A method and system for enhanced domain adaptive cross-domain fault diagnosis of a vibrating screen

Abstract

The application discloses a kind of enhanced field adaptation's vibrating screen cross-domain fault diagnosis method, comprising: the different field vibrating screen multidimensional comprehensive fault data set obtained is divided into source domain feature training set, target domain feature training set and target domain feature test set;Source domain feature training set and target domain feature training set are mapped to random feature space and are processed to joint distribution difference, to execute the field and class alignment of feature;Source domain feature training set is processed to field feature discrimination, to enhance the intra-class tightness and inter-class separability of feature;According to the joint distribution difference processing result and the field feature discrimination processing result, based on the enhanced field adaptation incremental random vector function chain network, a vibrating screen cross-domain fault diagnosis model is constructed, which is used to predict the fault class of unknown label in target domain. The application effectively overcomes the challenge of domain shift, and also significantly improves the performance and generalization ability of the vibrating screen fault diagnosis model.

Description

Technical Field

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

[0002] As a key piece of equipment in industrial sorting processes, the stability of vibrating screens directly impacts production efficiency and product quality. Because they often operate continuously for extended periods in harsh environments and withstand complex loads under varying conditions, vibrating screens inevitably experience various malfunctions. Failure to diagnose and address these malfunctions promptly can lead to decreased equipment performance and even safety accidents. Therefore, effective fault diagnosis of vibrating screens to ensure their normal operation and extend their service life is a crucial and urgent issue that needs to be addressed in the industrial sector.

[0003] Currently, fault diagnosis methods for vibrating screens mainly fall into three categories: mechanistic model-based methods, expert knowledge-based methods, and data-driven methods. However, these methods share a common premise in practical applications: the training data and test data must follow the same distribution. But due to factors such as the diversity of equipment models, environmental noise interference, and the natural aging of equipment, this assumption is often difficult, or even nearly impossible, to achieve. Therefore, researching cross-domain fault diagnosis methods for vibrating screens is particularly important.

[0004] Cross-domain fault diagnosis methods can be categorized into instance-based, model-based, and feature-based methods. Instance-based methods focus on effectively assigning weights to labeled data instances in the source domain, making the instance distribution in the source domain closer to that in the target domain, thereby constructing a high-accuracy and reliable learning model in the target domain. Model-based methods accelerate the training process for new tasks by reusing the parameters and structure of pre-trained models, assuming that the source and target tasks share common knowledge at the model level. Feature-based methods align source and target domain data in the feature space, minimizing the difference in metric feature distributions between the two domains, allowing the feature knowledge learned in the source domain to better adapt to the fault modes in the target domain. Domain adaptation, as a typical representative of this type of method, has been extensively studied.

[0005] However, most existing cross-domain fault diagnosis methods are performed in deep neural networks. For vibrating screens, the data contained in the collected vibration signals are complex and diverse, making the feature extraction process extremely complex and time-consuming. In addition, the difference in the distribution of fault data under different working conditions can also lead to domain shift problems in the complex and variable working environment of vibrating screens. Therefore, it is urgent to propose a cross-domain fault diagnosis method suitable for vibrating screens. Summary of the Invention

[0006] This invention proposes an enhanced domain-adaptive cross-domain fault diagnosis method and system for vibrating screens, which effectively overcomes the challenge of domain offset and significantly improves the performance and generalization ability of the vibrating screen fault diagnosis model, providing certain technical support for promoting the automation, informatization and intelligent development of coal washing and processing.

[0007] In a first aspect, the present invention discloses an enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens, the method comprising the following steps:

[0008] Step A: Divide the acquired multidimensional comprehensive fault dataset of vibrating screens from different fields into source domain feature training set, target domain feature training set, and target domain feature test set;

[0009] Step B involves mapping the source domain feature training set and the target domain feature training set to a random feature space for joint distribution difference processing to perform feature neighborhood and class alignment.

[0010] Step C involves performing domain feature discrimination processing on the source domain feature training set to enhance the intra-class tightness and inter-class separability of features;

[0011] Step D: Based on the joint distribution difference processing results and the domain feature discrimination processing results, a cross-domain fault diagnosis model for vibrating screen is constructed based on the enhanced domain adaptive incremental random vector function chain network to predict the fault category of unknown labels in the target domain.

[0012] Furthermore, in step A, the process of obtaining the multidimensional comprehensive fault dataset of vibrating screens from different fields includes the following steps:

[0013] In-depth analysis of the operating mechanism of vibrating screens under harsh working environments and heavy loads is conducted to identify and predict common failure types.

[0014] Based on common fault types of vibrating screens, acceleration-temperature composite sensors are installed at different key fault monitoring points of vibrating screens to collect comprehensive fault datasets of different monitoring points of vibrating screens in different fields, and generate multidimensional comprehensive fault datasets of vibrating screens in different fields.

[0015] Furthermore, the common failure types of the vibrating screen include at least:

[0016] The faults include: screen box body faults such as side plate cracking, lifting beam cracking, drive beam cracking, load-bearing beam cracking, small beam cracking, T-shaped steel cracking, small beam loosening, and T-shaped steel loosening; vibrator faults such as bearing damage and gear damage; drive component faults such as bearing damage and bearing seat loosening; screen plate assembly faults such as screen plate loosening, side pressure plate loosening, feed and discharge blind plates loosening, and rail seat loosening; and spring cracking faults.

[0017] Furthermore, the key fault monitoring points of the vibrating screen include at least the main motor, drive shaft, vibrating screen motor-side exciter, vibrating screen non-motor-side exciter, the middle of the vibrating screen motor-side side plate, the middle of the vibrating screen motor-side non-side plate, the vibrating screen non-motor-side feed end and discharge end.

[0018] Step A further includes:

[0019] Multi-domain combined statistical analysis and principal component analysis methods are used to perform multi-dimensional robustness processing on the multi-dimensional comprehensive fault dataset of vibrating screens in different fields in the time domain, frequency domain, and time-frequency domain. The generated highly robust feature set is divided into source domain feature training set, target domain feature training set, and target domain feature test set.

[0020] Furthermore, the process of performing multi-dimensional robustness processing on the multi-dimensional comprehensive fault dataset of vibrating screens from different fields using multi-domain combined statistical analysis methods and principal component analysis methods includes the following steps:

[0021] Step A1: Standardize the fault data using a multi-domain combined statistical analysis method to eliminate the influence of dimensions of data from different domains, and extract the statistical characteristics of the vibrating screen fault data using a multi-domain combined method.

[0022] Step A2: Use principal component analysis to remove redundant features and generate a highly robust feature set, which includes the following sub-steps:

[0023] Step A21: Calculate the covariance matrix of the feature set:

[0024]

[0025] In the formula, Let b represent the original feature set obtained in step A1, and b represent the total number of original features. This represents the standardized original feature set. express Transpose of;

[0026] Step A22: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​λ and corresponding eigenvectors ν. Arrange the eigenvalues ​​λ in descending order and selectively retain the first k principal components, where k is determined based on the principle that the cumulative contribution rate is greater than 95%. Then project the original feature matrix onto the selected principal components to obtain a highly robust feature set.

[0027] Step A3: Divide the robust feature set to generate the source domain feature training set. Target domain feature training set and target domain feature test set Where S represents the source domain, T represents the target domain, and N... S N is the number of samples in the source domain feature training set. t N is the number of samples in the training set for the target domain features. tu N is the number of test samples in the target domain feature set. t +N tu =N T And N tu Greater than N t ; Let i be the i-th sample in the source domain feature training set. Let be the label of the i-th sample in the source domain feature training set; For the i-th sample in the target domain feature training set, The label of the i-th sample in the target domain feature training set; Let be the i-th sample in the target domain feature test set; both the source domain and the target domain contain C fault categories, and each fault sample contains m features.

[0028] Step B further includes:

[0029] Step B1: Map the source domain feature training set and the target domain feature training set to the random feature space to generate source domain random feature data and target domain random feature data;

[0030] Step B2: Using the random feature data of the source domain and the random feature data of the target domain, calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain based on the difference of maximum mean and covariance, so as to generate the joint distribution difference.

[0031] Step C further includes:

[0032] The source domain random feature data generated in step B is subjected to domain feature discrimination processing. By introducing a domain discrimination metric, the feature space is optimized to enhance the intra-class compactness and inter-class separability of features in the random feature space.

[0033] Further, in step D, a vibrating screen fault diagnosis model is constructed based on an enhanced domain adaptive incremental random vector function chain network to predict the fault category of unknown labels in the target domain; the vibrating screen fault diagnosis model is trained using the source domain feature training set and the target domain feature training set to obtain the output weights of the vibrating screen fault diagnosis model, and fault diagnosis is performed on the target domain feature test set; specifically, it includes the following sub-steps:

[0034] Step D1: In the initialization phase of constructing the vibrating screen fault diagnosis model, set the upper limit of the maximum number of hidden layer nodes to L. max The network currently has L hidden layer nodes, the initial error of the model is set to e0, the tolerance error is ε, and a set of hyperparameters is defined. And the random parameters are randomly assigned within the interval [-δ, δ].

[0035] Step D2: Calculate the maximum mean difference matrix M0, the maximum covariance difference matrix Q0, and the source domain intra-class discriminant metric matrix L. tra-S and the inter-class discriminant metric matrix L ter-S ;

[0036] Step D3: Within the range of random parameter allocation, generate L hidden layer nodes sequentially. For each node generated, calculate the corresponding source domain hidden layer node output H. S and the output H of the hidden layer node in the target domain t And calculate H, H = [H S H t ];

[0037] Step D4: Calculate the hidden layer output H in the source domain S The class mean in the feature space is used to obtain the class centroid matrix of the source domain.

[0038] Step D5: Calculate the joint distribution difference matrix M;

[0039] Step D6: Solve for the model's output weights using a globally optimal approach:

[0040]

[0041] In the formula, I represents the identity matrix. H represents S The transpose of Y S Y represents the true output of the source domain feature training set. t This represents the true output of the target domain feature training set;

[0042] Step D7: Calculate the trade-off factor in the joint distribution difference:

[0043]

[0044] In the formula, d mar For marginal distribution differences, d con For conditional distribution differences;

[0045] Step D8: Calculate the residual e L =H t β * -Y t When the residual satisfies the tolerance error ε or the number of hidden layer nodes reaches L max If the model is established correctly, proceed to step D9; otherwise, return to step D3 and add hidden layer nodes to the network, continuously looping from step D3 to step D6 until the residuals meet the tolerance error ε or the number of hidden layer nodes reaches L. max Stop at this time;

[0046] Step D9: Use the vibrating screen fault diagnosis model to diagnose the target domain feature test set samples, and calculate the diagnostic accuracy, precision, and F1 score.

[0047] Secondly, this invention discloses an enhanced domain-adaptive cross-domain fault diagnosis system for vibrating screens, the system comprising:

[0048] The dataset construction module is used to obtain multidimensional comprehensive fault datasets of vibrating screens in different fields and divide them into source domain feature training set, target domain feature training set and target domain feature test set.

[0049] The joint distribution difference processing module is used to map the source domain feature training set and the target domain feature training set to a random feature space for joint distribution difference processing in order to perform feature neighborhood and class alignment.

[0050] The domain feature discrimination processing module is used to perform domain feature discrimination processing on the source domain feature training set to enhance the intra-class tightness and inter-class separation of features;

[0051] The model building and training module is used to construct a cross-domain fault diagnosis model for vibrating screens based on the joint distribution difference processing results and the domain feature discrimination processing results, and to predict the fault category of unknown labels in the target domain.

[0052] The beneficial effects of this invention are as follows:

[0053] First, the enhanced domain adaptive cross-domain fault diagnosis method and system for vibrating screens of the present invention, through in-depth analysis of the operating mechanism of the vibrating screen and analysis of the vibration signals collected by the vibrating screen according to its characteristics, proposes a robust feature extraction processing method. At the same time, by applying the enhanced domain adaptation method to an incremental random construct neural network, cross-domain fault diagnosis is achieved, which effectively solves the domain offset problem caused by the differences in fault data distribution under different working conditions (such as changes in equipment models, environmental noise interference, and natural aging of equipment in actual industry).

[0054] Second, the enhanced domain-adaptive cross-domain fault diagnosis method and system of the present invention achieves domain-class alignment by mapping the features of the source domain and the target domain to a random feature space for joint distribution difference processing, and performs domain feature discrimination processing on the source domain to enhance intra-class compactness and inter-class separation, thereby achieving domain-adaptive capability and significantly improving the performance and generalization capability of the fault diagnosis model. Attached Figure Description

[0055] Figure 1 This is a flowchart of the enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to the present invention;

[0056] Figure 2 This is a schematic diagram of the data distribution differences of the present invention, wherein (a) is a group boundary distribution diagram of the source domain vibrating screen and (b) is a group boundary distribution diagram of the target domain vibrating screen;

[0057] Figure 3 This is a schematic diagram of the enhanced domain adaptive incremental random vector function chain network structure in this invention. Detailed Implementation

[0058] The following embodiments are provided to enable those skilled in the art to more fully understand the present invention, but do not limit the invention in any way.

[0059] like Figure 1 As shown, the present invention provides an enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens, comprising the following steps:

[0060] S1: Obtain vibration acceleration and temperature data of vibrating screens from different fields, construct a multi-dimensional fault dataset, and generate a comprehensive fault dataset by fusing fault information from different monitoring points; Step S1 is detailed as follows:

[0061] S11: The operating mechanism of a vibrating screen mainly relies on the periodic excitation force generated by a vibrating motor. This force causes the screen body to vibrate in a reciprocating or rotating manner, resulting in the stratification and jumping of materials on the screen surface. During the vibration process, smaller particles fall through the screen holes to the lower layer, while larger particles remain on the screen surface and are eventually discharged, achieving material grading and screening according to particle size. As shown in Table 1, in actual use, due to the harsh working environment, the main components of the vibrating screen are prone to failure.

[0062] S12: Based on common faults, acceleration-temperature composite sensors are installed at eight points: the main motor, drive shaft, vibrating screen motor-side exciter, vibrating screen non-motor-side exciter, the middle of the vibrating screen motor-side side plate, the middle of the vibrating screen motor-side non-side plate, the inlet end and outlet end of the vibrating screen non-motor side. The vibration signal acquisition frequency is 51200Hz, and the acquisition time for each vibration signal is 320ms. Fault data from different measuring points of different vibrating screens are transmitted to the database via a wireless gateway. This represents the original fault data collected from the source domain. This represents the original fault data collected from the target domain.

[0063] S2: Perform multi-dimensional robustness processing on the comprehensive fault dataset in the time domain, frequency domain, and time-frequency domain to generate a highly robust feature set, and divide it into a source domain feature training set, a target domain feature training set, and a target domain feature test set; Step S2 is as follows:

[0064] S21: Standardize the fault data to eliminate the influence of dimensions from different domains; extract feature sets of the original fault data from the source and target domains in the time, frequency, and time-frequency domains. The total number of features in both the source and target domains is m. As shown in Table 2, this feature set consists of the mean, root mean square value, frequency centroid, mean square frequency, wavelet transform coefficients, etc.

[0065] S22: Use principal component analysis to remove redundant features and generate a highly robust feature set. The specific steps are as follows:

[0066] Calculate the covariance matrix of the feature set:

[0067]

[0068] In the formula, Let b represent the original feature set of the source domain obtained in step S21, and let b represent the total number of original features in the source domain. This represents the standardized original feature set of the source domain. express Transpose of;

[0069] Calculate the eigenvalues ​​λ of the covariance matrix S and eigenvector ν S , eigenvalue λ S Sort in descending order, selectively retaining the first k. S There are principal components, where k S The feature set is determined based on the principle that the cumulative contribution rate is greater than 95%; then, the original feature set of the source domain is projected onto the selected principal components to obtain a highly robust feature set. The same method is used to obtain the target domain feature set.

[0070] S23: Partition the highly robust feature set to generate a set containing N S Source domain feature training set of 1 feature sample N t A training set of target domain features from a small number of feature samples. and N tu Target domain feature test set of feature samples Where N t +N tu =N T Both the source and target domains contain C fault categories.

[0071] Table 1. Vibrating Screen Fault Types

[0072]

[0073] Table 2. Characteristics of Vibrating Screen Fault Data

[0074]

[0075]

[0076] S3: Based on the source domain and target domain feature training sets, map them to a random feature space for joint distribution difference processing to achieve feature neighborhood and class alignment; step S3 is as follows:

[0077] like Figure 2 As shown, due to differences between cross-domain data, the data from different vibrating screens are distributed differently on different coordinate axes; therefore, the features of the source and target domains are aligned in the random feature space; the specific steps are as follows:

[0078] S31: Map the source domain and target domain feature training sets to a random feature space to generate multiple sets of random feature data;

[0079] S32: Using the random feature data of the source and target domains, calculate the marginal distribution difference and conditional distribution difference between the source and target domains based on the difference in maximum mean and covariance, to generate a joint distribution difference; the joint distribution difference with the classifier is as follows:

[0080] L JMC =tr(β) T HMH T β)

[0081] Let the joint distribution difference matrix in, It is the balance factor between the two differences; H = [H S H T ], H S H represents the output matrix of the hidden layer in the source domain. T This represents the output matrix of the hidden layer in the target domain;

[0082] S4: Perform domain feature discrimination processing on the source domain feature training set to enhance the intra-class tightness and inter-class separability of features; step S4 is as follows:

[0083] To further effectively address the domain offset problem and learn more domain-specific discriminative features, we enhance intra-class tightness and inter-class separability by minimizing intra-class loss and maximizing inter-class loss, thereby improving the accuracy of diagnostic results. The source domain intra-class discriminative metric is calculated as follows:

[0084]

[0085] In the formula, L tra-S L represents the discriminant metric matrix within the source domain. tra-S ; H representsS Transpose of;

[0086] The inter-class discriminant metric of the source domain is calculated as follows:

[0087]

[0088] In the formula, This represents the class centroid matrix of the source domain in the feature space. express The transpose of L ter-S Represents the inter-class discriminant metric matrix;

[0089] S5: Based on the joint distribution difference and domain feature discrimination processing, an enhanced domain adaptive method is generated to construct a cross-domain fault diagnosis model for the vibrating screen; step S5 is as follows:

[0090] like Figure 3 As shown, this invention constructs a vibrating screen fault diagnosis model based on enhanced domain adaptation, namely, an enhanced domain adaptive incremental random vector function chain network, used to predict fault categories with unknown labels in the target domain. During model training, the source domain feature set and a small amount of the target domain feature set are used to construct the training set, while the remaining large amount of the target domain feature set is used to construct the test set. Finally, the output weights of the model are obtained and used for fault diagnosis of the target domain test set. The specific steps are as follows:

[0091] S51: Obtain the feature training set and feature test set through S2, including the source domain feature training set. Target domain feature training set Target domain feature test set Perform one-hot encoding on the training sample labels; initialize various parameters for building the model: the upper limit of the maximum number of hidden layer nodes is L. max The network currently has L hidden layer nodes, the initial error of the model is set to e0, the tolerance error is ε, and a set of hyperparameters is defined. And the random parameters are randomly assigned within the interval [-v, δ];

[0092] S52: Calculate the maximum mean difference matrix M0, the maximum covariance difference matrix Q0, and the source domain intra-class discriminant metric matrix L. tra-S and the inter-class discriminant metric matrix L ter-S ;

[0093] S53: Within the range of random parameter allocation, L hidden layer nodes are randomly generated sequentially. For each node generated, the corresponding source domain hidden layer node output H is calculated. S and the output H of the hidden layer node in the target domain t And calculate H;

[0094] S54: Calculate the hidden layer output H in the source domain SThe class mean in the feature space is used to obtain the class centroid matrix of the source domain.

[0095] S55: Calculate the joint distribution difference matrix M;

[0096] S56: Solve for the model's output weights using a globally optimal approach:

[0097]

[0098] In the formula, I represents the identity matrix. H represents S The transpose of Y S Y represents the true output of the source domain training set. t This represents the true output of the training set in the target domain;

[0099] S57: Calculating the trade-off factor in the joint distribution difference:

[0100]

[0101] In the formula, d mar For marginal distribution differences, d con For conditional distribution differences;

[0102] S58: Calculate residual e L =H t β * -Y t When the residual satisfies the tolerance error ε or the number of hidden layer nodes reaches L max If the model is built successfully, the process ends; otherwise, return to step S53 and add hidden layer nodes to the network. Repeat steps S53-S56 until the residuals meet the tolerance error ε or the number of hidden layer nodes reaches L. max Stop at this time;

[0103] S59: Using network models to diagnose target domain test set samples The hidden layer output weight H is calculated using the random parameters generated in step S53. tu The output matrix T of the model is obtained. tu =H tu β * Transform the output matrix and compare it with the true label Y of the test sample. tu Compare and calculate accuracy, precision, and F1 score.

[0104] To illustrate the performance of this invention in cross-domain fault diagnosis, a practical example of a vibrating screen is used. Experiments were conducted on three vibrating screens with the same fault, and their specific fault descriptions are shown in Table 3.

[0105] The experiment selected four methods—RVFL, IRVFL, UD-RVFL, and DAMR-RWNN—as the comparative methods for fault diagnosis in this paper. RVFL and IRVFL are base models without domain adaptation, trained only using labeled source domain feature sets, and then immediately evaluated on the target domain test set; UD-RVFL and DAMR-RWNN are cross-domain models that consider domain offset. The method of this invention was compared with the above four methods in terms of accuracy, precision, and F1 score. The diagnostic accuracy results of different methods are shown in Table 4, and Tables 5 and 6 show the diagnostic precision and F1 score of the five methods. As can be seen from the tables, in the experiments of the six cross-domain tasks, the method of this invention achieves the best values ​​in accuracy, precision, and F1 score, and compared with RVFL and IRVFL, the performance of the method of this invention is significantly improved, indicating that the cross-domain model can better solve the domain offset problem, thereby improving the accuracy of fault diagnosis. By comparing with other methods, the method of this invention performs best in all three evaluation metrics, proving the effectiveness of the method.

[0106] Table 3. Selected faults in the experiment

[0107]

[0108] Table 4. Diagnostic accuracy of different methods

[0109]

[0110] Table 5. Diagnostic accuracy of different methods

[0111]

[0112] Table 6 F1 scores for different methods

[0113]

[0114] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A cross-domain fault diagnosis method for an enhanced domain-adaptive vibrating screen, characterized in that, The method includes the following steps: Step A: Divide the acquired multidimensional comprehensive fault dataset of vibrating screens from different fields into source domain feature training set, target domain feature training set, and target domain feature test set; Step B involves mapping the source domain feature training set and the target domain feature training set to a random feature space for joint distribution difference processing to perform feature neighborhood and class alignment. Step C involves performing domain feature discrimination processing on the source domain feature training set to enhance the intra-class tightness and inter-class separability of features; Step D: Based on the joint distribution difference processing results and the domain feature discrimination processing results, a cross-domain fault diagnosis model for vibrating screen is constructed based on the enhanced domain adaptive incremental random vector function chain network to predict the fault category of unknown labels in the target domain. In step D, a vibrating screen fault diagnosis model is constructed based on an enhanced domain adaptive incremental random vector function chain network to predict the fault category of unknown labels in the target domain. The vibrating screen fault diagnosis model is trained using the source domain feature training set and the target domain feature training set to obtain the output weights of the model, and fault diagnosis is performed on the target domain feature test set. Specifically, this includes the following sub-steps: Step D1: In the initialization phase of constructing the vibrating screen fault diagnosis model, the upper limit of the maximum number of hidden layer nodes is set to... The current number of hidden layer nodes in the network is The initial error value of the model is set to The tolerance error is Define a set of hyperparameters and random parameters in Randomly assigned within the interval; Step D2: Calculate the maximum mean difference matrix Maximum covariance difference matrix Source domain intra-class discriminant metric matrix and inter-class discriminant metric matrix ; Step D3: Randomly generate parameters sequentially within the allocated range. For each hidden layer node generated, the corresponding source domain hidden layer node output is calculated. and target domain hidden layer node output and calculation , ; Step D4: Calculate the hidden layer output of the source domain The class mean in the feature space is used to obtain the class centroid matrix of the source domain. ; Step D5: Calculate the joint distribution difference matrix ; Step D6: Solve for the model's output weights using a globally optimal approach: ； In the formula, Represents the identity matrix. express transpose, This represents the true output of the source domain feature training set. This represents the true output of the target domain feature training set; Step D7: Calculate the trade-off factor in the joint distribution difference: ； In the formula, For marginal distribution differences, For conditional distribution differences; Step D8: Calculate the residuals When the residuals satisfy the tolerance error Or the number of hidden layer nodes reaches If the model is established correctly, proceed to step D9; otherwise, return to step D3 and add hidden layer nodes to the network, continuously looping from step D3 to step D6 until the residuals meet the tolerance error. Or the number of hidden layer nodes reaches Stop at this time; Step D9: Use the vibrating screen fault diagnosis model to diagnose the target domain feature test set samples, and calculate the diagnostic accuracy, precision, and... Fraction.

2. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 1, characterized in that, Step A involves obtaining a multidimensional comprehensive fault dataset of vibrating screens from different fields, which includes the following steps: In-depth analysis of the operating mechanism of vibrating screens under harsh working environments and heavy loads is conducted to identify and predict common failure types. Based on common fault types of vibrating screens, acceleration-temperature composite sensors are installed at different key fault monitoring points of vibrating screens to collect comprehensive fault datasets of different monitoring points of vibrating screens in different fields, and generate multidimensional comprehensive fault datasets of vibrating screens in different fields.

3. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 2, characterized in that, The common fault types of the vibrating screen include at least the following: The faults include: screen box body faults such as side plate cracking, lifting beam cracking, drive beam cracking, load-bearing beam cracking, small beam cracking, T-shaped steel cracking, small beam loosening, and T-shaped steel loosening; vibrator faults such as bearing damage and gear damage; drive component faults such as bearing damage and bearing seat loosening; screen plate assembly faults such as screen plate loosening, side pressure plate loosening, feed and discharge blind plates loosening, and rail seat loosening; and spring cracking faults.

4. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 2, characterized in that, The key fault monitoring points of the vibrating screen include at least the main motor, drive shaft, vibrating screen motor-side exciter, vibrating screen non-motor-side exciter, the middle of the vibrating screen motor-side side plate, the middle of the vibrating screen motor-side non-side plate, the vibrating screen non-motor-side feed end and discharge end.

5. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 1, characterized in that, Step A further includes: Multi-domain combined statistical analysis and principal component analysis methods are used to perform multi-dimensional robustness processing on the multi-dimensional comprehensive fault dataset of vibrating screens in different fields in the time domain, frequency domain, and time-frequency domain. The generated highly robust feature set is divided into source domain feature training set, target domain feature training set, and target domain feature test set.

6. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 5, characterized in that, The process of performing multi-dimensional robustness processing on multi-dimensional comprehensive fault datasets of vibrating screens from different fields using multi-domain combined statistical analysis methods and principal component analysis methods includes the following steps: Step A1: Standardize the fault data using a multi-domain combined statistical analysis method to eliminate the influence of dimensions of data from different domains, and extract the statistical characteristics of the vibrating screen fault data using a multi-domain combined method. Step A2: Use principal component analysis to remove redundant features and generate a highly robust feature set, which includes the following sub-steps: Step A21: Calculate the covariance matrix of the feature set: ； In the formula, This represents the original feature set obtained in step A1. This represents the total number of original features. This represents the standardized original feature set. express Transpose of; Step A22: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues. and the corresponding feature vector , eigenvalue Sort in descending order, selectively retaining the top few. One principal component, of which It is determined based on the principle that the cumulative contribution rate is greater than 95%; then the original feature matrix is ​​projected onto the selected principal components to obtain a highly robust feature set; Step A3: Divide the robust feature set to generate the source domain feature training set. Target domain feature training set and target domain feature test set ,in Represents the source domain. Indicates the target domain. It is the number of samples in the source domain feature training set. It is the number of samples in the training set for the target domain features. It is the number of test samples in the target domain feature set. ,and Greater than ; The first of the source domain feature training set One sample, The first of the source domain feature training set The labels of each sample; The first of the target domain feature training set One sample, The first of the target domain feature training set The labels of each sample; For the target domain feature test set, the first One sample; both the source and target domains contain... Each fault sample contains one fault category. One characteristic.

7. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 1, characterized in that, Step B further includes: Step B1: Map the source domain feature training set and the target domain feature training set to the random feature space to generate source domain random feature data and target domain random feature data; Step B2: Using the random feature data of the source domain and the random feature data of the target domain, calculate the marginal distribution difference and conditional distribution difference between the source domain and the target domain based on the difference of maximum mean and covariance, so as to generate the joint distribution difference.

8. The enhanced domain adaptive cross-domain fault diagnosis method for vibrating screens according to claim 1, characterized in that, Step C further includes: The source domain random feature data generated in step B is subjected to domain feature discrimination processing. By introducing a domain discrimination metric, the feature space is optimized to enhance the intra-class compactness and inter-class separability of features in the random feature space.

9. A cross-domain fault diagnosis system for a vibrating screen with enhanced domain adaptation based on the method described in any one of claims 1-8, characterized in that, The system includes: The dataset construction module is used to obtain multidimensional comprehensive fault datasets of vibrating screens in different fields and divide them into source domain feature training set, target domain feature training set and target domain feature test set. The joint distribution difference processing module is used to map the source domain feature training set and the target domain feature training set to a random feature space for joint distribution difference processing in order to perform feature neighborhood and class alignment. The domain feature discrimination processing module is used to perform domain feature discrimination processing on the source domain feature training set to enhance the intra-class tightness and inter-class separation of features; The model building and training module is used to construct a cross-domain fault diagnosis model for vibrating screens based on the joint distribution difference processing results and the domain feature discrimination processing results, and to predict the fault category of unknown labels in the target domain.

Images