Industrial process fault diagnosis method and system based on multi-sensing source feature selection

Through the multi-view feature selection method, combined with time domain, frequency domain, time frequency domain analysis and deep learning, feature extraction and recognition are optimized, and the troubleshooting difficulties of multi-sensor source data in industrial processes are solved, achieving more accurate and efficient Fault identification.

CN119002458AActive Publication Date: 2024-11-22LANZHOU UNIVERSITY OF TECHNOLOGY

Patent Information

Application Number
CN202411100515.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-22
Estimated Expiration
2044-08-12

AI Technical Summary

Technical Problem

The prior art has difficulty in effectively mining and feature selection from massive data from multi-sensor sources in industrial processes, resulting in inconsistent fault diagnosis effects, and traditional single-view feature selection methods are difficult to capture data patterns of different faults.

Method used

Using multi-view feature selection method, through time domain, frequency domain, time frequency domain analysis and deep learning feature extraction, data is mapped to the public information measurement space, the weight matrix of multi-view features is calculated, and the fault pattern recognizer and category boundaries are combined. Extend method, optimize the feature set to identify significantly different fault features.

Benefits of technology

It significantly improves the classification accuracy and efficiency of fault diagnosis, reduces feature dimensions, reduces information redundancy, enhances classification capabilities, and realizes accurate identification of industrial equipment failures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119002458A_ABST
    Figure CN119002458A_ABST
Patent Text Reader

Abstract

The invention discloses an industrial process fault diagnosis method and system based on multi-sensing source feature selection. The method comprises the following steps: S1, obtaining process data collected by each sensor; s2, performing feature extraction on the data acquired by each sensor under different feature views, and uniformly mapping the data to a public information measurement space to obtain a data feature set under each feature view; s3, in a public information measurement space, integrating views from multiple angles, analyzing features, identifying fault features with significant differences, and calculating and obtaining a weight matrix of the features of the multiple views to optimize a feature set; s4, a fault mode recognizer is introduced to be combined with the optimized fault features to obtain prediction categories corresponding to the fault features, and the prediction categories are converted into probability distribution to recognize and diagnose faults of the industrial equipment; according to the method, a multi-view feature selection and category boundary expansion method is introduced, so that the feature dimension is effectively reduced, the information redundancy is reduced, and meanwhile, the classification result accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of sensor diagnosis technology, and more particularly to an industrial process fault diagnosis method and system based on multi-sensor source feature selection. Background Art

[0002] Fault diagnosis is a key technology in intelligent manufacturing and has been widely studied and applied in industrial applications. Sudden failures in industrial processes may cause equipment damage and a significant decrease in production efficiency, while accurate fault diagnosis can significantly reduce maintenance costs and ensure the reliable operation of mechanical equipment. Therefore, using monitoring sensors to obtain various sensory information such as vibration, sound, and temperature to timely diagnose industrial equipment faults and take corresponding maintenance measures has important practical significance for ensuring its safe, reliable, and efficient operation.

[0003] At present, with the rapid development of modern sensor technology, there are more and more types of sensors, which provides a guarantee for improving the reliability and efficiency of equipment fault diagnosis.

[0004] However, in practical applications, it is difficult to mine sufficient and valuable fault information from the massive monitoring data of sensors, making it difficult to effectively separate faults. In addition, many high-dimensional original features are complex to extract, and the sample distribution is relatively sparse, which brings challenges to accurately establishing fault models. In the face of complex data types collected by sensors, traditional single-view feature selection algorithms have shown certain limitations. The reason is that the features extracted by traditional methods are too single, and this feature selection method is difficult to fully capture the data patterns of different faults. For example, data from different sensor sources contain different degrees of fault information, and the information volume of the same sensor data under different fault modes is quite different. Therefore, it is difficult to effectively select features for multi-sensor source information at the data level, resulting in inconsistent diagnostic results.

[0005] Therefore, how to perform feature selection based on multi-sensor source information to improve the reliability and efficiency of industrial process fault diagnosis is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0006] In view of this, the present invention provides an industrial process fault diagnosis method and system based on multi-sensor source feature selection to solve some of the technical problems mentioned in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] An industrial process fault diagnosis method based on multi-sensor source feature selection comprises the following steps:

[0009] S1. Obtain process data collected by each sensor;

[0010] S2. Extract features from the data collected by each sensor under different feature views and uniformly map them to a common information metric space to obtain a data feature set under each feature view;

[0011] S3. Integrate views from multiple angles in the common information metric space, analyze features and identify fault features with significant differences, and calculate and obtain the weight matrix of multi-view features to optimize the feature set;

[0012] S4. Introduce a fault pattern identifier combined with the optimized fault features to obtain the prediction category corresponding to the fault features and convert it into a probability distribution to identify and diagnose faults of industrial equipment.

[0013] Preferably, for step S1, each sample x i It is composed of data obtained by M sensors, represented by x i =[x i1 ,x i2 ,…,x iM ].

[0014] Preferably, for step S2, each sample is subjected to feature extraction in different feature views by using time domain, frequency domain, time-frequency domain analysis and deep learning feature extraction methods, and mapped to the common information metric space to generate corresponding feature representations as For the vth view, we have Among them, D v is the feature dimension of the vth view, N is the number of samples, and the total feature matrix of the dataset is expressed as X = [X 1 ,X 2 ,...,X V ], belongs to D×N dimensional space, where Represents the feature dimension of all views.

[0015] Preferably, the specific contents of step S3 include:

[0016] An independent weight matrix is ​​set for each view to measure the importance of the features in each view for various faults. The importance of the features of each view is determined by the corresponding weight matrix, which together constitute a comprehensive feature matrix.

[0017] Construct a difference metric loss function based on the parameters to be learned, combine regularization parameters to balance the difference and importance of feature views, and optimize the weight matrix by minimizing the loss function objective;

[0018] A sample difference metric loss function is established, a weight matrix is ​​used to map data from feature space to label space, and a category boundary extension method is adopted to introduce a forward adjustment matrix and a guidance matrix to optimize the label space.

[0019] Preferably, the difference metric loss function based on the parameters to be learned is specifically:

[0020] L D (W) = λR1 + βR2

[0021] Among them, λ and β are regularization parameters used to balance the difference and importance between feature views, R1 is the difference between feature views, and R2 is the importance between feature views.

[0022] Preferably, the sample difference metric loss function is specifically:

[0023]

[0024] Where L is the label matrix, which is obtained by one-hot encoding the fault category, and O is the guidance matrix, O∈{-1,+1} N×C , H is the positive adjustment matrix, L+O⊙H is the label matrix after the introduction of category boundary expansion, and ⊙ is the Hadamard operator.

[0025] Preferably, the prediction category in step S4 is:

[0026]

[0027] in, is the prediction matrix, is the sample to be predicted, W is the feature weight matrix, H is the adjustment matrix;

[0028] The probability distribution of the corresponding conversion of the predicted category is:

[0029]

[0030] in, for C scalars in the matrix;

[0031] Fault category It is expressed as:

[0032]

[0033] Among them, c (i) It indicates that the column with the largest category is i, and argmax() indicates the category with the largest predicted probability.

[0034] An industrial process fault diagnosis system based on multi-sensor source feature selection, based on the industrial process fault diagnosis method based on multi-sensor source feature selection, including a data acquisition module and a multi-sensor source feature selector model;

[0035] The multi-sensor source feature selector model includes multiple parallel multi-view feature extraction modules, a common feature space, a sensor source feature selector and a fault identifier;

[0036] A data acquisition module, used for acquiring process data collected by each sensor and inputting it into a multi-sensor source feature selector model;

[0037] Multiple parallel multi-view feature extraction modules are used to extract features from the data collected by each sensor under different feature views through time domain, frequency domain, time-frequency domain analysis and deep learning feature extraction methods, and uniformly map them to a common information metric space to obtain the data feature set under each feature view;

[0038] The sensor source feature selector is used to integrate views from multiple angles in the common information metric space, analyze features and identify fault features with significant differences, and calculate and obtain the weight matrix of multi-view features to optimize the feature set;

[0039] The fault identifier is used to combine the optimized fault features, obtain the prediction category corresponding to the fault features, and convert it into a probability distribution to identify and diagnose the faults of industrial equipment.

[0040] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements an industrial process fault diagnosis method based on multi-sensor source feature selection.

[0041] A processing terminal comprises a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor executes the computer program, an industrial process fault diagnosis method based on multi-sensor source feature selection is implemented.

[0042] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses an industrial process fault diagnosis method and system based on multi-sensor source feature selection, in order to realize the industrial process fault diagnosis under the condition of insufficient key fault information mining and feature learning mechanisms of different sensor sources in the field of intelligent diagnosis of industrial equipment, a variety of feature extraction methods of time domain, frequency domain, time-frequency domain and deep learning are comprehensively utilized, and they are integrated to form a new feature vector. This multi-angle feature extraction strategy is conducive to capturing richer information and enhancing feature diversity, thereby effectively improving classification performance;

[0043] The present invention introduces a multi-view feature selection method to construct a multi-sensor source feature selector, fully mines key fault information, and especially adopts a category boundary expansion method to significantly improve the accuracy of the algorithm in classification results. Experimental results confirm that the present invention effectively reduces feature dimensions and information redundancy, while enhancing classification capabilities and realizing multi-industrial process fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0045] Figure 1 A schematic diagram of an industrial process fault diagnosis method based on multi-sensor source feature selection provided by the present invention;

[0046] Figure 2 A schematic diagram of a label matrix and a guidance matrix list corresponding to different fault categories provided by the present invention;

[0047] Figure 3 A schematic diagram of the confusion matrix of the diagnosis results of the Random Forest method provided by the present invention;

[0048] Figure 4 A schematic diagram of the confusion matrix of the diagnosis results of the K-NN method provided by the present invention;

[0049] Figure 5 A schematic diagram of a confusion matrix of the diagnosis results of the method provided by the present invention;

[0050] Figure 6 A schematic diagram of the classification accuracy of different algorithm models provided by the present invention;

[0051] Figure 7 A schematic diagram of classification accuracy of different parameters provided by the present invention;

[0052] Figure 8 A schematic diagram of predicted values ​​of fault 0 in each view when λ=0 provided by the present invention;

[0053] Fig. 9 A schematic diagram of predicted values ​​of fault 0 in each view when λ=0.1 provided by the present invention;

[0054] Fig.10 A schematic diagram of classification without introducing the adjustment matrix H provided by the present invention;

[0055] Fig.11A classification diagram of the introduction of the adjustment matrix H provided by the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The embodiment of the present invention discloses an industrial process fault diagnosis method based on multi-sensor source feature selection, comprising the following steps:

[0058] S1. Obtain process data collected by each sensor;

[0059] S2. Extract features from the data collected by each sensor under different feature views and uniformly map them to a common information metric space to obtain a data feature set under each feature view;

[0060] S3. Integrate views from multiple angles in the common information metric space, analyze features and identify fault features with significant differences, and calculate and obtain the weight matrix of multi-view features to optimize the feature set;

[0061] S4. Introduce a fault pattern identifier combined with the optimized fault features to obtain the prediction category corresponding to the fault features and convert it into a probability distribution to identify and diagnose faults of industrial equipment.

[0062] In order to further implement the above technical solution, for step S1, each sample x i It is composed of data obtained by M sensors, represented by x i =[x i1 ,x i2 ,…,x iM ].

[0063] In order to further implement the above technical solution, for step S2, each sample is subjected to feature extraction in different feature views through time domain, frequency domain, time-frequency domain analysis and deep learning feature extraction methods, and mapped to the common information metric space to generate the corresponding feature representation as For the vth view, we have Among them, D v is the feature dimension of the vth view, N is the number of samples, and the total feature matrix of the dataset is expressed as X = [X 1 ,X 2 ,...,X V ], belongs to D×N dimensional space, where Represents the feature dimension of all views.

[0064] In this embodiment, time domain analysis is used to extract statistical parameter characteristics of the signal, such as mean and variance, which provides a basis for timely fault detection; frequency domain analysis extracts signal frequency characteristics, such as power spectral density, through Fourier transform, which helps to identify periodic faults; time-frequency domain analysis combines the advantages of the first two and uses wavelet packet transform to refine time-frequency characteristics, which is suitable for non-stationary fault analysis; deep learning feature extraction extracts deep features of the signal through convolutional neural network learning, and uses Adam optimizer for multiple iterative training to optimize model performance; in the multi-view feature space, it can fully reflect the information obtained by each sample from multiple sensors for subsequent analysis and processing. The multi-dimensional feature extraction strategy not only enhances the perception of the equipment fault state, but also improves the comprehensiveness and accuracy of the fault diagnosis process.

[0065] In order to further implement the above technical solution, the specific contents of step S3 include:

[0066] Set a separate weight matrix for each view It is used to measure the importance of features in each view for various types of faults. The feature importance of each view is determined by the corresponding weight matrix, which together constitutes a comprehensive feature matrix

[0067] Construct a difference metric loss function L based on the parameters to be learned D (W), combines the regularization parameter to balance the difference and importance of feature views, and optimizes the weight matrix by minimizing the loss function objective;

[0068] A sample difference metric loss function is established, a weight matrix is ​​used to map data from feature space to label space, and a category boundary extension method is adopted to introduce a forward adjustment matrix and a guidance matrix to optimize the label space.

[0069] In order to further implement the above technical solution, the difference metric loss function based on the parameters to be learned is specifically:

[0070] L D (W) = λR1 + βR2

[0071] Among them, λ and β are regularization parameters used to balance the difference and importance between feature views, R1 is the difference between feature views, and R2 is the importance between feature views.

[0072] In this embodiment, R1 is used to measure the difference between feature views, with the purpose of fully considering the difference information between different views and making the prediction structures of multiple views consistent, thereby ensuring consistent evaluation results between views, specifically:

[0073]

[0074] Among them, E v It is used to measure the degree of difference in feature space, and its expression is:

[0075]

[0076] in, is the average of the prediction results of each view to obtain a more comprehensive prediction mean, X v and X l represents the feature matrix, W v and W l is the corresponding weight matrix, so E v The purpose of is to adjust the weight matrix to reduce the difference between feature views, thereby reducing the redundant feature differences while retaining feature consistency;

[0077] R2 aims to combine two regularization strategies to maintain local structure and strengthen global consensus during feature selection, thereby learning the complementarity between views in multi-view tasks. Its expression is:

[0078]

[0079] Among them, the first is the accumulation of the weight of each view, W v Represents the feature weight corresponding to each view, N v is a diagonal matrix whose diagonal elements measure the importance of each view. This term can promote the uniform distribution of the weights of the model on different views by accumulating the weights of each view, which helps to avoid overfitting and maintain the integrity of the local structure. The second term Tr(W T MW) emphasizes the structural characteristics of the model weight matrix through the trace function, aiming to enhance the consensus between different feature views to maintain the global structure. Among them, M is also a diagonal matrix, which represents the relationship between features. The purpose is to enhance the consensus between different feature views in the model, that is, to strengthen those features that are generally important in all views; N v and M are diagonal matrices derived from W. The introduction of the regularization parameter α provides a trade-off between the two regularization strategies. A smaller value of α enhances the first term. The larger α value strengthens the second term Tr(W T MW).

[0080] In order to further implement the above technical solution, the sample difference measurement loss function is specifically:

[0081]

[0082] Where L is the label matrix, which is obtained by one-hot encoding the fault category, and O is the guidance matrix, O∈{-1,+1} N×C , H is the positive adjustment matrix, L+O⊙H is the label matrix after the introduction of category boundary expansion, and ⊙ is the Hadamard operator.

[0083] In this embodiment, H includes a series of stretching parameters ε ij , and O determines the direction of these parameters. Specifically: First, in order to create a larger gap between the positive and non-positive classes, the fault categories are encoded using one-hot encoding to obtain the label matrix L; then the label matrix Du is used to adjust the guidance matrix O. When an element L of the label matrix L ij Indicates the positive time, the corresponding adjustment matrix O element O ij is set to +1, otherwise, if L ij If it is not positive, then O ij is set to -1, and the label matrix and guidance matrix lists corresponding to different fault categories are as follows Figure 2 ,Right now:

[0084]

[0085] The label matrix after introducing the category boundary expansion is L+O⊙H, the operator ⊙ is the operator Hadamard, and the final objective function L M (W).

[0086] In order to further implement the above technical solution, step S4 introduces a fault pattern identifier, combined with the optimized features in the multi-source sensor data obtained by the weight matrix, to accurately identify and diagnose the faults of industrial equipment, wherein the fault identifier converts the output of the linear model into a probability distribution by applying the function Softmax to the output of the linear model for multi-class classification problems. The feature weight matrix W and the adjustment matrix H can be obtained by learning the sensor source feature selector. If is the sample to be predicted, Calculate the prediction matrix based on the prediction category formula As the input of the Soffmax model, the predicted category of the sample is obtained. Finally, the predicted categories of N samples are obtained;

[0087] The predicted category in step S4 is:

[0088]

[0089] in, is the prediction matrix, is the sample to be predicted, W is the feature weight matrix, H is the adjustment matrix;

[0090] The probability distribution of the corresponding conversion of the predicted category is:

[0091]

[0092] in, for C scalars in the matrix, o1, ..., o c is the probability distribution corresponding to the C scalars;

[0093] Fault category It is expressed as:

[0094]

[0095] Among them, c (i) It indicates that the column with the largest category is i, and argmax() indicates the category with the largest predicted probability.

[0096] In another embodiment, a nickel flash furnace system is used as an example for simulation. Flash smelting is to mix the concentrate with 200°C preheated air or oxygen-enriched air in a certain proportion in a concentrate nozzle and then spray it vertically into a reaction tower. The concentrate is in a suspended state and is heated by the high-temperature furnace. It reacts with the reaction air within 2 to 3 seconds, quickly reaches the required temperature and melts rapidly, and the molten material falls into a sedimentation tank for sedimentation and separation. In order to fully melt the nickel, a large number of various sensors are distributed on the flash furnace body to constitute a variable measurement system required to control the production process. This embodiment selects 16 representative sensors for simulation verification. These sensors are used to measure key parameters such as temperature, pressure, sound and vibration.

[0097] 200 samples were collected for each flash furnace condition, each sample contained 16 channels of multi-sensor data, and each channel contained 2560 data points. The feature sets of each view were extracted from the fault data, and the feature sets of the filtered flash furnace data were randomly divided into 80% samples for training and 20% samples for model testing.

[0098] Compared with random forest, the feature selection method of the present invention has an advantage in efficiency, mainly because it does not need to build multiple decision trees and the parameter optimization time is short; in addition, the present invention has a lower dependence on internal parameters, compared with the K-NN method, the latter requires more pre-set internal parameters; Figures 3 to 5The confusion matrices of the flash furnace fault diagnosis results based on the RandomForest method, the K-NN method and the three different classification algorithms of the present invention are shown. These matrices clearly show the correspondence between the true labels and the predicted labels. The total number of test samples is 400, and each fault type contains 40 samples. The classification accuracy is obtained by calculating the ratio of the sum of the diagonal elements of the confusion matrix to the total number of samples, thereby reflecting the performance of the fault diagnosis classifier.

[0099] like Figure 6 , showing that the method proposed in the present invention effectively improves the accuracy of fault classification compared with other methods in the field of multi-view feature selection: the method of the present invention shows better performance in comparison with the RFS algorithm, which verifies the effectiveness of comprehensively considering complementary information and consensus information in multi-view feature selection; in contrast, although the MVML method considers the complementarity between views, it ignores the consistency, which may cause conflicts between views in the feature selection process, thereby reducing the accuracy; the MFSICC method considers both the complementarity and consistency of information, and performs better than other multi-view feature selection methods in fault diagnosis scenarios, but does not fully consider the distribution of different categories of data.

[0100] The feature view difference metric loss function proposed in the present invention has two key parameters, λ and β, which are used to control the consistency and complementarity of information. In the field of machine learning, the regularization term is used to adjust the fitting degree of the model to prevent overfitting, thereby limiting the hypothesis space of the model within the range of the defined consistency and complementarity regularization terms. If the definitions of these regularization terms meet actual needs, the model will be more sensitive to them. Appropriate adjustment of these parameters can improve the performance of the model. Therefore, a common method to quantify the advantages of the regularized model is to check the sensitivity of the parameters.

[0101] Specifically, by setting different values ​​of λ and β, such as {0.001, 0.01, 0.1, 1, 10, 100, 1000} and {0.01, 0.1, 1, 10, 100, 1000, 10000} respectively, the obtained classification results are different, e.g. Figure 7As shown, it can be clearly seen that the method of the present invention is very sensitive to the settings of these parameters, which means that information complementarity and consistency have an important influence on the performance of feature selection in fault diagnosis; when the parameters (λ, β) are (0.1, 0.1), the method proposed in this paper obtains the best classification accuracy. In addition, for parameter β, when β is approximately 1, the classification accuracy is relatively unstable with the change of λ; and when λ is approximately 100, the change of classification accuracy is also relatively unstable with the change of , which indicates that by adjusting the parameters (λ, β) to control the classification accuracy, since λ controls the consistency between views and β controls the complementarity between views, the feature view difference metric loss function is more sensitive to β, which means that in the feature selection of fault diagnosis, the complementary information between views plays a more important role.

[0102] To verify the impact of consistency on the output results, taking fault 0 as an example, some samples are selected and different λ parameter values ​​are set to visualize the prediction results of the four extracted views. Figure 8 and Fig. 9 , is the predicted value of fault 0 in each view: when λ is set to 0.1, the impact of consistency on the result is considered, while when λ is set to 0, the impact of consistency on the result is not considered. Each line in the figure represents the prediction of each view of a sample, and the color represents the predicted value. Figure 8 It can be observed that the predictions of the four views have obvious differences in each category, especially the Figure 4 The difference in predictions produced is most obvious, however, compared to Figure 8 , when considering the effect of consistency on the results, Fig. 9 It shows that the differences in the predicted outputs of the four views are significantly reduced.

[0103] In order to more intuitively demonstrate the impact of the optimization algorithm used in the present invention on the classification performance, the t-distributed stochastic neighbor embedding (t-SNE) algorithm is used to perform two-dimensional visualization of the fault classification results. Fig.10 is the classification diagram without introducing the adjustment matrix H, Fig.11In order to introduce the classification diagram of the adjustment matrix H, 0-9 represents ten types of faults. The predicted values ​​of each category are distinguished by different colors. The data points of the same color are concentrated in the same area, and the distance between the data points of different colors is large and not mixed. This further verifies the significant effect and accuracy of this method in flash furnace fault classification. From the comparison of the two figures, it can be seen that the fault category difference measurement loss function has a significant impact on the classification results. When the "category boundary expansion" method is not introduced, there is overlap between different categories. On the contrary, the classification effect is significantly improved; the reason is that the introduction of this method can adjust the regression target gap between different categories in the multi-classification regression problem. By appropriately increasing the gap between the targets, the accuracy of multi-classification regression can be improved. By adjusting the weights of different categories in the process of solving the objective function, the prediction is more accurate and better adapted to the distribution characteristics of the data.

[0104] An industrial process fault diagnosis system based on multi-sensor source feature selection, based on an industrial process fault diagnosis method based on multi-sensor source feature selection, including a data acquisition module and a multi-sensor source feature selector model;

[0105] The multi-sensor source feature selector model includes multiple parallel multi-view feature extraction modules, a common feature space, a sensor source feature selector and a fault identifier;

[0106] A data acquisition module, used for acquiring process data collected by each sensor and inputting it into a multi-sensor source feature selector model;

[0107] Multiple parallel multi-view feature extraction modules are used to extract features from the data collected by each sensor under different feature views through time domain, frequency domain, time-frequency domain analysis and deep learning feature extraction methods, and uniformly map them to a common information metric space to obtain the data feature set under each feature view;

[0108] The sensor source feature selector is used to integrate views from multiple angles in the common information metric space, analyze features and identify fault features with significant differences, and calculate and obtain the weight matrix of multi-view features to optimize the feature set;

[0109] The fault identifier is used to combine the optimized fault features, obtain the prediction category corresponding to the fault features, and convert it into a probability distribution to identify and diagnose the faults of industrial equipment.

[0110] Each of the multiple parallel multi-view feature extraction modules includes a time domain feature extraction unit, a frequency domain feature extraction unit, a time-frequency domain analysis feature extraction unit, and a deep learning feature extraction unit; the time domain feature extraction unit is used to extract the statistical parameter characteristics of the signal, such as the mean and variance, which provides a basis for timely fault detection; the frequency domain feature extraction unit extracts the signal frequency characteristics, such as power spectrum density, through Fourier transform, which helps to identify periodic faults; the time-frequency domain analysis feature extraction unit uses wavelet packet transform to refine the time-frequency features, which is suitable for non-stationary fault analysis; the deep learning feature extraction unit extracts the deep features of the signal through convolutional neural network learning, and uses the Adam optimizer for multiple iterative training to optimize the model performance.

[0111] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements an industrial process fault diagnosis method based on multi-sensor source feature selection.

[0112] A processing terminal comprises a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor executes the computer program, an industrial process fault diagnosis method based on multi-sensor source feature selection is implemented.

[0113] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0114] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An industrial process fault diagnosis method based on multi-sensor source feature selection, characterized in that: The following steps are involved: S1. Obtain process data collected by each sensor; S2. Extract features from the data collected by each sensor under different feature views and uniformly map them to a common information metric space to obtain a data feature set under each feature view; S3. Integrate views from multiple angles in the common information metric space, analyze features and identify fault features with significant differences, and calculate and obtain the weight matrix of multi-view features to optimize the feature set; S4. Introduce a fault pattern identifier combined with the optimized fault features to obtain the prediction category corresponding to the fault features and convert it into a probability distribution to identify and diagnose faults of industrial equipment.

2. The industrial process fault diagnosis method based on multi-sensor source feature selection according to claim 1 is characterized in that: For step S1, each sample x i It is composed of data obtained by M sensors, represented by x i =[x i1 ,x i2 ,…,x iM ].

3. The industrial process fault diagnosis method based on multi-sensor source feature selection according to claim 2 is characterized in that: For step S2, each sample is extracted from different feature views using time domain, frequency domain, time-frequency domain analysis, and deep learning feature extraction methods, and mapped to the common information metric space to generate the corresponding feature representation: For the vth view, we have Among them, D v is the feature dimension of the vth view, N is the number of samples, and the total feature matrix of the dataset is expressed as X = [X 1 ,X 2 ,...,X V ], belongs to D×N dimensional space, where Represents the feature dimension of all views.

4. The industrial process fault diagnosis method based on multi-sensor source feature selection according to claim 1 is characterized in that: The specific contents of step S3 include: An independent weight matrix is ​​set for each view to measure the importance of the features in each view for various faults. The importance of the features of each view is determined by the corresponding weight matrix, which together constitute a comprehensive feature matrix. Construct a difference metric loss function based on the parameters to be learned, combine regularization parameters to balance the difference and importance of feature views, and optimize the weight matrix by minimizing the loss function objective; A sample difference metric loss function is established, a weight matrix is ​​used to map data from feature space to label space, and a category boundary extension method is adopted to introduce a forward adjustment matrix and a guidance matrix to optimize the label space.

5. The industrial process fault diagnosis method based on multi-sensor source feature selection according to claim 4 is characterized in that: The difference metric loss function based on the parameters to be learned is specifically: L D (W)=λR1+βR2 Among them, λ and β are regularization parameters used to balance the difference and importance between feature views, R1 is the difference between feature views, and R2 is the importance between feature views.

6. The industrial process fault diagnosis method based on multi-sensor source feature selection according to claim 4 is characterized in that: The sample difference metric loss function is specifically: Where L is the label matrix, which is obtained by one-hot encoding the fault category, and O is the guidance matrix, O∈{-1,+1} N×C , H is the positive adjustment matrix, L+O⊙H is the label matrix after the introduction of category boundary expansion, and ⊙ is the Hadamard operator.

7. The industrial process fault diagnosis method based on multi-sensor source feature selection according to claim 1 is characterized in that: The predicted category in step S4 is: in, is the prediction matrix, is the sample to be predicted, W is the feature weight matrix, H is the adjustment matrix; The probability distribution of the corresponding conversion of the predicted category is: in, for C scalars in the matrix; Fault category It is expressed as: Among them, c (i) It indicates that the column with the largest category is i, and argmax() indicates the category with the largest predicted probability.

8. An industrial process fault diagnosis system based on multi-sensor source feature selection, characterized in that: An industrial process fault diagnosis method based on multi-sensor source feature selection according to any one of claims 1 to 7, comprising a data acquisition module and a multi-sensor source feature selector model; The multi-sensor source feature selector model includes multiple parallel multi-view feature extraction modules, a common feature space, a sensor source feature selector and a fault identifier; A data acquisition module, used for acquiring process data collected by each sensor and inputting it into a multi-sensor source feature selector model; Multiple parallel multi-view feature extraction modules are used to extract features from the data collected by each sensor under different feature views through time domain, frequency domain, time-frequency domain analysis and deep learning feature extraction methods, and uniformly map them to a common information metric space to obtain the data feature set under each feature view; The sensor source feature selector is used to integrate views from multiple angles in the common information metric space, analyze features and identify fault features with significant differences, and calculate and obtain the weight matrix of multi-view features to optimize the feature set; The fault identifier is used to combine the optimized fault features, obtain the prediction category corresponding to the fault features, and convert it into a probability distribution to identify and diagnose the faults of industrial equipment.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for diagnosing industrial process faults based on multi-sensor source feature selection as described in any one of claims 1 to 7 is implemented.

10. A processing terminal, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the computer program, the industrial process fault diagnosis method based on multi-sensor source feature selection as described in any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Bearing health monitoring method based on fault feature fusion

    CN114509266A

  • Information fusion three-phase motor fault diagnosis method based on game mapping learning

    CN114970744A

  • Rotating machinery variable working condition fault diagnosis method and system

    CN115165366A

  • Bearing fault diagnosis method based on federal local transfer learning

    CN116992336A

  • Fault diagnosis method for mechanical equipment

    CN117648611A

Cited By

  • Fault diagnosis method and system for fault large-scale unit based on sharing and specific representation learning

    CN119441893A

  • Industrial instrument fault detection and self-calibration method based on intelligent sensing fusion

    CN120445284A

  • An industrial instrument fault detection and self-calibration method based on intelligent sensing fusion

    CN120445284B