the source domain and the target domain are preprocessed, and the preprocessing includes time domain feature extraction and frequency domain feature extraction of the data, and then the features extracted by the time domain feature extraction and the frequency domain feature extraction are simultaneously input into the domain adaptive network.

[0014] Further preferably, after preprocessing the data in the source domain and the target domain, a feature extraction network is used to extract features from the data in the source domain and the target domain.

[0015] Further preferably, the feature extraction adopts a deep residual network.

[0016] Further preferably, the loss function of the prototype learning module is:

[0017]

[0018] Among them, L Y is the loss function of the prototype learning module, is the function that maximizes the mean difference loss, i is the number of the source domain, n is the total number of source domains, is the feature norm loss function, C is the total number of fault categories involved in the source domain or target domain, c is the number of the fault type, and D s is the source domain, D t is the target domain, is the kernel function, x t is the target domain sample, is the data sample of the i-th source domain, They are source domain samples and target domain samples The corresponding sample weight of the cth fault type.

[0019] Further preferably, the feature norm loss function The formula is as follows:

[0020]

[0021] Among them, n s , n t are the total number of data in the source domain and the target domain, i is the number of the source domain, x i is a data sample of source domain i, that is, the loss function calculates the sum of the distances between each sample in the source domain and the target domain, D s is the source domain, D t is the target domain, L d is the distance of the L2 regularized norm, h is a function in the function class H, H is all possible function combinations of the L2 norm operator and the deep feature representation module, θ y are the model parameters updated in the current iteration of the fault classifier, and θ0 represents the model parameters of the last iteration.

[0022] Further preferably, the loss function of the classifier is:

[0023]

[0024]

[0025] Among them, L F is the loss function of the classifier, i is the source domain number, n1 is the total number of source domain data, represents the loss function of each source domain data, m, k represent different source domain health status categories, n2 is the number of health status categories in the source domain, is the label of the source domain sample, I{y=m} is the selection function, θ kare the basic parameters of the classifier; is the fully connected layer feature of the i-th source domain.

[0026] According to another aspect of the present invention, a system for diagnosing a multi-source domain ball screw fault according to the above-mentioned method for diagnosing a multi-source domain ball screw fault with missing partial source domain data is provided, the system comprising a data completion module and a domain adaptation module, wherein:

[0027] The data completion module is used to complete the data of the vibration signal;

[0028] The domain adaptation module includes a feature extraction module and a domain adaptation module, wherein the feature extraction module is used to extract depth-sensitive features of input data;

[0029] The domain adaptation module is used to perform training using data in the source domain and the target domain, and then to train a classifier using the trained data, and finally to classify the vibration signal to be diagnosed using the trained classifier to obtain the corresponding fault type.

[0030] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:

[0031] 1. The present invention provides a screw fault diagnosis method. Compared with the traditional fault diagnosis method based on transfer learning, the source domain data in the traditional method is generally complete. When a single source domain does not have all types of fault data, all source domain data are generally directly fused to obtain all fault categories. The method provided by the present invention uses a data completion module to complete the data type of the fault type under each working condition, so that all fault types are included under each working condition. At the same time, multiple source domains are constructed using different working conditions. Compared with the existing method that only uses data from one source domain for training, the domain adaptive network obtained by training has higher diagnostic accuracy and more accurate results.

[0032] 2. The loss function of the prototype learning module in the domain adaptive network of the present invention includes the maximum mean difference loss function (LMMD) and the feature norm loss function. By introducing the local maximum mean difference loss function, the distribution difference of similar fault features under different working conditions can be reduced. By introducing the L2 regularization loss function, the features of the target domain can be transferred to the area with a larger feature norm through mapping and keep consistent with the feature norm of the source domain, which not only accelerates the convergence speed of the network but also enhances the domain adaptive effect.

[0033] 3. The present invention introduces deep residual network feature extraction before domain adaptive network training to reduce the loss of effective features, thereby greatly improving its generalization ability and prediction performance, and the extracted deep features are more representative. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of a multi-source domain ball screw fault diagnosis method with partial source domain data missing constructed according to a preferred embodiment of the present invention;

[0035] Figure 2 It is a flow chart of a data completion model constructed according to a preferred embodiment of the present invention;

[0036] Figure 3 is a schematic diagram of a fault diagnosis method model constructed according to a preferred embodiment of the present invention;

[0037] Figure 4 is a test instance confusion matrix constructed according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0039] like Figure 1 As shown, a multi-source domain ball screw fault diagnosis method with partial source domain data missing includes the following steps:

[0040] (1) Collect the vibration signal during the operation of the screw and preprocess the signal; input the preprocessed training data into the data completion module to train the model, and use the module to complete the missing health status data in the source domain;

[0041] The main analysis object is the vibration signals of various healthy states when the motor is running. Due to different working conditions, it is very likely that the length of each sample data is not fixed, and the sample needs to be truncated to meet the input samples of the same length in the network.

[0042] Due to the limited health status categories of the public data sets and the lack of certain health status categories during the signal acquisition process, it is impossible to obtain data on all health status categories; however, simply using a simple union between different data sets will reduce the training effect or make it impossible to train the network due to the large differences in the working conditions of the data sets. Therefore, the data completion module is used to complete data in different source domains.

[0043] The data completion module uses a cyclic adversarial network (CycleGAN), which effectively fills in missing samples to meet the needs of subsequent domain adaptation. The structure is shown in Figure 2 As shown in Figure 2, take two source domains as an example, assuming that X A With XB Representing the public health status category data of different source domains, the data completion module consists of two generators (G AB , G BA ), two domain discriminators (D A , D B ). The training process is mainly as follows: A As the generator G AB Input and generate B-source domain pseudo data X B ′, and then use it as the generator G BA Input regeneration A-source domain pseudo data X A ′, through the domain discriminator D B Identify B-source domain pseudo data X B ′ and source domain data X B Similarity: If the two are not similar, the pseudo data will be given a low score (the lowest score is 0); if the two are similar, the pseudo data will be given a high score (the highest score is 1); B-source domain data X B The operation method is similar. This module improves the generator's generation ability through cross-training confrontation between the generator and the domain discriminator, making the generated pseudo data more reliable.

[0044] The module network parameters are optimized using the source domain public health status category data. The loss function mainly includes the following three parts:

[0045] Since the data completion module improves the generator’s generation ability through the confrontation between the domain discriminator and the generator, the confrontation loss function between the domain discriminator and the generator is defined as:

[0046]

[0047] in, Denotes the generator G AB With the domain discriminator D B The adversarial loss function between Denotes the generator and its distinguisher D B The adversarial loss function between They are The i-th data, for The number of data.

[0048] In the process of cross-domain data generation, the generated data should be stored in the same features as the original data, and the original features of the data should not be lost in the case of reverse generation. Constraints are imposed by adding a reconstruction consistency loss function between the data, which is defined as follows:

[0049]

[0050] In the data, the same fault category should maintain identity consistency between different source domains, and the identity unification loss is defined as follows:

[0051]

[0052] The total loss function is expressed as follows:

[0053]

[0054] After the module training is completed, the data of the unique categories in a source domain (i.e., the missing categories in other source domains) are input into the network to fill the missing data in the source domain, such as Figure 3 Shown on the left.

[0055] (2) The completed multi-source domain training data is preprocessed and input into the multi-source domain adaptation module to train the model; the optimized model is tested using the test data to obtain the diagnosis results.

[0056] In one embodiment of the present invention, since the original data length is too large, the number of network parameters will increase, and the data is first preprocessed. Data pre-feature processing is to extract time-frequency domain features from the data, wherein the time domain features mainly include variance, skewness, kurtosis, waveform factor, peak factor, and pulse factor, and the frequency domain features mainly include center frequency, root mean square frequency, frequency standard deviation, frequency variance, and spectral kurtosis. Time-frequency analysis mainly includes empirical mode decomposition (EMD) and wavelet transform, and the features extracted by preprocessing are used as module input for verification.

[0057] The multi-source domain adaptation module is divided into feature extraction network and domain adaptation network, such as Figure 3 As shown on the right, the former extracts data features adaptively, and the latter predicts the health status of the ball screw based on the extracted features.

[0058] First, the feature extraction network uses a deep residual network, which is mainly used to learn depth-sensitive features as much as possible from the input sample data. Different from time-frequency domain features, depth-sensitive features have stronger abstraction capabilities and are suitable for processing complex data, which can achieve better results.

[0059] Secondly, the domain adaptation network adopts an integrated thinking, using the characteristics of each source domain and the target domain for domain adaptation operations and then classification, and uses the voting method to predict the final health status of the target domain data.

[0060] The preprocessed multi-source domain data and target domain data are simultaneously input into the feature extraction network to obtain deep sensitive features; each source domain sensitive feature and target domain sensitive feature are further extracted, and the maximum mean difference loss function (LMMD) and feature norm loss function are used in this process to make the distributions between the same subclasses in the two domains as similar as possible; finally, a classifier is used to classify the health status of the features; since there are multiple classifiers, the principle of raising hands and voting is adopted, and the category with the most predicted health status in the classifier is the health status finally predicted by the network.

[0061] This paper uses multiple loss functions to optimize the module network parameters, which mainly includes the following three parts:

[0062] First, the network module should have a good classification effect on the source domain data. Here we define the classifier loss function L f :

[0063]

[0064] In the formula, i is the source domain number, n1 is the total number of source domain data, m and k represent different source domain health status categories, n2 is the number of health status types in the source domain, is the label of the source domain sample, I{y=m} represents the selection function, that is, if y=m, the value is 1, otherwise it is 0; θ k Represents the basic parameters of the classifier; is the fully connected layer feature of the i-th source domain.

[0065] Secondly, by minimizing the maximum mean difference loss, the distribution of the same subclasses between the source domain and the target domain is made as similar as possible to achieve domain adaptation. The formula is as follows:

[0066]

[0067] In the formula, i is the number of the source domain, C represents the total number of categories involved in the domain, c is the number of the fault type, and D s represents the source domain, D t represents the target domain, Represents the kernel function, you can choose linear kernel function, Gaussian kernel function, etc. is the data sample of the i-th source domain, x t is the target domain sample, They are source domain samples and the target domain x t The weight of a certain fault type state, and for each sample, the sum of the weights of all fault types is 1, and the calculation formula is as follows:

[0068]

[0069] In the formula, y pc is the label vector Y p In the source domain, the real label vector is used; in the target domain, since there is no real label, the pseudo label vector generated by the network is used for calculation.

[0070] The hypothesis of too small feature norm indicates that the effect of domain transfer basically depends on the features in the target task that have too little information and too small feature norm. Assuming that in the case of loose alignment, adjusting the target features away from the area with smaller feature norm can still complete the safe transfer. Based on this hypothesis, we hope that the features of the target domain can be transferred to the area with larger feature norm through mapping and keep consistent with the feature norm of the source domain. Its loss function can be expressed as follows:

[0071]

[0072] Where i is the number of the source domain, n s , n t Indicates the number of data in the source domain and the target domain, D s, D t represents the source domain and target domain datasets, L d represents the distance of the L2 regularized norm, h is a function in the function class H, H is all possible function combinations of the L2 norm operator and the deep feature representation module, θ y represents the model parameters updated in the current iteration of the fault classifier; θ0 represents the model parameters of the last iteration.

[0073] The total loss function is expressed as follows:

[0074]

[0075] In the formula, i is the number of the source domain, and n represents the number of source domains.

[0076] (3) Input the test data into the network to predict the health status, obtain the diagnosis results, and compare and analyze them with their true labels.

[0077] The present invention will be further described below with reference to specific embodiments.

[0078] This embodiment provides a multi-source domain ball screw fault diagnosis method with partial source domain data missing, and the specific implementation steps are as follows:

[0079] S1 data completion

[0080] The motor vibration signals during the test bench processing are collected and marked with health status categories (normal, slight wear of the screw, misalignment between the screw and the guide rail, local pitting of the screw), and then the data is preprocessed.

[0081] In order to obtain experimental data, a test bench was built independently, and a series of simulation experiments of different health states of the lead screw were carried out. The lead screw uses the THK precision linear motor series BNK1520-3G0-521LC5Y, and the x-axis servo motor uses the Panasonic MHMF042L1U2M. In order to simulate different working conditions, the motor speed is set to 120rmp and 160rmp. The load is changed by changing the load of the moving platform on the lead screw. There are two types of 0HP and 1HP. The test is repeated 4 times, and each experiment requires 10 reciprocating motions. The acceleration sensor is Dytran 3263A2, the acquisition card is NI 9234, the sampling frequency is 25600Hz, and the z-direction vibration signal of the motor end is saved for each reciprocating motion. Due to the different motor speeds in different working conditions, there is a large difference in the time used for a reciprocating stroke, resulting in different vibration signal lengths. The present invention truncates the data to ensure that the input data length is consistent. In order to reduce the loss of useful information, this paper intercepts the middle part of each data segment and extracts 16,000 points of data from each sample as a new sample for subsequent algorithm verification.

[0082] In order to meet the needs of the scenario of the present invention, a part of the working conditions randomly lacking a certain type of health status data is used as the source domain (the field where data and labels are easily obtained), and the remaining working conditions are used as the target domain (the field where data labels are difficult to obtain).

[0083] Table 1 Example data set

[0084]

[0085]

[0086] The verification of the present invention is carried out, and the specific settings are shown in Table 1:

[0087] The preprocessed training data is input into the data completion module to train the module.

[0088] Considering the lack of some healthy state samples in the source domain data, the present invention effectively fills in the missing samples through the data filling module (CycleGAN) to meet the subsequent domain adaptation. Its structure is as follows: Figure 3 As shown on the left. Assume that X A With X B Represents the public health status category data of source domains A and B, that is, normal or slight wear (taking normal as an example, the optimization module should refer to the total loss of all public health status category data). A As the generator G AB Input and generate B-source domain normal pseudo data X B ′, and then use it as the generator G BA Input regenerate A-source domain normal pseudo data X A′. Through the domain identifier D B Determine B- normal pseudo data X in the source domain B ′ and B-source domain normal data X B The similarity between the generator and the domain discriminator is improved by the confrontation between the generator and the domain discriminator; B- The operation method of normal data in the source domain is similar. The specific parameters of the generator and the domain discriminator in the data completion module are shown in Table 2:

[0089] Table 2 Data completion module model structure parameters

[0090]

[0091]

[0092] The present invention uses multiple loss functions to optimize the network parameters of the data completion module, and optimizes the network parameters of the data completion module by minimizing the sum of the three losses. AB The pseudo screw and guide misalignment data of the B-source domain are generated to complete the source domain B. Similarly, the screw local pitting data using the B-source domain data is generated using the generator G BA The pseudo screw local pitting data of the A-source domain is generated to complete the source domain A.

[0093] S2 preprocesses the padded training data and inputs it into the multi-source domain adaptation module to train the model.

[0094] Since the input data length is too long, the number of network parameters will increase, which increases the difficulty of network training. The present invention first pre-processes the data. Data pre-feature processing is to extract time-frequency domain features of the data. The original signal is transformed into 5 layers of wavelet packets and 3 layers of empirical mode decomposition (EMD) to extract the time domain and frequency domain features of each layer of data as the input of the network; the time domain features mainly include variance, skewness, kurtosis, waveform factor, peak factor, and pulse factor, and the frequency domain features mainly include center frequency, root mean square frequency, frequency standard deviation, frequency variance, and spectral kurtosis. A total of 120 features are extracted here.

[0095] The multi-source domain adaptation module is divided into a feature extraction network and a domain adaptation network. Its structure is shown in the figure below. Figure 3 As shown on the right. The former extracts data features adaptively, and the latter predicts the health status of the ball screw based on the extracted features. The feature extraction network uses a deep residual network (one-dimensional Resnet18), which is mainly used to learn deeper sensitive features from the input sample features as much as possible. The domain adaptation network adopts an integrated thinking, and predicts the sensitive features of each source domain and target domain after domain adaptation, and uses the voting method to predict the final health status of the target domain data.

[0096] The present invention uses the above three loss functions to optimize the network parameters of the multi-source domain adaptation module.

[0097] S3 uses the test data to test the optimized model and obtain the diagnosis results.

[0098] The optimized model is tested using the target domain dataset C, and its confusion matrix is ​​as follows: Figure 4 As shown in the figure, we can see that the accuracy of each health status is over 90%, and the average accuracy is 94.36%, achieving the expected effect.

[0099] In order to verify the superiority of this method, this section compares this method with the more advanced transfer learning methods. Due to the partial loss of domain data, different source domain data are generally fused as a single source domain for domain adaptation. Specifically, the methods compared in this paper include domain adversarial neural network (DANN), multi-adversarial domain adaptation network (MADA), and deep sub-domain adaptation network (DSAN). The main structural settings of these methods are consistent, and their accuracy is shown in Table 3.

[0100] Table 3 Diagnostic accuracy of different methods

[0101]

[0102] The results show that this method has a good detection effect when the source domain data is partially missing. Since this method is applicable to more scenarios, it has a broader application prospect.

[0103] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-source domain ball screw fault diagnosis method with partial source domain data missing, characterized in that: The method comprises the following steps: Offline training: For vibration signals with some fault categories missing under different working conditions during the operation of the screw, the vibration signals under any working condition are supplemented to obtain vibration signals of various fault types, and all vibration signals of the working condition and the fault label corresponding to each vibration signal are used as the source domain under the working condition; in this way, the source domain corresponding to each working condition is obtained; The source domain and the target domain corresponding to each working condition are simultaneously input into the domain adaptive network to train the domain adaptive network, wherein the domain adaptive network includes a prototype learning module and a classifier. During the training process, the prototype learning module adopts the method of minimizing the maximum mean difference loss function and the feature norm loss function to shorten the feature distribution distance between each source domain and the target domain. After shortening the feature distribution distance between the source domain and the target domain, the classifier classifies the data in the source domain and the target domain and outputs the fault type with the highest probability corresponding to the target domain; Online Application: The vibration signal to be diagnosed is input into the classifier of the domain adaptive network to obtain the fault type corresponding to the vibration signal to be diagnosed.

2. A multi-source domain ball screw fault diagnosis method with partial source domain data missing as claimed in claim 1, characterized in that: The data completion module includes a generator and a domain identifier. The generator is used to generate source domain pseudo data from data in the source domain. The domain identifier is used to identify the similarity between the source domain pseudo data and the data in the domain source, and adjust the parameters in the network for generating source domain pseudo data according to the similarity result, so that the similarity between the final source domain pseudo data domain and the source domain data meets the preset requirements.

3. A multi-source domain ball screw fault diagnosis method with partial source domain data missing as claimed in claim 1 or 2, characterized in that: The data completion module adopts a cyclic adversarial network.

4. The multi-source domain ball screw fault diagnosis method with partial source domain data missing according to claim 1, characterized in that: Before the source domain and the target domain corresponding to each working condition are simultaneously input into the domain adaptive network, the data in the source domain and the target domain are preprocessed, and the preprocessing includes time domain feature extraction and frequency domain feature extraction of the data, and then the features extracted by the time domain feature extraction and the frequency domain feature extraction are simultaneously input into the domain adaptive network.

5. A multi-source domain ball screw fault diagnosis method with partial source domain data missing as claimed in claim 4, characterized in that: After preprocessing the data in the source domain and the target domain, a feature extraction network is used to extract features from the data in the source domain and the target domain.

6. A multi-source domain ball screw fault diagnosis method with partial source domain data missing as claimed in claim 1, characterized in that: The feature extraction adopts a deep residual network.

7. The multi-source domain ball screw fault diagnosis method with partial source domain data missing according to claim 1, characterized in that: The loss function of the prototype learning module is: Among them, L Y is the loss function of the prototype learning module, is the function that maximizes the mean difference loss, i is the number of the source domain, n is the total number of source domains, is the feature norm loss function, C is the total number of fault categories involved in the source domain or target domain, c is the number of the fault type, and D s is the source domain, D t is the target domain, is the kernel function, x t is the target domain sample, is the data sample of the i-th source domain, They are source domain samples and target domain sample x t The corresponding sample weight of the cth fault type.

8. A multi-source domain ball screw fault diagnosis method with partial source domain data missing as claimed in claim 7, characterized in that: The feature norm loss function The formula is as follows: Where i is the number of the source domain, n s , n t are the total number of data in the source domain and the target domain, respectively, i is a data sample, that is, the loss function calculates the sum of the distances between each sample in the source domain and the target domain, D s is the source domain, D t is the target domain, L d is the distance of the L2 regularized norm, h is a function in the function class H, H is all possible function combinations of the L2 norm operator and the deep feature representation module, θ y are the model parameters updated in the current iteration of the fault classifier, and θ0 is the model parameter of the last iteration.

9. A multi-source domain ball screw fault diagnosis method with partial source domain data missing as claimed in claim 7 or 8, characterized in that: The loss function of the classifier is: Among them, L F is the loss function of the classifier, i is the source domain number, n1 is the total number of source domain data, represents the loss function of each source domain data, m, k represent different source domain health status categories, n2 is the number of health status categories in the source domain, is the label of the source domain sample, I{y=m} is the selection function, θ k are the basic parameters of the classifier; is the fully connected layer feature of the i-th source domain data.

10. A system for diagnosing a multi-source domain ball screw fault according to the method for diagnosing a fault with partial source domain data missing according to any one of claims 1 to 9, characterized in that: The system includes a data completion module and a domain adaptation module, wherein: The data completion module is used to complete the data of the vibration signal; The domain adaptation module includes a feature extraction module and a domain adaptation module, wherein the feature extraction module is used to extract depth-sensitive features of input data; The domain adaptation module is used to perform training using data in the source domain and the target domain, and then to train a classifier using the trained data, and finally to classify the vibration signal to be diagnosed using the trained classifier to obtain the corresponding fault type.

Images