Partial domain fault diagnosis method of mixed subdomain distribution difference based on dynamic weight

By introducing dynamic weighting and mixed subdomain distribution difference modules into deep neural networks, the problem of negative domain migration in rotary mechanical fault diagnosis is solved, the fault recognition accuracy is improved, and efficient fault diagnosis is achieved under different working conditions.

CN118916766APending Publication Date: 2024-11-08CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411160241.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing partial domain adaptation method has negative domain migration problems when dealing with rotating machinery fault diagnosis, especially because the subdomain distribution differences and subdomain number differences are not effectively considered, resulting in low recognition accuracy and in the case of inconsistent data distribution. It is difficult to effectively diagnose faults.

Method used

Using a deep neural network model based on dynamic weights of mixed subdomain distribution differences, a feature extractor, a dynamic adversarial weighting module and a mixed subdomain distribution difference module are constructed by sliding window slice data, and the weighted cross entropy loss and conditional entropy loss are calculated. Deep neural networks can reduce the distribution differences between the source domain and the target domain and improve the accuracy of fault identification.

Benefits of technology

It effectively improves the accuracy of fault recognition, can perform mechanical equipment fault diagnosis more accurately under different working conditions, reduces the noise of source domain data weight, and enhances the robustness and recognition capabilities of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916766A_ABST
    Figure CN118916766A_ABST
Patent Text Reader

Abstract

The invention provides a partial domain fault diagnosis method of mixed subdomain distribution difference based on dynamic weight, and relates to the field of mechanical intelligent fault diagnosis and computer artificial intelligence, bearing fault data is collected, and a source domain and a target domain are formed; slicing the data set and carrying out normalization processing on each sample; inputting a source domain sample and a target domain sample into a feature extractor and a dynamic adversarial weighting module respectively to obtain corresponding feature vectors, and solving a source domain data weight; inputting the weighted source domain data features and target data domain features into a label classifier and a mixed subdomain distribution difference module, calculating weighted cross entropy loss and conditional entropy loss, and calculating mixed subdomain distribution differences; and training the established deep neural network model, and obtaining a final deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution difference after a preset number of iterations is reached for fault diagnosis. The method has the beneficial effect that the fault identification precision is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of mechanical intelligent fault diagnosis and computer artificial intelligence, and in particular to a partial domain fault diagnosis method based on hybrid subdomain distribution differences of dynamic weights. Background Art

[0002] Modern industry is developing rapidly, and industrial system integration and information technology are also developing continuously. Highly reliable components and systems are the guarantee for the safe operation of aerospace. Undetected aviation faults may lead to catastrophic events in aviation machinery. As the core component of rotating machinery such as aircraft engines and space shuttles, mechanical transmission systems are prone to various faults under long-term high-speed, heavy-load and harsh operating conditions, which directly affect the safe operation of mechanical systems. Rotating machinery represented by aviation equipment has attracted the attention of researchers. Early fault detection and diagnosis technology can predict the development trend of faults and play a key role in preventing faults in mechanical engineering transmission systems. Therefore, in order to avoid the occurrence of subsequent major accidents caused by minor faults, various industries attach great importance to intelligent fault diagnosis systems on rotating machinery.

[0003] Deep learning is a type of machine learning algorithm based on artificial neural networks. In the years of research and development, a large number of algorithm models have emerged, gradually becoming the main representative algorithm of machine learning and achieving success in many fields. Due to its special advantages in massive data processing, feature learning and pattern recognition, deep learning technology has been proven to be a promising tool in many applications in the manufacturing industry. Intelligent models are built by mapping the relationship between the health status of industrial equipment and industrial data from end to end. Intelligent fault diagnosis using deep learning methods can process massive monitoring data and determine the health status of the machine, which can improve the reliability and safety of industrial production. Compared with the traditional machine learning methods commonly used in the past, these deep learning models do not require expert experience, train the entire model parameters in an adaptive manner, automatically learn key features and predict results. However, deep learning technology has its limitations, which inhibits its further progress, advancement and application in complex real-world scenarios. For example, the success of most intelligent fault diagnosis methods based on deep learning depends on the following two conditions: first, a large amount of labeled fault data is required for model training; second, the training and test data should satisfy the same probability distribution. However, in the case of rotating machinery data sampling, the equipment usually runs smoothly for a long time, so it is difficult to obtain a sufficient amount of sample data with fault labels. Furthermore, machine operating conditions, signal sampling frequency, and workload may result in significant differences in the distribution of acquired training and test data.

[0004] Transfer learning technology has been proven to be effective in solving the problems of sparse labeling and significant distribution differences of sample data in the field of intelligent diagnosis. This technology extracts feature information from target samples by learning feature knowledge from a source domain with sufficient labeled data and then transferring this knowledge to a target domain with limited labeled data. Among the existing methods, domain adaptation (DA) is a subset of transfer learning methods and has achieved remarkable results in cross-domain transfer tasks. However, in practical situations, the target domain space is usually much smaller than the source domain space. This is particularly evident in a specific DA scenario, namely partial domain adaptation (PDA).

[0005] Most of the current PDA methods are based on classifiers or discriminators to learn the source data weights. Then, they use the weighted distribution alignment loss defined on the target data and the weighted source data to train the feature extractor, which results in the weighted feature distribution alignment method being not robust to the "noise" of the source data weights, i.e., some source-only data are incorrectly assigned non-zero weight alignment losses. Secondly, existing PDA methods usually notice the global distribution differences between the two domains. After global domain adaptation, the distribution differences between the domains are greatly reduced, but because the differences in subdomain distribution or the number of subdomains are not considered, the subdomains may overlap, making it difficult to effectively improve the recognition accuracy. Summary of the invention

[0006] In order to solve the problem of domain negative transfer in the field of partial domain adaptation in transfer learning, the present invention provides a partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights, which can be used for intelligent fault diagnosis of partial domain mechanical equipment under cross-working conditions. The method mainly includes:

[0007] S1: Collect bearing failure data from different working conditions of mechanical equipment to form source domain and target domain data sets;

[0008] S2: Slice the data set through a sliding window to generate samples, and normalize each sample;

[0009] S3: Construct a deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution difference, which includes a feature extractor, a dynamic adversarial weighting module, a label classifier and a mixed subdomain distribution difference module;

[0010] Input the source domain and target domain samples into the feature extractor respectively to obtain the corresponding feature vectors;

[0011] Input the source domain and target domain features into the dynamic adversarial weighting module to find the source domain data weight;

[0012] Input the weighted source domain data features and target data domain features into the label classifier and calculate the weighted cross entropy loss and conditional entropy loss;

[0013] Input the weighted source domain data and target domain data into the mixed subdomain distribution difference module to calculate the mixed subdomain distribution difference;

[0014] S4: Train the established deep neural network model to determine whether the preset number of iterations of the deep neural network model is reached. If so, end the iteration; if not, return to S3 until the preset number of iterations is reached to obtain the final deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution differences for fault diagnosis.

[0015] A storage device stores instructions and data for implementing a partial domain fault diagnosis method based on hybrid subdomain distribution differences of dynamic weights.

[0016] A partial domain fault diagnosis device based on hybrid subdomain distribution difference of dynamic weights, comprising: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights.

[0017] The technical solution provided by the present invention has the following beneficial effects: the present invention collects bearing fault data to form a source domain and a target domain; slices the data set and normalizes each sample; inputs the source domain and target domain samples into the feature extractor and the dynamic adversarial weighting module respectively to obtain the corresponding feature vectors and calculate the source domain data weight; inputs the weighted source domain data features and target data domain features into the label classifier and the mixed subdomain distribution difference module to calculate the weighted cross entropy loss and conditional entropy loss and the mixed subdomain distribution difference; trains the established deep neural network model, and obtains the final deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution difference after reaching the preset number of iterations for fault diagnosis. The present invention takes into account the difference in subdomain distribution or the difference in the number of subdomains, so that each subdomain may overlap and overlap, effectively improving the accuracy of fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0019] Figure 1 It is a flow chart of a partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights in an embodiment of the present invention.

[0020] Figure 2 It is a model structure framework diagram in an embodiment of the present invention.

[0021] Figure 3 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0023] Please refer to Figure 1 , Figure 1 : is a flow chart of a partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights in an embodiment of the present invention, the method comprising:

[0024] S1: Collect bearing failure data from different working conditions of mechanical equipment to form source domain and target domain data sets;

[0025] S2: Slice the data set through a sliding window to generate samples, and normalize each sample;

[0026] S3: Construct a deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution difference, which includes a feature extractor, a dynamic adversarial weighting module, a label classifier and a mixed subdomain distribution difference module;

[0027] Input the source domain and target domain samples into the feature extractor respectively to obtain the corresponding feature vectors;

[0028] Input the source domain and target domain features into the dynamic adversarial weighting module to find the source domain data weight;

[0029] Input the weighted source domain data features and target data domain features into the label classifier and calculate the weighted cross entropy loss and conditional entropy loss;

[0030] Input the weighted source domain data and target domain data into the mixed subdomain distribution difference module to calculate the mixed subdomain distribution difference;

[0031] S4: Train the established deep neural network model to determine whether the preset number of iterations of the deep neural network model is reached. If so, end the iteration; if not, return to S3 until the preset number of iterations is reached to obtain the final deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution differences for fault diagnosis.

[0032] The present invention is described below from three parts: bearing fault data collection under variable working conditions, construction of a deep neural network model based on hybrid subdomain distribution differences with dynamic weights, and model training strategy.

[0033] (1) Bearing failure data collection under variable operating conditions

[0034] In bearing fault diagnosis, different domains refer to different data distributions caused by changes in factors such as speed, load, or operating conditions. The source domain refers to the data used when the model is trained under specific conditions, and the target domain refers to the data faced by the model during testing or actual application. The purpose of domain adaptation is to make the data in the source domain and the target domain more similar in the feature space through techniques such as feature alignment and adversarial training, thereby improving the performance and reliability of the model in the target domain.

[0035] In actual working conditions, an accelerometer is used to collect bearing vibration signals with known fault information under one working condition as source domain data, and then bearing vibration signals with unknown fault information under other working conditions are collected as target domain data. For the collected bearing vibration signals, they are segmented and generated into samples through a sliding window. The size of the sliding window is generally selected to be close to a multiple of 2 of the sampling points of two cycles of bearing rotation. In this embodiment, 4096 sampling points are selected as the time window. Normalization is performed on each collected sample. There are multiple fault types under each working condition, including: normal state, outer ring fault, inner ring fault and rolling element fault, among which the fault types may have different damage sizes and different composite situations. The application scenario of the present invention is that the label space of the target domain data is a subset of the label space of the source domain data, and the distribution of the two domains is different. The purpose of the present invention is to find a suitable mapping to reduce the distribution difference between the source domain and the target domain, and to make the distribution of the source domain and the target domain as consistent as possible by training a deep neural network model.

[0036] In the partial domain adaptation (PDA) problem, the labeled source domain dataset is defined as The unlabeled target domain dataset is defined as in and represents the i-th source domain data and the i-th target domain data, and and represents the label corresponding to the i-th source domain data, and and denote the source domain and target domain label space respectively, N S and N T Represent the number of samples in the source domain and the target domain respectively. For a task It consists of a label space and a mapping function f(·). The mapping function f(·) represents the relationship between the samples of the dataset X and the prediction results. Let P(X,Y) and Q(X,Y) represent the probability distribution of the data in the source domain and the target domain respectively. In partial domain adaptation, P(X,Y)≠Q(X,Y) is further obtained. The goal of the present invention is to develop a method to improve the Performance of the above model. PDA methods usually use a feature extractor F(·) to extract features, a label classifier C(·) to predict labels, and a discriminator D(·) to distinguish domains.

[0037] (2) Constructing a deep neural network model based on dynamic adversarial weighting and hybrid subdomain distribution differences

[0038] The deep neural network consists of four modules, namely feature extractor, dynamic adversarial weighted module, label classifier and mixed subdomain distribution difference module.

[0039] ①Build a feature extractor

[0040] The original one-dimensional vibration data of the source domain and the target domain are divided into samples and Then it is processed by CNN. CNN is a neural network specially used for processing data, with a known grid-like topology. Sparse interactions, parameter sharing and equivalent representations are achieved through convolution and pooling operations on CNN. The present invention uses CNN as a feature extractor F(·) to reduce the dimension of the original data and extract effective features for subsequent use. CNN mainly includes three layers, including convolutional layers, Maxpooling layers and classifier layers. Convolutional layers and Maxpooling layers are used for feature learning, and the classifier layer divides the learned features into different categories. The present invention uses two convolutional layers with large-size kernels and one convolutional layer with a small-size kernel, and each convolutional layer is followed by a maximum pooling layer. Batch normalization and LeakyRELU functions are used after each convolutional layer, and dropout functions are used after some fully connected layers.

[0041] The feature extractor F(·) transforms x S and x T By f s =F(x s ) and f t =F(x t ) is mapped to a common feature space. S and x T Respectively represent the source domain and target domain sample sets that are divided, Represents the F-dimensional feature vector of the source domain and the target domain.

[0042] ②Construct dynamic adversarial weighted module

[0043] To address the negative domain shift problem in PDA, we construct a dynamic weighting module in which we design a dynamic adversarial weighting method that learns to reweight the source domain data in an adversarial manner to align the distributions of the source and target domains.

[0044] Assume shared class The source domain data belongs to the pure source class The source domain data is closer to the target domain data. In order to reduce the importance of source class data alone and reduce domain shift, the weight of the source domain data is adversarially learned by minimizing the Wasserstein distance between the reweighted source domain distribution and the target domain distribution. The dual form of the Wasserstein distance is used for learning. The weight learning process is formulated as a dynamic adversarial weighted model.

[0045] The Wasserstein distance is a measure of the difference between two distributions. The Wasserstein distance between distributions μ and v is defined as where Π is the coupled set of distributions μ and v, i.e., Π = {π|∫π(x,y)dy = μ(x), ∫π(x,y)dx = ν(y)}, x represents a random variable, y represents the label corresponding to the random variable x, and ||·|| is the l2 norm. Using the Kantorovich-Rubinstein duality, the Wasserstein distance has the dual form in, It means taking the expected value of the mapping function f(x) according to the distribution μ, It means taking the expected value of the mapping function f(x') according to the distribution v. f(x) represents the value acting on the random variable x, and f(x') is the value acting on the random variable x'.

[0046] In order to calculate the Wasserstein distance, the present invention uses the parameter θ D The random variable is parameterized by the discriminator D(·), which uses a neural network. Then, the Wasserstein distance becomes:

[0047]

[0048] The maximization in formula (1) is for all 1-Lipschitz functions where θ = {θ D :||D(·;θ D )|| L ≤1},||D(·;θ D )|| L is D(·;θ D) norm in the L-norm space. Formula (1) allows deep neural network models to use gradient-based optimization algorithms to calculate Wasserstein distance on large-scale data sets, and uses gradient penalty technology (see the paper "Imoroved training of wasserstein definition" for details) to constrain formula (1). Compared with other popular statistical distances (such as JS divergence), the Wasserstein distance calculated by the present invention has better continuity for learning distribution.

[0049] The dynamic adversarial weighted method is defined in the feature space, and the extracted features are represented as source domain data features and target domain data characteristics and represents the i-th source domain data and the j-th target domain data, θ F represents the parameters of the feature extractor F(·). The distribution of the weighted target domain data is recorded as where δ(·) is the Dirac function, n t represents the number of samples in the target domain that are divided. The weighted source domain distribution is recorded as n s Represents the number of source domain samples after segmentation. Based on the previous assumption, the source domain data of the non-shared class (the class that does not exist in the target domain) is farther away from the target domain data, and the source domain data of the shared class is closer to the target domain data, that is, only the source domain class data is farther away from the target domain data than the source domain data of the shared class.

[0050] The Wasserstein distance between the weighted source and target domain distributions is minimized, and the weights are learned as follows:

[0051]

[0052] Where W() represents the Wasserstein distance, w is the source domain data weight, is the solution space, And w i ≥0, i=1,2,3,…,n s , n s Indicates the number of source domain samples that are split.

[0053] By approximating the dual form in equation (1), equation (2) is transformed into the following adversarial weight model:

[0054]

[0055] where n s 、n tRespectively represent the number of samples in the source domain and the target domain that are divided, θ D represents the parameters of the discriminator, D(·) represents the discriminator, Θ={θ D :||D(·;θ D )|| L ≤1}, Represents the source domain data characteristics, represents the target domain data features, ||D(·;θ D )|| L Denotes D(·;θ D ) in the L-norm space, where L represents the norm space.

[0056] In formula (3), the discriminator is trained to maximize (or minimize) the average value of its output on the source domain (or target domain) to distinguish the source domain from the target domain. Conversely, the source domain data weights are learned to minimize the weighted average of the discriminator output on the source domain. Therefore, source data with smaller discriminator outputs (closer to the target domain) will be assigned larger weights.

[0057] ③Build a label classifier

[0058] The label classifier is a classification model used to map input features to specific class labels. It can be a simple linear classifier or a complex neural network. In the present invention, the label classifier consists of a fully connected layer, the number of input dimensions is the number of dimensions after flattening, and the output dimension is the number of bearing fault categories.

[0059] Send the feature vector to the label classifier for prediction, and get z s =C(f s ),z t =C(f t ).in, is the score vector, C(·) represents the label classifier, and f s 、f t Represents the D-dimensional feature vector of the source domain and the target domain. To ensure that the model has good classification ability for source domain samples, z s And the real data labels The combination of the weighted cross entropy loss on the source domain and the conditional entropy loss on the target domain is defined as:

[0060]

[0061] in yes The cross entropy loss is defined as follows: for a distribution p = (p1, p2, ..., p K ) T , H(·) is defined as H(p) = -∑ k pk log p k The conditional entropy of is the weight that specifies the importance of the source domain data, so that θ F and θ C are the parameters of F(·) and C(·), respectively, and the superscript T indicates the transpose. Minimizing the weighted cross entropy loss can force the recognition network to predict the label of the input data. Minimizing the conditional entropy loss can promote low-density separation between classes in the target domain.

[0062] ④ Hybrid subdomain distribution difference module

[0063] The goal of previous domain adaptation methods is to align the global distribution of different data, and they mainly study the problem of global domain shift. After global domain adaptation, the overall distribution of the two domains is roughly the same. However, this method does not consider the data classification problem of the two domains. After global domain matching, it is easy to cause confusion between different types of data in the two domains, destroy the discriminant structure of the data, and reduce the recognition ability of the model. To address these problems, a mixed distribution adaptation method is used to approximate the difference in the distribution of mixed subdomains between the two domains using the marginal probability and conditional probability of the source domain and the target domain, so as to obtain more comprehensive and accurate difference information between different domains.

[0064] Therefore, the present invention designs a measurement method of hybrid subdomain distribution differences based on dynamic weights (HSDW). HSDW first filters the source domain data according to the source domain data weights calculated by the dynamic adversarial weighted model, and then divides the deep features extracted by the model into different subdomains according to the true labels of the source domain and the predicted labels of the target domain, and measures the difference in the mixed distribution of the same subdomains in the two domains. Among them, the mixed distribution of the source domain is represented by the mixed distribution between the true labels and features of the source domain samples, and the mixed distribution of the target domain is approximately represented by the mixed distribution between the predicted labels and features of the source domain samples. By refining the subdomain division and comprehensive mixed distribution difference measurement, HSDW can more comprehensively and effectively evaluate the difference information between different domains, thereby achieving more accurate domain matching. HSDW divides the feature space extracted by the deep neural network model into subdomains according to the categories of the source domain, and performs refined subdomain difference measurement. The mixed distribution of the original data is approximately represented by the mixed distribution of the feature space of the target domain and the prediction space. HSDW can effectively measure the difference in mixed subdomain distribution between the target domain and the source domain.

[0065] In order to calculate HSDW, we first need to obtain the empirical mixed distribution of the dataset between the two domains. Since the true distribution of the dataset is unknown, it cannot be directly obtained. Studies have shown that the kernel embedding of the mixed distribution of the dataset can be regarded as a covariance operator, where a finite number of samples can be used to estimate the embedding. Given a sample The number of samples is given as N S According to P(X,Y), the covariance of the sample is estimated by the following formula:

[0066]

[0067] in represents the inner product operator, φ(·) represents the feature map of the original data into the Hilbert space, x i represents the i-th sample feature, y i Indicates the label corresponding to the i-th sample, N S Indicates the sample size.

[0068] In traditional shallow methods, it is difficult to achieve mixed distribution adaptation. Considering the outstanding performance of Maximum Mean Discrepancy (MMD) in transfer learning in matching domain edge probability distribution, the present invention embeds the mixed distribution into the Hilbert space through HSDW and calculates the mixed subdomain distribution difference of two variables.

[0069] Will and Divided into M subdomains and Where m∈{1,2,…,M} represents the subdomain label, P M (X,Y) and Q M (X,Y) represent subdomains respectively and The distribution of deep neural network models at each layer The activations in are and P M (X,Y) and Q M The mixed distribution of (X,Y) can be obtained by and HSDW embeds the mixed distribution of the source domain and the target domain into RKHS and calculates the difference in the distribution of mixed subdomains between different domains. and The difference in the mixed subdomain distribution can be expressed as:

[0070]

[0071] in, is the reproducing kernel Hilbert space (RKHS) with characteristic kernel l. and Indicates subdomain and A mixed distribution of represents the inner product operator, represents the number of layers of the deep neural network model, M represents the number of subdomains, represent the eigenvectors that reproduce the mixture distribution of the source domain and the target domain in the kernel Hilbert space, respectively.

[0072] Empirical estimate of Calculated as the squared distance between embeddings of the empirical kernel mean, HSDW can be calculated by the following formula:

[0073]

[0074] in, and express and The weight of category m in , and the weight is processed by the weight obtained by adversarial dynamic weighting, and represents the activation value of the i-th source domain sample in the l-th layer network; Represents the activation value of the jth source domain sample in the lth layer network. represents the activation value of the i-th target domain sample in the l-th layer network; represents the activation value of the jth target domain sample in the lth layer network. The kernel function in Non-uniform weights are used to better reflect other layers. The influence of other variables in the kernel k represents k(X s ,X t )=<φ(X s ),φ(X t )>, <·,·> represents the inner product of vectors.

[0075] (3) Model training strategy

[0076] The model framework of the method of the present invention is as follows Figure 2As shown in the figure, the backbone of the model is a classification model for predicting the target domain label, which consists of a one-dimensional convolutional neural network (1D-CNN) as a feature extractor and a fully connected layer as a label classifier. The feature extractor is used to map the source domain data and the target domain data into the feature space. The dynamic adversarial weighting module automatically calculates the importance of the source domain data to match the distribution of the target domain in the feature space to reduce the importance of the source domain class data alone. And a weighted cross entropy loss is defined on the weighted source domain data distribution, and a conditional entropy loss is defined on the target domain data to learn a transferable recognition network for the target domain. The source domain data weights calculated by this module are then passed to the hybrid subdomain distribution difference module, which helps to better achieve the subdomain adaptation task.

[0077] In order to train the recognition network to minimize the loss of the model, the present invention adopts an alternating training method to train the model. The present invention mainly optimizes the network parameters θ F and θ C , and learn the weight w by fixing other known parameters. For all i, by w i = 1 to initialize the weight w. Then, the alternating training mode is adopted when training the network:

[0078] ① Update θ with a fixed weight w F and θ C . Fix weight w and update θ F and θ C To minimize the loss in formula (4).

[0079] ②Use a fixed θ F and θ C Update weight w. Fix θ F and θ C , extract the features of all training data in the source domain and the target domain, and learn w in formula (3). Since formula (3) is a minimum-maximum optimization problem, the weights w and parameters θ of the domain discriminator can be optimized alternately by fixing the other known D Therefore, first fix w for all i i =1Optimize θ D To maximize the objective function in (4). Formula (3) uses the gradient penalty technique, then, the domain discriminator is fixed and w is optimized as follows. and The optimization problem of weight w becomes:

[0080]

[0081] Formula (8) is a second-order cone procedure. The present invention uses the CVXPY package to solve formula (8). Wherein ρ is a parameter that is automatically adjusted during the training process and is used to control the parameters of the constraints and loss function in the optimization problem.

[0082] The appropriate size of ρ is important for the method of the present invention. If the ratio of the label space size of the target domain to the source domain is If ρ is large, then ρ needs to be small so that more source data contributes to the weighted cross entropy loss, and vice versa. Since the target label space is unknown, ρ is automatically adjusted to be within the preset interval [A low ,A up ] The Wasserstein distance (the loss value in formula (3) is expressed as A ρ ). To this end, the present invention selects the initial value ρ of ρ 0 and a constant c>1, and adjust ρ as follows:

[0083] If A ρ >A up (A ρ low ), then adjust ρ by ρ = cρ (ρ = ρ / c) and solve equation (3). Adjust until A low ≤A ρ ≤A up In this embodiment, A up =5.0,A low = -5.0, c = 1.2 and ρ 0 =5.0.

[0084] In summary, the overall optimization objective loss function of the deep neural network model can be expressed as:

[0085]

[0086] in, represents the total loss, represents the classification loss, θ F and θ C represents the optimized network parameters, w represents the source domain data weight, is the HSDW loss, and λ is the balance factor between classification loss and HSDW.

[0087] (4) Simulation experiment

[0088] ① Experimental setup

[0089] In order to verify the feasibility of the present invention, a simulation experiment is conducted according to the content of the present invention. The data set selected by the present invention is the Xi'an Jiaotong University data set (XJTU).

[0090] ​The XJTU-SY bearing dataset collects full life cycle bearing information with different degradation modes (such as internal faults, external faults, and combined faults). The experimental platform of the XJTU-SY bearing experimental dataset is mainly composed of an AC motor, a hydraulic loading system, a motor speed regulator, a rotating shaft, a support bearing, and a test bearing. The bearing is an LDK UER204 rolling bearing. The experimental sampling frequency is set to include three working conditions, namely, the radial force applied at a rated speed of 2100r / min is 12kN, the radial force applied at a rated speed of 2250r / min in working condition two is 11kN, and the radial force is 10kN at a speed of 2400r / min in working condition three. The sampling frequency is 25.6kHz, the sampling interval is 1min, and the duration of each sampling is 1.28s. A total of five types of fault locations are included in each type of operating condition.

[0091] In the industrial big data environment, using historical labeled data collected from related scenarios to transfer from a large-scale but redundant source domain to an unknown small-scale target domain is a promising solution. In actual engineering applications, it is more common that the label space between the two domains is different, that is, the categories in the source domain contain the categories in the target domain. The details of XJTU-SY are shown in Table 1, where OR represents the outer ring, IR represents the inner ring, C represents the cage, N represents the normal state, and M represents the mixed state of the first three positions and the rolling element position.

[0092] Table 1 XJTU-SY dataset details

[0093]

[0094]

[0095] This experiment uses the fault part of the XJTU-SY dataset for subsequent research. According to the different categories of the source domain and the target domain, the partial domain migration task is designed, as shown in Table 2. The source domain of the experiment contains four categories. For example, the source domains of T1-T4 are internal faults, normal conditions, cage faults extracted from bearings, and external faults. The target domain can have two or three categories.

[0096] Table 2 XJTU-SY dataset migration task list

[0097]

[0098] For a fair comparison, we choose 1D-CNN as the base network, and all domain adaptation methods are adjusted based on this network. We use the stochastic gradient descent (SGD) optimizer to train the parameters in the model, with momentum set to 0.9 and weight decay set to 5×10 -4 The learning rate is given by the formula η θ=η0 / (1+αθ) β Dynamic adjustment, where θ is the training progress that changes linearly from 0 to 1, η0=0.01, α=10, β=0.75. The batch size is set to 32 and the epochs are set to 500. In order to avoid excessive impact on the labeled classifier in the early stage of training, the balance factor λ is gradually adjusted by 2 / (1+exp(-γθ))-1, γ=3. The experiments were run in a Linux server environment, using an Intel(R) Xeon(R) Gold5117CPU and an NVIDIA GeForce RTX 3090 graphics card. The PyTorch deep learning framework was used to build the model, and GPU was used to accelerate the calculation.

[0099] ②Fault diagnosis results

[0100] In order to verify the effectiveness of the model of the present invention, the present invention designed a comparative experiment. 1D-CNN is the most basic model and does not set any modules that can optimize domain migration. The DANN model adds a domain discriminator to the basic model to try to distinguish whether the feature is from the source domain or the target domain. The goal of the feature extractor is to learn features that are useful for both domains, while the domain classifier strives to distinguish the sources of these features. This setting creates an adversarial relationship, prompting the feature extractor to generate features that are increasingly difficult to distinguish between domains. DAW is a weighted model that is also used to calculate the weight of source domain data, thereby reducing domain negative migration.

[0101] Table 3 Experimental results of XJTU-SY dataset

[0102]

[0103] The experimental results of the XJTU-SY dataset are shown in Table 3. This dataset collects information on bearings with different degradation modes such as internal faults, external faults and combined faults throughout their life cycle. Compared with the diagnosis task of artificially introduced faults, the diagnosis task of bearing faults throughout their life cycle is usually more difficult. In the process of domain adaptation, the common features of different domains are not obvious, which increases the difficulty of domain adaptation. All models achieved high accuracy in tasks T2 and T8, while the accuracy of other tasks was low. For all tasks, the method proposed in the present invention has improved accuracy over 1D-CNN, DANN and DAW. Specifically, for task T2, the other three models have been able to achieve relatively high accuracy, but the model proposed in the present invention has increased the accuracy to 100%. For task T8, the other three models have also been able to achieve relatively high accuracy, but the model proposed in the present invention makes the results more stable (standard deviation becomes smaller) while achieving the same accuracy. For tasks T1, T3, T5 and T7, the accuracy of the model proposed in the present invention is higher than that of the other three models. For task T4, although the accuracy of the model proposed by the present invention is higher than that of 1D-CNN and DAW, it is 3.9% lower than that of DANN. For task T6, although the accuracy of the model proposed by the present invention is slightly lower than that of DAW, the difference in results is very small, and the accuracy is more stable (the standard deviation is smaller than that of DAW). In general, the average accuracy of the model proposed by the present invention is the highest, reaching 87.8±11.4, which proves its good domain adaptability.

[0104] See also Figure 3 , Figure 3 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a partial domain fault diagnosis device 401 based on hybrid subdomain distribution difference of dynamic weight, a processor 402 and a storage device 403.

[0105] A partial domain fault diagnosis device 401 based on hybrid subdomain distribution difference with dynamic weights: The partial domain fault diagnosis device 401 based on hybrid subdomain distribution difference with dynamic weights implements the partial domain fault diagnosis method based on hybrid subdomain distribution difference with dynamic weights.

[0106] Processor 402: The processor 402 loads and executes instructions and data in the storage device 403 to implement the partial domain fault diagnosis method based on hybrid sub-domain distribution difference with dynamic weight.

[0107] Storage device 403: the storage device 403 stores instructions and data; the storage device 403 is used to implement the partial domain fault diagnosis method based on hybrid sub-domain distribution difference of dynamic weights.

[0108] 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, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A partial domain fault diagnosis method based on hybrid subdomain distribution differences with dynamic weights, characterized by: include: S1: Collect bearing failure data from different working conditions of mechanical equipment to form source domain and target domain data sets; S2: Slice the data set through a sliding window to generate samples, and normalize each sample; S3: Construct a deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution difference, which includes a feature extractor, a dynamic adversarial weighting module, a label classifier and a mixed subdomain distribution difference module; Input the source domain and target domain samples into the feature extractor respectively to obtain the corresponding feature vectors; Input the source domain and target domain features into the dynamic adversarial weighting module to find the source domain data weight; Input the weighted source domain data features and target data domain features into the label classifier and calculate the weighted cross entropy loss and conditional entropy loss; Input the weighted source domain data and target domain data into the mixed subdomain distribution difference module to calculate the mixed subdomain distribution difference; S4: Train the established deep neural network model to determine whether the preset number of iterations of the deep neural network model is reached. If so, end the iteration; if not, return to S3 until the preset number of iterations is reached to obtain the final deep neural network model based on dynamic adversarial weighting and mixed subdomain distribution differences for fault diagnosis.

2. A partial domain fault diagnosis method based on hybrid subdomain distribution differences of dynamic weights as claimed in claim 1, characterized in that: In S1, an accelerometer is used to collect a bearing vibration signal with known fault information as source domain data, and then bearing vibration signals with unknown fault information under other working conditions are collected as target domain data; fault types include: normal state, outer ring fault, inner ring fault and rolling element fault.

3. The partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights according to claim 1, characterized in that: In S3, source domain data features Target domain data characteristics θ F represents the parameters of the feature extractor F(·), and Represents the i-th source domain data and the j-th target domain data.

4. The partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights according to claim 3, characterized in that: The dynamic adversarial weighting module is used to minimize the Wasserstein distance between the reweighted source and target domain distributions. The specific formula is: Among them, n s 、n t Respectively represent the number of samples in the source domain and the target domain that are divided, θ D represents the parameters of the discriminator, D(·) represents the discriminator, Θ = θ D :||D(·;θ D )|| L ≤1, is the solution space, And w i ≥0, i=1,2,3...,n s , w is the source domain data weight, Represents the source domain data characteristics, represents the target domain data features, ||D(·;θ D )|| L Denotes D(·;θ D ) in the L-norm space, where L represents the norm type.

5. The partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights according to claim 1, characterized in that: In S3, the feature vector is sent to the label classifier for prediction, and z is obtained. s =C(f s ),z t =C(f t ), where z s 、z t represents the score vector, C(·) represents the label classifier, and f s 、f t Represents the D-dimensional feature vector of the source domain and the target domain.

6. The partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights according to claim 1, characterized in that: The weighted cross entropy loss on the source domain is: The conditional entropy loss on the target domain is: Among them, F(·) and C(·) represent the feature extractor and label classifier respectively, θ F and θ C are the parameters of F(·) and C(·), and represents the i-th source domain data and the j-th target domain data, Represents the label corresponding to the i-th source domain data.

7. The partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights according to claim 1, characterized in that: The difference in the hybrid subdomain distribution is: in, With characteristic core The reproducing kernel Hilbert space of represents the difference in the distribution of mixed subdomains, P M , Q M Respectively represent subdomains and The distribution of represents the inner product operator, represents the number of layers of the deep neural network model, M represents the number of subdomain labels, They represent the eigenvectors of the mixed distribution of the source domain and the target domain in the reproducing kernel Hilbert space, P represents the source domain, and Q represents the target domain.

8. The partial domain fault diagnosis method based on hybrid subdomain distribution difference of dynamic weights according to claim 1, characterized in that: In S4, the optimization objective loss function of the deep neural network model is: in, represents the total loss, represents the classification loss, θ F and θ C represents the parameters of the feature extractor F(·) and the discriminator C(·), w represents the source domain data weight, is the HSDW loss, and λ is the balance factor between classification loss and HSDW.

9. A storage device, characterized in that: The storage device stores instructions and data for implementing the partial domain fault diagnosis method based on dynamic weighted hybrid subdomain distribution differences as described in any one of claims 1 to 8.

10. A partial domain fault diagnosis device based on hybrid subdomain distribution difference of dynamic weight, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the partial domain fault diagnosis method based on dynamic weighted hybrid subdomain distribution differences as described in any one of claims 1 to 8.

Citation Information

Patent Citations

  • Partial migration fault diagnosis method based on multi-scale weight selection adversarial network

    CN113837071A

  • Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework

    CN114048568A

  • Rolling bearing fault diagnosis method and system under variable working condition based on improved depth subdomain adaptation network

    CN114429150A

  • Bearing transfer learning fault diagnosis method based on domain dynamic adversarial self-adaption

    CN115600134A

  • Gear case fault diagnosis method based on deep transfer learning

    CN116894187A