Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data

Abstract

The application provides a mechanical equipment unsupervised transfer intelligent fault diagnosis method for releasing source data, which comprises the following steps: S1, collecting equipment vibration monitoring data under different working conditions; S2, dividing into a source field training sample set and a target field test sample set; S3, performing data standardization; S4, establishing an unsupervised transfer fault diagnosis model for releasing source data; S5, collecting the vibration monitoring data of the mechanical equipment, inputting into the unsupervised transfer fault diagnosis model, and realizing mechanical health state discrimination. The kernel norm maximization is introduced to further encourage the distinguishability and diversity of the prediction output, and the self-training mechanism is interacted to improve the transfer performance of the source model to the target domain, so that the health state recognition of the unlabeled target sample under the condition of releasing the source data is realized. The application considers the cost of expensive data storage and transmission and privacy protection in the actual industrial scene, so that the unsupervised transfer has the cross-domain diagnosis capability, and the demand for source field data in the model transfer process is released.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis, in particular to a mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data. BACKGROUND

[0002] Rotating machinery, as one of the most important pillars in industrial manufacturing, its safety and reliability are attracting more and more attention. However, due to long-term operation, the key components will inevitably degrade and damage. Therefore, fault diagnosis technology has been widely studied in academia, and has gradually become an indispensable technology in modern industry.

[0003] With the vigorous development of the Internet, the Internet of Things and industrial big data, deep learning models have triggered a wave of intelligent diagnosis methods and achieved remarkable achievements. This technology breaks the conventional way of model-based or signal analysis-based diagnosis methods, and regards the diagnosis problem as an end-to-end health pattern recognition problem. Although the performance of these models is very attractive, the prerequisite is to meet the independent and identically distributed assumption of data. Due to the changes in working conditions and the uncertainty of monitoring environment, the data feature distribution of the training domain and the test domain always inevitably has differences. Therefore, due to the failure to meet the above prerequisites, a serious performance decline will be observed. In other words, it is difficult to directly apply the trained model to a new task with domain distribution differences.

[0004] In view of the above problems, in recent years, diagnosis technology based on unsupervised migration has appeared. This method aims to reduce the domain bias and align the labeled source data and unlabeled target data in a common representation space, so that the model trained on the source domain can be generalized to the target domain. A common strategy is to use different metrics to measure and reduce the differences in feature distribution, and another popular pattern is to use the concept of generative adversarial network to learn the transferable knowledge between domains.

[0005] However, although the above methods have achieved certain success, they must always work under the strict condition that the source data is available during the training process. The following two conditions limit the unsupervised migration diagnosis to be unsuitable in many real scenarios. First, a key reason is that the source data is not always accessible due to privacy protection or intellectual property issues. Another key reason is that long-term monitoring, especially high-frequency sampling, means huge storage and loading costs. Therefore, it is necessary to study the transfer model without source data, which can complete the target domain diagnosis task without accessing the source data. SUMMARY

[0006] To overcome the shortcomings of the existing technologies mentioned above, the purpose of this invention is to propose a Source-Free Transfer Diagnosis (SFTD) network for cross-domain fault diagnosis of mechanical equipment. Combined with the transfer diagnosis network, it can complete fault diagnosis without relying on source data. SFTD consists of two stages: a source model generation stage and a source model transfer stage. In the source model generation stage, labeled source domain data is used as input, and a labeled smooth cross-entropy is designed as the loss function to fully mine source domain diagnostic information while suppressing the overfitting and weak fit problems of traditional cross-entropy. In the source model transfer stage, source model transfer is achieved using only unlabeled target domain data. This invention first proposes a novel self-training mechanism to fully explore the application of unlabeled data, while considering the expensive data storage and transmission costs and privacy protection costs in actual industrial scenarios. This allows unsupervised transfer to achieve cross-domain diagnostic capabilities while reducing the need for source domain data during model transfer.

[0007] Specifically, to achieve the above objectives, the solution adopted by the present invention is as follows:

[0008] A method for unsupervised migration intelligent fault diagnosis of mechanical equipment releasing source data, comprising the following steps:

[0009] S1. For rotating machinery with multiple health states, collect equipment vibration monitoring data under different working conditions, and construct a sample data space and its corresponding labels based on the monitoring data;

[0010] S2. Divide the equipment vibration monitoring data under different working conditions in step S1 into a source domain training sample set and a target domain test sample set.

[0011] S3. Standardize the sample data of the source domain training sample set and the target domain test sample set obtained in step S2 to obtain the standardized source domain training sample set and target domain test sample set.

[0012] S4. Establish an unsupervised migration fault diagnosis model for releasing source data, which includes the following sub-steps:

[0013] S41. Construct an unsupervised transfer diagnostic model, wherein the unsupervised transfer diagnostic model f includes a feature encoding module. and classification module Where d is the dimension of the encoded features, and K is the total number of categories; the feature encoding module includes a multi-layer convolutional module and a fully connected layer, and the classification module includes a fully connected layer and a non-linear activation layer;

[0014] S42. Generate source model: Utilize a supervised learning paradigm to generate a source model, and introduce label smoothing cross-entropy L. sls As the loss function for training the source model;

[0015] S43. Generating pseudo-tags using class prototypes includes the following sub-steps:

[0016] S431. Calculate the class prototype of the target domain using the following formula:

[0017]

[0018] In the formula, σ k (·) represents the k-th element in the softmax output, X t Represents the target domain space; This represents the target model that has been learned in the previous iteration, which is obtained by initializing the source model.

[0019] S432. Obtain the initial pseudo-label through distance measurement:

[0020]

[0021] In the formula, These are the initial pseudo-tags;

[0022] S433. Update the class prototype using the initial pseudo-tag, optimize the initial pseudo-tag, and obtain the optimized pseudo-tag. The specific process is as follows:

[0023]

[0024]

[0025] When the parameter is true, the indicator function ζ(·) equals 1; For initial pseudo-labels; x t represents the target domain sample; k represents the predicted category of the target domain sample; For the classifier in the target model;

[0026] S434. Introduce normalized symmetric cross-entropy to reduce interference from noisy pseudo-labels:

[0027] The formula for normalized symmetric cross-entropy is:

[0028]

[0029] The right side represents the normalized cross-entropy and its inverse representation. The normalized cross-entropy is specifically:

[0030]

[0031] In the formula, p is the prediction probability, i.e., p = σ(f(x)); Indicates a pseudo-tag;

[0032] S435. Based on the normalized symmetric cross-entropy from step S434, the following self-training objective loss function is formed:

[0033]

[0034] S44. Utilize the F-norm maximization of the output matrix to provide reliable target output prediction and discrimination capabilities, and add a kernel norm maximization loss function;

[0035] S45. Using the standardized target domain training sample set obtained in step S3, perform model transfer training on the unsupervised transfer fault diagnosis model. Fix the network parameters of the classification module, combine the self-training target loss function in step S43 and the nuclear norm maximization loss function in step S44, use stochastic gradient descent to optimize the model feature extraction module, iterate until the model converges, and complete the source model transfer.

[0036] S5. Collect vibration monitoring data of mechanical equipment and input it into the unsupervised migration fault diagnosis model to identify the health status of the machinery.

[0037] Preferably, the loss function in step S42 is:

[0038]

[0039] q' k =(1-α)q k +α(1 / K)

[0040] In the formula, x s For the source domain sample, y s σ represents the corresponding label of the source neighborhood sample; σ(·) denotes the softmax function. This represents the k-th element in the softmax output of the K-dimensional nonnormalized probability vector; if and only if y s When q = k, k If q is 1, then q is 1; otherwise, q is 1. k =0; α is the smoothing parameter, with a default value of 0.1; f s :X s →Y s Represents the source model, f t :X t →Y t For the learned objective function, inference

[0041] Preferably, the expression for maximizing the norm of the output matrix F in step S44 is:

[0042]

[0043] In the formula, P = [p ij ] B×K This represents the softmax output matrix of a batch of size B.

[0044] Preferably, the loss function for maximizing the nuclear norm in step S44 is:

[0045] L nnm =-||P|| * / B

[0046] In the formula, the convex hull of the rank of matrix P is its nuclear norm ||P||. * .

[0047] Preferably, the feature encoding module in step S41 includes four convolutional layers, a batch-normalization layer, a non-linear activation function, a pooling layer, a fully connected layer, and a Dropout layer; the classification module includes a fully connected layer and a non-linear activation layer.

[0048] Preferably, in step S2, it is assumed that the source region D s Contains n s A labeled sample in Target Domain D t Contains n t Unlabeled samples ,in

[0049] Preferably, the formula for data standardization in step S3 is as follows:

[0050]

[0051] In the formula: x i For the i-th data sample; μ i For x i The average value; σ i For x i Standard deviation; x i (j) is x i The j-th element.

[0052] Preferably, the mechanical equipment in step S1 includes a multi-stage acceleration transmission device bearing failure simulation test bench.

[0053] Preferably, the sampling frequency of the vibration data of the multi-stage acceleration transmission device is 10kHz.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] 1) This invention uses the original time-domain vibration signal as data input and adaptively extracts relevant information on the health status of key mechanical components through a diagnostic network model, thereby establishing an end-to-end mechanical fault diagnosis model, reducing the need for human resources and expert experience in the diagnostic decision-making process, and ensuring the efficiency of automated diagnosis and the accuracy of diagnostic results.

[0056] 2) This invention enables the identification of complex mechanical health status when the target domain samples are unlabeled and the data distribution of different domains differs. It completes the migration of labeled source domain data to unlabeled target domain data, makes up for the shortcomings of conventional deep learning diagnostic models, and improves the accuracy of fault identification.

[0057] 3) This invention takes into account that most deep migration fault diagnosis methods rely on the availability of source data. In view of the technical defects of traditional methods, which are always impossible to implement in many practical industrial scenarios due to the high cost of data storage, transmission and privacy protection, this invention establishes a migration method that does not require source data, and realizes deep migration fault diagnosis without the need for source data.

[0058] 4) This invention utilizes a pseudo-label learning strategy oriented towards class prototypes and a standardized symmetric cross-entropy concatenation deployment to promote robust training; at the same time, it introduces the maximization of the nuclear norm to further improve the discriminability and diversity of the predicted output, which interacts with the self-training mechanism to improve the transfer performance of the source model to the target domain, and realizes the identification of the health status of unlabeled target samples without the need for source data. Attached Figure Description

[0059] Figure 1 This is a flowchart of an intelligent fault diagnosis method for mechanical equipment based on SFTD, according to an embodiment of the present invention.

[0060] Figure 2 This is a flowchart of the training steps for the unsupervised migration fault diagnosis model for releasing source data in an embodiment of the present invention.

[0061] Figure 3 This describes the network structure and parameter settings of the feature encoding module of SFTD in this embodiment of the invention.

[0062] Figure 4 This describes the network structure and parameter settings of the classification module of SFTD in this embodiment of the invention;

[0063] Figure 5 This is a schematic diagram of the multi-stage acceleration transmission experimental device in this embodiment of the invention;

[0064] Some of the attached labels in the figure are as follows:

[0065] 1-Drive motor; 2-Coupling; 3-Planetary gearbox; 4-Fixed shaft gearbox; 5-Tachometer; 6-Bearing housing; 7-Vibration sensor; 8-Magnetic powder brake. Detailed Implementation

[0066] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0067] This invention provides an unsupervised intelligent fault diagnosis method for mechanical equipment that releases source data, such as... Figures 1-4 As shown, the specific steps include:

[0068] S1. For rotating machinery with multiple health states, collect equipment vibration monitoring data under different operating conditions, so that the collected data contains different health state information. Then, construct a sample data space and its corresponding labels based on the collected monitoring data.

[0069] S2. Divide the equipment monitoring data under different operating conditions in step S1 into two groups: the source domain training sample set and the target domain test sample set. Source domain training sample set D s Contains n s A labeled sample in Target domain test sample set D t Contains n t Unlabeled samples in The source domain training sample set mentioned above is a labeled dataset, and the target domain test sample set mentioned above is an unlabeled dataset. Furthermore, the datasets can be cross-validated under different working conditions.

[0070] S3. Standardize the sample data of the source domain training sample set and the target domain test sample set obtained in step S2 to obtain the standardized source domain training sample set and target domain test sample set. The specific standardization formula is as follows:

[0071]

[0072] In the formula, x i For the i-th data sample; μ i It is x i The average value; σ i It is x i Standard deviation; x i (j) represents x i The j-th element.

[0073] S4. Construct an unsupervised transfer diagnosis model (SFTD) for releasing source data. Initially train an SFTD feature encoding module (composed of a convolutional neural network) and an SFTD classification module (composed of fully connected layers and non-linear activation layers) using a standardized source domain training sample set. Then, perform model transfer training on the SFTD feature encoding module only based on a standardized target domain test sample set to complete the transfer from source domain knowledge to target domain knowledge. The specific steps are as follows:

[0074] S41. Construct the SFTD framework. This framework uses a one-dimensional convolutional neural network (1D-CNN) to build the overall network framework, which includes a feature encoding module and a classification module. It performs feature learning on source and target domain data to achieve knowledge transfer between different domains. The two domains include the source domain and the target domain. The feature encoding module includes four convolutional layers, a batch-normalization layer, a non-linear activation function, a pooling layer, a fully connected layer, a batch-normalization layer, and a dropout layer. (Example...) Figure 3 The diagram shows the network structure and parameter settings of the feature encoding module in SFTD. Here, "Maps" represents the number of convolutional kernels in the current convolutional layer; "Pad" indicates zero-padding, which aims to maintain the feature dimensions before and after convolution; "BN" indicates batch normalization; "ReLU" indicates the rectified linear unit activation function; "Maxpool:2×1" indicates a 2×1 max pooling operation; "Stride2" indicates a stride of 2 for the pooling operation; "Linear" represents a fully connected layer; and "Drop" indicates Dropout. The classification module consists of one fully connected layer and one non-linear activation layer, as shown below. Figure 4 The diagram shows the network structure and parameter settings of the classification module in SFTD.

[0075] S42. Generating the Source Model: In the source model generation stage, to enable the network model to correctly classify the source domain training sample set, a source model can be generated using a standard supervised learning paradigm based on the given source data and its corresponding true labels. Simultaneously, to suppress the overfitting and weak fitting problems inherent in traditional cross-entropy, a label-smoothed cross-entropy L is introduced. s ls The loss function used for training the source model is expressed as follows:

[0076]

[0077] In the formula, x s For the source domain sample, y sσ represents the corresponding label of the source neighborhood sample; σ(·) denotes the softmax function. This represents the k-th element in the softmax output of the K-dimensional nonnormalized probability vector; if and only if y s When q = k, k If q is 1, then q is 1; otherwise, q is 1. k =0; α is the smoothing parameter, with a default value of 0.1. f s :X s →Y s Represents the source model, f t :X t →Y t For the learned objective function, inference f represents the proposed diagnostic model. The diagnostic model is initially trained using the standardized source domain training sample set obtained in step S3. The model is then optimized using stochastic gradient descent and iterated until the model converges, completing the first stage of source model generation.

[0078] S43. After the source model is generated, the source data becomes inaccessible. During the source model transfer process, the classification module remains fixed, and only the feature encoding module is updated. Due to the lack of source data, it is not feasible to use a general solution that explicitly matches the distributions of the source and target domains. Therefore, this invention proposes a simple and competitive source-domain-free data model transfer strategy using a self-training mechanism. Unlike traditional strategies, it first utilizes the computation of class prototypes to generate more reliable pseudo-labels. The calculation of the target domain class prototype is as follows:

[0079]

[0080] In the formula, σ k (·) represents the k-th element in the softmax output, X t Represents the target domain space; This represents the target model that has been learned in the previous iteration, which was initialized from the source model.

[0081] Then, pseudo-labels are obtained through distance measurement:

[0082]

[0083] In the formula, These are the initial pseudo-tags.

[0084] After this, the class prototype is updated again using the initial labels, and the pseudo-labels are optimized using the updated class prototype to calculate more accurate pseudo-labels. The process is as follows:

[0085]

[0086] When the parameter is true, the indicator function ζ(·) equals 1; For initial pseudo-labels; x t represents the target domain sample; k represents the predicted category of the target domain sample; is the classifier in the target model.

[0087] No pseudo-label generation algorithm can completely avoid the interference of noisy labels, which leads to underlearning and overfitting problems in standard cross-entropy. To improve noise tolerance, this invention introduces a variant, Normalized Symmetric Cross-Entropy (NSCE), to replace the original CE for robust self-training. The formula for NSCE is:

[0088]

[0089] The right side represents the normalized cross-entropy (NCE) and its inverse representation. The formula for calculating the normalized cross-entropy (NCE) is as follows:

[0090]

[0091] In the formula, p is the prediction probability, i.e., p = σ(f(x)); This indicates a pseudo-tag.

[0092] Then, based on the above normalized cross-entropy, the robust self-training objective function is formed as follows:

[0093]

[0094] S44. To enhance the model's ability to distinguish between different categories, the F-norm of the output matrix is ​​maximized to provide reliable prediction and discrimination capabilities for the target output. The F-norm of the output matrix is ​​maximized as follows:

[0095]

[0096] In the formula, P = [p ij ] B×K This represents the softmax output matrix of a batch of size B.

[0097] Given that incorrect pseudo-labels can impair prediction diversity and hinder model transfer, a constraint is established to maximize the matrix rank (P). Directly calculating the matrix rank is an NP-hard and non-convex problem; however, if ||P|| F ≤1, the convex hull of the rank of matrix P is its nuclear norm ||P|| * Therefore, maximizing the nuclear norm can simultaneously guarantee the discriminativeness and diversity of the target output. A nuclear norm maximization loss function is deployed in the model, with the following form:

[0098] Lnnm =-||P|| * / B (10)

[0099] S45. Using the standardized target domain training sample set obtained in step S3, perform model transfer training on the model, fix the network parameters of the classification module, combine the loss functions in steps S43 and S44, optimize the model feature extraction module using stochastic gradient descent, iterate until the model converges, and complete the source model transfer in the second stage.

[0100] S5. Collect vibration monitoring data of mechanical equipment and input it into the unsupervised migration network of release source data to achieve identification of mechanical health status. Specific Implementation

[0102] The working principle of the present invention will be further explained below with reference to a specific embodiment. In this embodiment, the mechanical device is a rolling bearing. The sampling frequency for collecting the vibration data of the rolling bearing is 10kHz.

[0103] In the verification process of the embodiments, the diagnostic methods based solely on source domain data (Method 1), the transfer diagnostic method based on maximum mean difference (Method 2), the transfer diagnostic method based on relevance rearrangement (Method 3), the adversarial transfer diagnostic method based on domain discriminator (Method 4), and the source hypothesis transfer diagnostic method (Method 5) were selected and compared with the methods of the present invention to verify the effectiveness of the method of the present invention. Among them, Method 1 is a typical representative of deep learning diagnostic models. This model is trained only on labeled source domain data and then directly applied to the target domain. In order to ensure the effectiveness of the comparison, the network framework of Method 1 is the same as the feature encoding module and classification module of the SFTD proposed in the present invention. Method 2 and Method 3 are typical representatives of two distance metric-based transfer methods. Method 2 uses the maximum mean difference of multi-Gaussian kernels to constrain the distribution difference; while Method 3 uses the difference of second-order correlation statistics to achieve feature distribution rearrangement. Method 4 is an adversarial transfer diagnosis method based on a domain discriminator. The feature encoding module and classification module are the same as those in this invention's SFTD, but an additional domain discriminator module consisting of three fully connected layers is added. This module uses the output of the feature encoding module as input and engages in an adversarial game with the feature encoding module. Method 5 is an excellent unsupervised transfer method that releases source data, achieving impressive results in image classification. Therefore, it is applied to the field of transfer diagnosis as a comparative method of this invention.

[0104] The invention was verified by applying it to an embodiment of bearing fault diagnosis in a multi-stage acceleration transmission device.

[0105] Specifically, regarding bearing fault diagnosis in multi-stage acceleration transmission devices, this invention uses experimental data from bearings in multi-stage transmission systems for verification in practical applications. A schematic diagram of the experimental setup is shown below. Figure 5 As shown, the experimental setup includes a drive motor 1, a coupling 2, a planetary gearbox 3, a fixed-axis gearbox 4, a tachometer 5, a bearing housing 6, a vibration sensor 7, and a magnetic powder brake 8. Power is provided by the drive motor 1, and after acceleration via the planetary gearbox 3 and fixed-axis gearbox 4, the power reaches the magnetic powder brake 8 via the bearing test base. The magnetic powder brake 8 can apply different loads to the system. This invention uses vibration data with a sampling frequency of 10kHz for analysis. The health status of rolling bearings includes seven types: normal state, inner ring crack failure, inner ring wear failure, inner and outer ring crack failure, outer ring crack failure, outer ring wear failure, and cage crack failure. During the experiment, each bearing operated under four different conditions (speed approximately 1200 rpm, loads of 0.7 Nm, 1.2 Nm, 2 Nm, and 5 Nm respectively). Data was collected using an accelerometer mounted on the top of the bearing housing at a sampling frequency of 10kHz. For each bearing, 1000 samples were collected under each operating condition, with each sample size being 1024. To simulate a real, noisy industrial environment, noise was mixed into the measurement data to reduce the signal-to-noise ratio. Four domains were constructed based on differences in working conditions, labeled A1, A2, A3, and A4, with different data distributions across the four domains. Therefore, 12 transfer diagnostic tasks were arranged across the four domains, as shown in Table 1. Here, A1→A2 indicates that data samples obtained under working condition A1 were used as the source domain, and data collected under working condition A2 were used as the target domain; the target domain data does not contain label information. During the model training phase, 70% of the samples from each of the two domains were randomly selected for model training. After training, the remaining samples in the target domain were tested and analyzed.

[0106] Table 1 Bearing Migration Diagnosis Tasks for Multistage Transmission Systems

[0107]

[0108]

[0109] Based on this dataset, the proposed method and various comparative methods were executed, and the diagnostic results obtained are shown in Table 2. In this table, "SF" indicates a source-free setting, and "A1→A2" indicates that the data in A1 is the source domain and A2 is the target domain. The best results for each task are shown in bold for visual clarity. Observing these results, we can draw the following conclusions: (1) Without source data for model transfer, the method of the present invention still achieved an average accuracy of 96.89% in 12 tasks, which is better than all existing comparative methods. (2) Undoubtedly, Method1 performed poorly in all diagnostic tasks because it could not do anything about domain drift. In contrast, the four transfer-based methods were able to address data distribution differences and obtain better diagnostic results. (3) Evaluating the existing similar source data release method Method5, it can be found that the method of the present invention has achieved a significant performance improvement and has better model transfer capabilities. Although the diagnostic effect of Method5 is not as good as the adversarial transfer diagnostic method (Method4), it is significantly better than Method2 and Method3. This also shows that the results of Method5 are quite competitive because it does not require source domain data during the migration process, but its results are still far inferior to those of this patent. Therefore, the diagnostic effect of the method of this patent is extremely excellent.

[0110] Table 2. Classification accuracy and standard deviation (%) of various methods for bearing diagnosis in multi-stage transmission systems.

[0111]

[0112]

[0113] By comparing the identification results with those of diagnostic methods based solely on source domain data (Method 1), transfer diagnostic methods based on maximum mean difference (Method 2), transfer diagnostic methods based on relevance rearrangement (Method 3), adversarial transfer diagnostic methods based on domain discriminators (Method 4), and source hypothesis transfer diagnostic methods (Method 5), the mechanical fault identification rate of the present invention is superior to several other existing intelligent diagnostic methods based on deep learning models.

[0114] This invention provides a passive transfer diagnostic method for rotating machinery. Unlike traditional unsupervised transfer diagnostic methods, SFTD does not require source data during the transfer process, thus meeting the needs of more practical diagnostic scenarios where only the source model is available. In SFTD, we introduce a pseudo-label generation algorithm and normalized system cross-entropy loss to achieve more robust self-training. We also deploy an output optimization oriented towards maximizing the nuclear norm to improve the discriminability and diversity of predictions. Ultimately, this achieves cross-domain complex machinery health status identification under multiple operating conditions, unlabeled target samples, and different domain distribution differences, while releasing source domain data. This invention overcomes the shortcomings of conventional deep transfer diagnostic models, reduces the need for source domain data during the diagnostic model transfer process, improves the accuracy of fault identification, and simultaneously reduces the costs of data storage and transmission, as well as privacy protection, in practical engineering scenarios.

[0115] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A mechanical equipment unsupervised migration intelligence fault diagnosis method for releasing source data, characterized in that: It comprises the following steps: S1, for rotating machinery with multiple health states, collecting equipment vibration monitoring data under different working conditions, so that the collected data contains different health state information, and constructing a sample data space and its corresponding label through the monitoring data; S2, the equipment vibration monitoring data under different working conditions in step S1 is divided into source field training sample set and target field test sample set; S3, data standardization is performed on the sample data of the source field training sample set and the target field test sample set obtained in step S2, to obtain the standardized source field training sample set and the target field test sample set; S4, an unsupervised transfer fault diagnosis model releasing source data is established, which comprises the following sub-steps: S41, constructing an unsupervised transfer diagnosis model comprising a feature encoding module and a classification module wherein d is the dimension of the encoded features, and K is the total number of classes; the feature encoding module comprises a multi-layer convolution module and a fully connected layer, and the classification module comprises a fully connected layer and a nonlinear activation layer; S42, generating a source model, generating a source model using a supervised learning paradigm and introducing label smoothed cross entropy as a loss function for training the source model; The loss function in step S42 is: ； ； where, is the source domain sample, is the corresponding label of the source domain sample; denotes the softmax function, denotes the k-th element in the softmax output of the K-dimensional non-normalized probability vector; if and only if , is 1, otherwise, is 0; is a smoothing parameter with a default value of 0.1; denotes the source model, is the learned target function to infer ; S43, generating pseudo-labels using class prototypes, which specifically comprises the following sub-steps: S431, calculating the class prototype of the target field using the following formula: ； In the formula, denotes the kth element in the softmax output, denotes the target domain space; denotes a target model that has been learned in the last iteration, which is obtained by initializing the source model. S432, obtaining the initial pseudo-label by distance measurement: ； In the formula, is the initial pseudo label; S433, update the class prototype using the initial pseudo label, optimize the initial pseudo label, and obtain an optimized pseudo label The specific process is as follows: ； wherein, when the parameter is true, the indicator function equals 1 ; is an initial pseudo-label; denotes a target domain sample; denotes a class that the target domain sample is predicted to belong to; is a classifier in the target model; S434, introducing normalized symmetric cross-entropy to reduce the noise label interference of the pseudo-label: The formula of the normalized symmetric cross-entropy is: ； Where the right side represents the normalized cross-entropy and its inverse representation, and the normalized cross-entropy is specifically: ； In the formula, is the predicted probability, i.e. ; denotes the pseudo label; S435, based on the normalized symmetric cross-entropy of step S434, the following self-training target loss function is formed: ； S44, using the output matrix F norm maximization to provide reliable target output prediction discrimination ability, and adding a kernel norm maximization loss function; S45, using the standardized target field training sample set obtained in step S3 to perform model transfer training on the unsupervised transfer fault diagnosis model, fixing the network parameters of the classification module, combining the self-training target loss function in step S43 and the kernel norm maximization loss function in step S44, using the stochastic gradient descent method to optimize the model feature extraction module, and iterating until the model converges, completing the source model transfer; S5, collecting vibration monitoring data of mechanical equipment, inputting into the unsupervised transfer fault diagnosis model, and realizing mechanical health state discrimination.

2. The mechanical equipment unsupervised migration intelligence fault diagnosis method for releasing source data according to claim 1, characterized in that: The output matrix F norm maximization expression in step S44 is: ； wherein denotes the softmax output matrix of a batch of size B.

3. The mechanical equipment unsupervised migration intelligence fault diagnosis method of release source data according to claim 2, characterized in that: The kernel norm maximization loss function in step S44 is: ； In the formula, the convex envelope of the rank of the matrix P is the nuclear norm of P .

4. The mechanical equipment unsupervised migration intelligence fault diagnosis method of release source data according to claim 1, characterized in that: The feature encoding module in step S41 comprises four convolutional layers, a Batch-standardization layer, a nonlinear activation function, a pooling layer, a fully connected layer, and a Dropout layer; the classification module comprises a fully connected layer and a nonlinear activation layer.

5. The method of claim 1, wherein the method further comprises: The source domain in step S2 is assumed comprising one labeled sample wherein , ; Target field comprising one unlabeled sample wherein .

6. The mechanical equipment unsupervised migration intelligence fault diagnosis method of release source data according to claim 1, characterized in that: The formula for data standardization in step S3 is specifically: ； In the formula: For the first One data sample; for The average value; for Standard deviation; for The j-th element.

7. The mechanical equipment unsupervised migration intelligence fault diagnosis method of release source data according to claim 1, characterized in that: The mechanical equipment in step S1 comprises a multi-stage accelerating transmission device bearing fault simulation test bench.

8. The mechanical equipment unsupervised migration intelligence fault diagnosis method of release source data according to claim 7, characterized in that: The sampling frequency of the multi-stage accelerating transmission device vibration data is 10 kHz.

Images