Vibrating screen cross-domain fault diagnosis method and system for open scene

Abstract

The invention discloses a vibrating screen cross-domain fault diagnosis method and system oriented to an open scene, and belongs to the technical field of vibrating screen fault diagnosis. The method comprises the following steps: collecting fault data of different models of vibrating screens, and constructing a fault data set through data balance processing and multi-dimensional vibration feature extraction and selection; designing a constraint Brown generator, and introducing direction constraint and intra-domain intra-class-inter-class constraint to generate a pseudo-unknown sample which keeps an effective distance with a known class sample of a source domain; a weighted feature distribution difference calculation strategy is designed by using dynamic time warping and openness, and different domain alignment weights are given to known / unknown category samples; and establishing a dual-path knowledge domain adaptive fault diagnosis model by adopting an incremental random weight network, and completing cross-domain fault diagnosis of the vibrating screen in an open set environment. The problem of superposition of category difference and distribution difference in an open set scene is effectively solved, the precision, robustness and generalization ability of the diagnosis model are improved, and the method is suitable for various types of vibrating screens and working conditions.

Description

Technical Field

[0001] This invention belongs to the field of vibrating screen fault diagnosis technology, specifically relating to a cross-domain fault diagnosis method and system for vibrating screens in open scenarios. Background Technology

[0002] Vibrating screens are core equipment for material screening and grading in coal washing and processing. Their stable operation directly determines the continuity, production efficiency, and product quality of the coal washing production line. In actual operation, vibrating screens operate under harsh environments of high-frequency vibration and material impact, making them prone to various malfunctions such as drive shaft failure, vibrator failure, loose pressure bars, and loose screen plates. Once a malfunction occurs, it can lead to reduced screening efficiency and substandard product quality, or even cause cascading damage to equipment, production line shutdown, and even safety accidents, resulting in huge economic losses for enterprises. Therefore, achieving real-time, accurate, and reliable fault diagnosis of vibrating screens is of significant engineering and economic value for the safe production and efficient operation of the coal washing industry.

[0003] In recent years, deep learning-based intelligent fault diagnosis technology has made breakthrough progress in the field of fault diagnosis for rotating machinery such as bearings and gearboxes, thanks to its powerful automatic feature extraction and pattern recognition capabilities. However, when applying this technology directly to the fault diagnosis of vibrating screens, it faces the dual challenges of domain offset and open-set faults. On the one hand, vibration signals collected from different models of vibrating screens, under different operating conditions, and at different sensor installation locations exhibit significant differences in data distribution, i.e., domain offset. This causes a diagnostic model trained on one vibrating screen (source domain) to have a significantly reduced diagnostic accuracy when deployed to another vibrating screen (target domain). On the other hand, vibrating screens may encounter unknown fault types not seen during the training phase during actual operation, i.e., open-set faults. Traditional closed-set fault diagnosis models may incorrectly classify unknown faults as known faults, resulting in extremely high false alarm and false negative rates, which cannot meet the actual engineering needs.

[0004] To address the aforementioned issues, existing technologies have proposed solutions such as domain adversarial networks, open-set recognition models based on classifier confidence, and open-set domain adaptation methods. However, these methods still have several shortcomings when facing the problem of overlapping category differences and domain offsets in the open-set scenario of vibrating screens: First, existing domain adaptation methods are difficult to effectively align with the feature distribution of vibrating screens under different operating conditions, resulting in low classification accuracy for known category faults; Second, open-set recognition models cannot effectively distinguish between inter-domain distribution differences and category differences, easily misclassifying unknown category faults as known categories, leading to a high false alarm rate; Third, most methods rely on the strong assumption of a complete match between the source and target domain fault categories, resulting in poor generalization ability and an inability to adapt to the complex and ever-changing actual operating conditions of vibrating screens. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a cross-domain fault diagnosis method and system for vibrating screens in open scenarios. This effectively solves the challenge of superimposed category differences and distribution differences in open set scenarios, and significantly improves the accuracy, robustness and generalization ability of the vibrating screen fault diagnosis model.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a cross-domain fault diagnosis method for vibrating screens in open scenarios, the method comprising the following steps: Step S1, collecting fault data of different models of vibrating screens, performing intra-class sample balancing processing through a data balancing strategy, and constructing a fault dataset by combining multi-dimensional vibration feature extraction and selection.

[0007] Step S2: Design a constrained Brownian generator, introduce directional constraints and intra-domain and inter-domain constraints to optimize the data generation process, and ensure that pseudo-unknown samples are generated that maintain an effective distance from known class samples in the source domain while satisfying the source domain distribution.

[0008] Step S3: Quantize the similarity of fault features between the source and target domains using dynamic time warping, introduce the concept of openness, design a weighted feature distribution difference calculation strategy, and assign different domain alignment weights to samples of known and unknown categories.

[0009] Step S4: An incremental random network is used to establish a dual-path knowledge domain adaptive fault diagnosis model, thereby completing the cross-domain fault diagnosis of the vibrating screen in an open set environment.

[0010] As a further technical solution of the present invention: In step S1, the process of sampling fault data includes the following steps: Step S11, using a combined sampling method to perform balanced sampling of the vibrating screen fault data, and constructing a balanced dataset. The dataset was then normalized to obtain the vibration dataset. .

[0011] Step S12: Extract vibration dataset using statistical analysis. The vibration feature set is obtained by statistically analyzing the time domain, frequency domain, and time-frequency domain. Robust features were selected using the constructed feature evaluation index to form a robust vibration sample set. .

[0012] As a further technical solution of the present invention: In step S12, the constructed feature evaluation index expression is as follows: In the formula, Indicates the first 1 eigenvector Indicates the first 1 eigenvector Indicates the number of features. Indicates the first Each feature vector and fault category label The grey relational coefficient between them Indicates except the first The set of other eigenvectors besides the eigenvectors. Represents a set Number of features Indicates that in the known Under the conditions, the first The conditional entropy of each eigenvector.

[0013] As a further technical solution of the present invention: Step S2 involves inputting known class samples from the source domain. With various weight parameters, output out-of-distribution samples and pseudo-source domain sample set Repeat the out-of-region sample generation steps until the source domain unknown class sample categories are expanded to... There are various weight parameters, including the number of unknown categories of the target. Number of samples generated per class Boundary distance offset distance Boundary softness Guiding weight Noise intensity The specific steps are as follows: Step S21, use an autoencoder to process known class samples from the source domain. Mapped to latent space In the process, principal components of the latent space are calculated, and a sample is randomly selected from the latent space. Let the initially generated data points The minimum distance to a known class sample is .

[0014] Step S22, for the selected sample Iteratively add offsets, primarily based on directional constraints, and perform constrained Brownian motion until the transformed data points are reached. minimum distance Exceeding the boundary distance Meanwhile, considering the boundary softness Soft boundaries are implemented using multiple conditions containing intra-domain and inter-class constraints and likelihood rejection, i.e., when near There is a certain probability that the process will stop early and save the generated samples that meet the conditions.

[0015] Step S23: Use a decoder to decode the generated samples from the latent space to obtain the final pseudo-source domain sample set. .

[0016] As a further technical solution of the present invention: the offset mainly based on direction constraints, the intra-domain intra-class-inter-class constraints, and the multiple conditions for rejection likelihood in step S22 are expressed as follows: the offset mainly based on direction constraints: Intra-domain and intra-class / inter-class constraints: Reject the likelihood condition: ; Update the location of generated samples: In the formula, This indicates that the mean is 0 and the variance is . Gaussian distribution, For the first in the latent space of the source domain Class center of class sample, For the generated pseudo-unknown samples, This represents the minimum distance between pseudo-unknown class samples and known class samples in the latent space of the source domain. Indicates the distance within the current class. Represents the minimum inter-class distance. This indicates a rejection of the likelihood function. , , , and These are the control parameters in boundary distance, offset distance, guided weight, noise intensity, and rejection likelihood, respectively.

[0017] As a further technical solution of the present invention: Step S3 is as follows: Step S31, Dynamic Time Warping (DTW) is used to calculate the similarity between feature samples of the source domain and the target domain, and the calculation formula is as follows: In the formula, Indicates the first The feature sample and the first The distance between corresponding elements in each feature sample.

[0018] Step S32: Establish an openness assessment model, using the number of fault categories in the source and target domains to calculate openness. The calculation formula is as follows: In the formula, This indicates the number of known fault categories. This represents the total number of known samples in the source domain. Represents the total number of samples in the target domain, where Represents the source domain. If the target domain is represented, then Indicates that there is in the source domain Each fault category Indicates that the target domain contains Fault categories.

[0019] Step S33: Design dynamic weights using dynamic time warping and openness. In the formula, This represents the balancing parameter, used to balance the impact of openness and similarity on dynamic weights.

[0020] Step S34: Design a dynamic weighted feature distribution difference calculation strategy.

[0021] As a further technical solution of the present invention: the formula for calculating the dynamic weighted feature distribution difference in step S34 is as follows: ; ; In the formula, and Let them represent the pseudo-source domain sample set and the target domain sample set, respectively. This represents the balance parameter between global distribution variance and weighted local distribution variance. and These represent the number of samples in the pseudo-source domain and the number of samples in the target domain, respectively. Represents a space mapping function, Indicates dynamic weights, and Let represent the centering matrices of the pseudo-source domain and the target domain, respectively.

[0022] As a further technical solution of the present invention: step S4 further includes: step S41, using the pseudo-source domain sample set obtained in step S2 The objective function of the first-stage knowledge model is as follows: In the formula, the first term is the regularization term, the second term is the empirical error term for the pseudo-source domain samples, and the third term is the intra-class and inter-class difference term. The output weights of the first-stage knowledge model are shown in the upper right corner. " indicates transpose, and For the first One pseudo-source domain sample and its corresponding label, For feature mapping function, and These are the intra-class distance matrix and inter-class distance matrix within the domain. and To weigh the parameters.

[0023] Step S42: Using a small number of target domain samples and combining the knowledge from the first-stage knowledge model in step S41, a second-stage knowledge domain adaptation model is established. The objective function of the constructed model is as follows: In the formula, the first term is the regularization term, the second term is the empirical error loss term for the target domain, the third term is the knowledge distillation loss term, and the fourth term is the weighted distribution difference loss term. The output weights of the second-stage knowledge domain adaptation model are used. The number of samples in the target domain. and For the first Each target domain sample and its corresponding ground truth label For the sample sets of pseudo-source and target domains, the first... One sample, The weighted distribution difference matrix, , , and To weigh the parameters.

[0024] As a further technical solution of the present invention: the update formula for the output weights in step S42 is constructed using the following formula: In the formula, It is the identity matrix. , , These are the hidden layer outputs of the sample sets from the target domain, pseudo-source domain, pseudo-source domain, and target domain, respectively. For the actual output of the target domain, and The soft output (probability output predicted by the knowledge model in the first stage) and the true output of the pseudo-source domain sample are described respectively.

[0025] A cross-domain fault diagnosis system for vibrating screens in open scenarios based on the above method is provided. The system includes: a data acquisition and processing module for acquiring fault data of different models of vibrating screens, performing intra-class sample balancing on the acquired raw fault data, and extracting and selecting multi-scale vibration features from the balanced data to construct a robust fault dataset.

[0026] Data generation module: Used to generate pseudo-unknown fault samples that are missing from the source domain data. Based on directional constraints and intra-domain and inter-domain constraints, a Brownian generator generates pseudo-unknown samples that maintain an effective distance from known class samples in the source domain.

[0027] Model training module: Based on the constructed pseudo-source domain samples and a small number of target domain samples, it constructs a cross-domain fault diagnosis model for vibrating screens using a dual-path knowledge domain adaptation network, thereby completing the cross-domain fault diagnosis of vibrating screens in an open-set environment.

[0028] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. The present invention generates pseudo-unknown samples that satisfy the source domain distribution and maintain an effective distance from known class samples by constraining the Brownian generator, which effectively solves the problem of the difference between the source domain and the target domain categories in open scenarios, makes up for the lack of unknown fault category data, and provides a data foundation for open set fault diagnosis.

[0029] 2. This invention combines dynamic time warping and openness to design a dynamically weighted feature distribution difference calculation strategy, which assigns different domain alignment weights to samples of known and unknown categories, achieving accurate domain alignment between the source and target domains, effectively solving the domain offset problem, and improving the model's cross-domain adaptability.

[0030] 3. This invention adopts a dual-path knowledge domain adaptation model. The first path learns source domain fault knowledge, and the second path realizes the transfer of source domain knowledge to the target domain. By combining knowledge distillation and weighted distribution difference loss, the source domain knowledge is effectively preserved and accurate domain adaptation is achieved, which significantly improves the accuracy of cross-domain fault diagnosis of vibrating screen in open set scenarios.

[0031] 4. The method and system of the present invention are applicable to various types of vibrating screens and different operating conditions, and have good versatility, robustness and scalability. They are easy to promote and apply in the coal washing industry, and can effectively improve the operation and maintenance efficiency of vibrating screen equipment, extend the service life of equipment, reduce the operating and maintenance costs of enterprises, and reduce economic losses caused by downtime due to failure. Attached Figure Description

[0032] Figure 1 This is a flowchart of the cross-domain fault diagnosis method for vibrating screens in open scenarios according to the present invention.

[0033] Figure 2 This is a distribution diagram of the imbalance data collected from different vibrating screens.

[0034] Figure 3 A distribution map of private fault data generated for different tasks by the soft Brown generator.

[0035] Figure 4 This is a bar chart showing the diagnostic results of the method proposed in this invention. Detailed Implementation

[0036] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figures 1-4 As shown, this invention proposes a technical solution from two aspects: hardware device design and method implementation. The system serves as the hardware carrier of the method, and the method achieves full-process management of hardware breakpoints through the collaborative work of various modules of the system.

[0038] A cross-domain fault diagnosis method for vibrating screens in open scenarios includes the following steps: S1: Obtain fault data of different models of vibrating screens, and to address the problem of unbalanced distribution of intra-class samples in the data, adopt a data balancing strategy to balance intra-class samples, and construct a fault dataset by combining multi-dimensional vibration feature extraction and selection.

[0039] Step S1 is as follows: S11, use a single-axis vibration acceleration sensor to collect fault data of different models of vibrating screens, such as... Figure 2 As shown, the data collection points include the motor, drive shaft, vibrator, middle of the side plate, feed end, discharge end, and spring components. A combined sampling method (random oversampling and random undersampling) is used to perform balanced sampling on the vibrating screen data to construct a balanced dataset. For balanced datasets Normalization is performed to obtain the vibration dataset. .

[0040] S12, extract vibration dataset using statistical analysis. The statistical characteristics in the time domain, frequency domain, and time-frequency domain include 65 indicators such as maximum value, minimum value, mean, root mean square value, standard deviation, frequency band energy distribution, spectral peak value, and wavelet coefficients, resulting in a vibration feature set. We construct feature evaluation indicators, rank these indicators, and select features whose indicator values ​​are greater than the overall indicator mean to form a robust vibration sample set. The final sample consists of 100 samples for each type of fault.

[0041] The expression for the feature evaluation index is as follows: In the formula, Indicates the first 1 eigenvector Indicates the first 1 eigenvector Indicates the number of features. Indicates the first Each feature vector and fault category label The grey relational coefficient between them Indicates except the first The set of other eigenvectors besides the eigenvectors. Represents a set Number of features Indicates that in the known Under the conditions, the first The conditional entropy of each eigenvector.

[0042] S2: Design a constrained Brownian generator, introduce directional constraints and intra-domain and intra-class constraints to optimize the data generation process, and ensure that pseudo-unknown samples are generated that maintain an effective distance from known class samples in the source domain while satisfying the source domain distribution.

[0043] Step S2 is as follows: Input: Sample set of known classes in the source domain Number of unknown target categories Number of samples generated per class Boundary distance offset distance Boundary softness Guiding weight Noise intensity .

[0044] Output: Out-of-distribution samples, i.e., the sample set of pseudo-unknown samples in the source domain. Pseudo-source domain sample set (Includes known class samples from the source domain and generated pseudo-unknown class samples).

[0045] S21, using a sparse autoencoder to map known class samples from the source domain to the latent space. In the process, principal components of the latent space are calculated, and a sample is randomly selected from the latent space. Let the initially generated data points The minimum distance to a known class sample is .

[0046] S22, for the selected sample Iteratively add offsets, primarily based on directional constraints, and perform constrained Brownian motion until the transformed data points are reached. minimum distance Exceeding the boundary distance Meanwhile, considering the boundary softness Soft boundaries are implemented using multiple conditions containing intra-domain and inter-class constraints and likelihood rejection, i.e., when near There is a certain probability of stopping early; the offset mainly based on direction constraints, intra-domain and intra-class / inter-class constraints, and multiple conditions for rejection likelihood are expressed as follows: Offset mainly based on direction constraints: Intra-domain and intra-class / inter-class constraints: Reject the likelihood condition: ; Update the location of generated samples: In the formula, This indicates that the mean is 0 and the variance is . Gaussian distribution, For the first in the latent space of the source domain Class center of class sample, For the generated pseudo-unknown samples, This represents the minimum distance between pseudo-unknown class samples and known class samples in the latent space of the source domain. Indicates the distance within the current class. Represents the minimum inter-class distance. This indicates a rejection of the likelihood function. This indicates the rejection of the control parameters in the likelihood.

[0047] S23, the decoder is used to decode the generated samples from the latent space to obtain the final pseudo-source domain sample set. (Includes known class samples from the source domain and generated pseudo-unknown class samples).

[0048] S3 utilizes Dynamic Time Warping (DTW) to quantify the similarity of fault features between the source and target domains, and introduces the concept of openness to design a dynamically weighted feature distribution difference calculation strategy. Different domain alignment weights are assigned to known-class samples and unknown-class samples. Step S3 is as follows: S31, DTW is used to calculate the similarity between feature samples in the source and target domains. The calculation formula is as follows: In the formula, Indicates the first The feature sample and the first The distance between corresponding elements in each feature sample.

[0049] S32. Establish an openness assessment model, using the number of fault categories in the source and target domains to calculate openness. The calculation formula is as follows: In the formula, This indicates the number of known fault categories. This represents the total number of known samples in the source domain. This represents the total number of samples in the target domain.

[0050] S33, designing dynamic weights using dynamic time warping and openness. Balance parameters The selection range is .

[0051] S34, Design a dynamic weighted feature distribution difference calculation strategy, the expression of which is as follows: ; ; In the formula, and Let them represent the pseudo-source domain sample set and the target domain sample set, respectively. This represents the balance parameter between global distribution variance and weighted local distribution variance. and These represent the number of samples in the pseudo-source domain and the number of samples in the target domain, respectively. Represents a space mapping function, and For random weights and random biases, Indicates dynamic weights. and Let represent the centering matrices of the pseudo-source domain and the target domain, respectively.

[0052] S4, establish a dual-path knowledge domain adaptive fault diagnosis model to complete the cross-domain fault diagnosis of the vibrating screen in an open-set environment. Step S4 is as follows: S41, use the pseudo-source domain sample set obtained in step S2. The objective function of the first-stage knowledge model is as follows: In the formula, the first term is the model complexity, the second term is the empirical error term for pseudo-source domain samples, and the third term is the intra-class and inter-class difference term. The output weights of the first-stage knowledge model (top right corner) "Indicates transpose" and For the first One pseudo-source domain sample and its corresponding label, For feature mapping function, and These are the intra-class distance matrix and inter-class distance matrix within the domain. and To weigh the parameters.

[0053] S42, using a small number of target domain samples and combining the knowledge from the first-stage knowledge model in step S41, establish the second-stage knowledge domain adaptation model. The objective function of the constructed model is as follows: In the formula, the first term is the model complexity, the second term is the empirical error loss of the target domain, the third term is the knowledge distillation loss, and the fourth term is the weighted distribution difference loss. The output weights of the second-stage knowledge domain adaptation model are used. The number of samples in the target domain. and For the first Each target domain sample and its corresponding ground truth label For the sample sets of pseudo-source and target domains, the first... One sample, The weighted distribution difference matrix, , , and To weigh the parameters.

[0054] The output weights of step S42 are constructed using the following formula: In the formula, It is the identity matrix. , , The hidden layer outputs are respectively the target domain, pseudo-source domain, pseudo-source domain, and target domain sample sets. For the actual output of the target domain, and The soft output (probability output predicted by the first path model) and the true output of the pseudo-source domain samples are described respectively; the above... , , , and Trade-off parameters from Select from, parameters The selection range is .

[0055] To illustrate the performance of this invention in cross-domain fault diagnosis of vibrating screens in open-set scenarios, a real-world example of a vibrating screen is used. Four vibrating screens were selected for the experiment. Table 1 shows the information of the vibrating screen fault dataset, and Table 2 shows the information of the open-set fault diagnosis task for the vibrating screen dataset. Figure 3 The diagram shows the distribution of unknown faults and original fault data for different tasks generated by the soft Brownian offset generator. It can be seen that the method proposed in this invention can generate unknown fault data missing in the source domain effectively, and the generated data quality is high even when the migration task is complex. Furthermore, this invention demonstrates the diagnostic results of the proposed cross-domain fault diagnosis method for vibrating screens in open scenarios, such as... Figure 4 As shown, this includes the overall model mean accuracy (OS), the mean accuracy for known fault classes (OS*), the mean accuracy for unknown fault classes (UNK), and the harmonic mean (HOS). From Figure 4 As can be seen, the present invention can achieve high accuracy in different migration tasks, all exceeding 0.9, and can also classify unknown faults while accurately identifying known fault types.

[0056] Table 1: Information on Vibrating Screen Fault Data Set

[0057]

[0058] Table 2: Information regarding the open set diagnosis task for vibrating screen fault datasets

[0059]

[0060] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0061] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment have been appropriately combined to form other embodiments that are easy for those skilled in the art to understand.

Claims

1. A method for cross-domain fault diagnosis of vibrating screens in open scenarios, characterized in that, The method includes the following steps: Step S1: Collect fault data of different types of vibrating screens, perform intra-class sample balancing processing through data balancing strategy, and construct fault dataset by combining multi-dimensional vibration feature extraction and selection. Step S2: Design a constrained Brownian generator, introduce directional constraints and intra-domain and inter-domain constraints to optimize the data generation process, and ensure that pseudo-unknown samples are generated that maintain an effective distance from known class samples in the source domain while satisfying the source domain distribution. Step S3: Quantify the similarity of fault features between the source domain and the target domain using dynamic time warping, and introduce the concept of openness to design a weighted feature distribution difference calculation strategy, assigning different domain alignment weights to known category samples and unknown category samples. Step S4: An incremental random network is used to establish a dual-path knowledge domain adaptive fault diagnosis model, thereby completing the cross-domain fault diagnosis of the vibrating screen in an open set environment.

2. The method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 1, characterized in that, In step S1, the process of sampling fault data includes the following steps: Step S11: Use a combined sampling method to perform balanced sampling on the vibrating screen fault data to construct a balanced dataset. The dataset was then normalized to obtain the vibration dataset. ; Step S12: Extract vibration dataset using statistical analysis. The vibration feature set is obtained by statistically analyzing the time domain, frequency domain, and time-frequency domain. Robust features were selected using the constructed feature evaluation index to form a robust vibration sample set. .

3. The method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 2, characterized in that, In step S12, the constructed feature evaluation index expression is as follows: ； In the formula, Indicates the first 1 eigenvector Indicates the first 1 eigenvector Indicates the number of features. Indicates the first Each feature vector and fault category label The grey relational coefficient between them Indicates except the first The set of other eigenvectors besides the eigenvectors. Represents a set Number of features Indicates that in the known Under the conditions, the first The conditional entropy of each eigenvector.

4. The method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 1, characterized in that, Step S2 involves inputting known class samples from the source domain. With various weight parameters, output out-of-distribution samples and pseudo-source domain sample set ; Repeat the out-of-distribution sample generation steps until the source domain unknown class sample categories are expanded to... There are various weight parameters, including the number of unknown categories of the target. Number of samples generated per class Boundary distance offset distance Boundary softness Guiding weight Noise intensity The specific steps are as follows: Step S21, use an autoencoder to process known class samples from the source domain. Mapped to latent space In the process, principal components of the latent space are calculated, and a sample is randomly selected from the latent space. Let the initially generated data points The minimum distance to a known class sample is ; Step S22, for the selected sample Iteratively add offsets, primarily based on directional constraints, and perform constrained Brownian motion until the transformed data points are reached. minimum distance Exceeding the boundary distance Meanwhile, considering the boundary softness Soft boundaries are implemented using multiple conditions containing intra-domain and inter-class constraints and likelihood rejection, i.e., when near If necessary, stop early and save the generated samples that meet the conditions; Step S23: Use a decoder to decode the generated samples from the latent space to obtain the final pseudo-source domain sample set. .

5. The method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 4, characterized in that, In step S22, the offset, mainly based on direction constraints, the intra-domain intra-class and inter-class constraints, and the multiple conditions for rejection likelihood are represented as follows: Offset primarily constrained by direction: ; Intradomain intraclass-interclass constraints: ; Reject the likelihood condition: ; Update the location of generated samples: ; In the formula, This indicates that the mean is 0 and the variance is . Gaussian distribution, For the first in the latent space of the source domain Class center of class sample, For the generated pseudo-unknown samples, This represents the minimum distance between pseudo-unknown class samples and known class samples in the latent space of the source domain. Indicates the distance within the current class. Represents the minimum inter-class distance. This indicates a rejection of the likelihood function. , , , and These are the control parameters in boundary distance, offset distance, guided weight, noise intensity, and rejection likelihood, respectively.

6. The method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 1, characterized in that, Step S3 is as follows: Step S31: Calculate the similarity between feature samples in the source and target domains using Dynamic Time Warping (DTW). The calculation formula is as follows: ； In the formula, Indicates the first The feature sample and the first The distance between corresponding elements in each feature sample; Step S32: Establish an openness assessment model, using the number of fault categories in the source and target domains to calculate openness. The calculation formula is as follows: ； In the formula, This indicates the number of known fault categories. This represents the total number of known samples in the source domain. Represents the total number of samples in the target domain, where Represents the source domain. If the target domain is represented, then Indicates that there is in the source domain Each fault category Indicates that the target domain contains One fault category; Step S33: Design dynamic weights using dynamic time warping and openness. ; In the formula, This represents the balancing parameter, used to balance the impact of openness and similarity on the dynamic weights; Step S34: Design a dynamic weighted feature distribution difference calculation strategy.

7. A method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 6, characterized in that, The formula for calculating the difference in the dynamically weighted feature distribution in step S34 is as follows: ； ； ； In the formula, and Let them represent the pseudo-source domain sample set and the target domain sample set, respectively. This represents the balance parameter between global distribution variance and weighted local distribution variance. and These represent the number of samples in the pseudo-source domain and the number of samples in the target domain, respectively. Represents a space mapping function, Indicates dynamic weights, and Let represent the centering matrices of the pseudo-source domain and the target domain, respectively.

8. A method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 1, characterized in that, Step S4 further includes: Step S41, using the pseudo-source domain sample set obtained in step S2 The objective function of the first-stage knowledge model is as follows: ； In the formula, the first term is the regularization term, the second term is the empirical error term for the pseudo-source domain samples, and the third term is the intra-class and inter-class difference term. The output weights of the first-stage knowledge model are shown in the upper right corner. " indicates transpose, and For the first One pseudo-source domain sample and its corresponding label, For feature mapping function, and These are the intra-class distance matrix and inter-class distance matrix within the domain. and To weigh the parameters; Step S42: Using a small number of target domain samples and combining the knowledge from the first-stage knowledge model in step S41, a second-stage knowledge domain adaptation model is established. The objective function of the constructed model is as follows: ； In the formula, the first term is the regularization term, the second term is the empirical error loss term for the target domain, the third term is the knowledge distillation loss term, and the fourth term is the weighted distribution difference loss term. The output weights of the second-stage knowledge domain adaptation model are used. The number of samples in the target domain. and For the first Each target domain sample and its corresponding ground truth label For the sample sets of pseudo-source and target domains, the first... One sample, The weighted distribution difference matrix, , , and To weigh the parameters.

9. A method for cross-domain fault diagnosis of vibrating screens in open scenarios according to claim 8, characterized in that, The update formula for the output weights in step S42 is constructed using the following formula: ； In the formula, It is the identity matrix. , , The hidden layer outputs are respectively the target domain, pseudo-source domain, pseudo-source domain, and target domain sample sets. For the actual output of the target domain, and Describe the soft output and true output of the pseudo-source domain sample, respectively.

10. A cross-domain fault diagnosis system for vibrating screens in open scenarios, based on the cross-domain fault diagnosis method for vibrating screens in open scenarios as described in any one of claims 1-9, characterized in that, The system includes: Data acquisition and processing module: used to collect fault data of different types of vibrating screens, perform intra-class sample balancing on the collected raw fault data, and extract and select multi-scale vibration features from the balanced data to construct a robust fault dataset. Data generation module: used to generate pseudo-unknown fault samples where the source domain data is missing; based on the Brownian generator with directional constraints and intra-class and inter-class constraints, it generates pseudo-unknown samples that maintain an effective distance from the known class samples in the source domain. Model training module: Based on the constructed pseudo-source domain samples and a small number of target domain samples, it constructs a cross-domain fault diagnosis model for vibrating screens using a dual-path knowledge domain adaptation network, thereby completing the cross-domain fault diagnosis of vibrating screens in an open-set environment.

Images