Data-driven fault diagnosis system and method based on texture feature fusion

Abstract

The invention discloses a data-driven fault diagnosis system and method based on texture feature fusion, and the method comprises the steps: obtaining easily-measured variable data and a corresponding fault type label in a monitored process, and forming a training data set; performing feature extraction on the training data set easy-to-measure variable samples to obtain training data initial features; performing texture information analysis on the extracted initial features to obtain training data texture features; fusing the initial features and the texture features of the training data, and determining a feature fusion scale of the training data; establishing full connection between the training data features and the corresponding sample fault type labels; data texture information is fused, deep features of training data are further mined, separable recognition of data in overlapping distribution or overlapping distribution groups is achieved, and the problem of fault type misclassification is solved.

Description

Technical Field

[0001] The invention relates to a data-driven fault diagnosis system and method based on texture feature fusion. Background Art

[0002] Process industry fault diagnosis technology can quickly identify fault types and ensure safe and reliable operation of production processes in industrial process control. Data-driven fault diagnosis technology mines hidden feature information of easily measurable data to establish a nonlinear mapping relationship between data features and fault type labels, so as to achieve online rapid diagnosis and analysis of the fault type of the detected process. The use of data-driven fault diagnosis technology will be conducive to dealing with problems such as long processes, nonlinearity, and dynamic time-space correlation in process industries, thereby effectively reducing equipment maintenance costs, improving product yields, saving energy and reducing emissions, and improving production efficiency on the basis of ensuring safe production and product quality, thereby increasing the core competitiveness of enterprises.

[0003] With the continuous improvement of the informationization and intelligence level of the production process, the traditional process monitoring mechanism based on statistical analysis methods can no longer meet the needs of enterprises for feature extraction of complex production processes in the new era. With the help of machine learning artificial intelligence technology, the description of the spatiotemporal information characteristics of data can be enhanced by designing a neural network structure. However, the easy-to-measure data of complex production processes generally have a high dimensionality, and the characteristics of fault-related information are hidden deep. The existing fault diagnosis methods based on deep network structures have the following limitations:

[0004] (1) The ability to mine deep information from data is positively correlated with the complexity of the network structure. The ability of existing artificial intelligence feature extraction methods to extract hidden complex nonlinear features depends on the complexity of their network structure, which leads to an increase in the types of functional parameters to be optimized and may pose a risk of insufficient model generalization ability. This is specifically manifested in constructing optimal phenomena, thereby increasing the difficulty of parameter optimization.

[0005] (2) When fault data cannot be separated, detection is more difficult. Existing fault diagnosis methods have a good recognition effect on linearly separable data. The essential reason is that there is no cross-overlapping area in the data distribution between groups in the original space or simple nonlinear implied space. When there are overlapping areas in the data distribution, the fault pattern recognition method based on cluster analysis is prone to misclassification, which affects the data decoupling and fault isolation results, and directly increases the equipment production and maintenance costs.

[0006] (3) The data feature description mechanism is relatively simple. Although the use of artificial intelligence methods can improve the impact of overlapping distribution conditions between groups of data to a certain extent, when the fault data is similar to or exactly the same as the regular data distribution, the traditional single distribution form parameter characteristics and artificial intelligence hidden layer characteristics may not be able to detect the occurrence of faults.

[0007] In summary, the existing data-driven fault diagnosis technology cannot effectively identify fault data between groups with overlapping or overlapping distributions in a single data feature flow form, and as the model complexity increases, the model parameters are prone to local optimality, which makes it impossible to accurately judge the fault type in complex process industries. Therefore, in order to meet the increasingly complex process industry process monitoring needs, it is urgent to develop a fault diagnosis method based on multi-source feature fusion. Summary of the invention

[0008] The present invention aims to solve the problems existing in the above-mentioned prior art and provides a data-driven fault diagnosis system and method based on texture feature fusion.

[0009] The technical solutions adopted in the present invention are:

[0010] A data-driven fault diagnosis method based on texture feature fusion, comprising:

[0011] S1: Obtain the easily measurable variable data and corresponding fault type labels in the monitored process to form a training data set;

[0012] S2: Extract features from samples of easily measurable variables in the training data set to obtain initial features of the training data;

[0013] S3: Analyze the texture information of the extracted initial features to obtain the texture features of the training data;

[0014] S4: Fusing the initial features of the training data with the texture features to determine the feature fusion scale of the training data;

[0015] S5: Establish a full connection between the training data features and the corresponding sample fault type labels;

[0016] S6: By incorporating data texture information, we can further explore the deep features of the training data and realize the separability identification of data between overlapping or coincident distribution groups to solve the misclassification of fault types.

[0017] Furthermore, in S1, n c The set of easily measurable variables of a discrete process under mode c Corresponding modal fault type l c =c;

[0018] The spline difference is used to supplement or eliminate missing data values. After eliminating data outliers through the 3σ method, the rolling window is used to calculate the discrete process easy-to-measure variable set d c Split the sample and get N c The dimension is w×dim easy-to-measure variable window samples; each easy-to-measure variable window sample is c Constitute a training sample in the training data set;

[0019] Wherein, c = 1, 2, 3...C, C is the number of fault mode types, including the fault-free mode under normal working conditions, dim is the dimension of the easily measurable variable; and w is the window length.

[0020] Furthermore, in S2, N c ×C easy-to-measure variables in training samples As input, a convolutional neural network is used to perform L-layer convolution pooling operations to obtain the initial features of the L layer Among them, N l is the number of convolution kernels, M l It is the number of rows and columns of the feature matrix after convolution and pooling.

[0021] Furthermore, in S3, the initial feature texture information requirement vector is set

[0022] S=[s L ,s L-1 ,…,s2,s1] T

[0023] Among them, s l Binarized variable representing the texture information analysis requirement of the convolution feature of the lth layer (l = 1, 2, ..., L)

[0024]

[0025] Among them, s l = 0 means that there is no need to perform texture analysis on the l-th layer convolutional features, s l =1 means that texture analysis needs to be performed on the lth layer of convolutional features.

[0026] Furthermore, in S3, l = 1, the hidden layer convolution feature performs texture analysis from the perspective of its numerical curve, and uses the Glam matrix to represent the texture feature information, specifically:

[0027] Select s according to the reference variable vector l = 1 hidden layer initial convolution feature F l , calculate its Glam matrix

[0028]

[0029] in, is the i-th row and j-th column element corresponding to the Glam matrix in the l-th layer.

[0030] Furthermore, in S4, fusing the initial features of the training data with the texture features includes:

[0031] Stretch and stack the top initial features of the training data to obtain a 1-dimensional initial feature vector

[0032] Stretch and stack the selected l-layer texture features to obtain a 1-dimensional texture feature vector And define the overall texture features according to the texture information requirement vector

[0033]

[0034] Determine feature F * L 、F * Fusion scale coefficients λ1, λ2, where λ1 represents F * L Fusion ratio, λ2 represents F * Fusion ratio, and λ1+λ2=1.

[0035] Furthermore, in S5, F * L 、F * Build L f Layer fully connected network, construct fusion feature loss function, and obtain the mapping relationship between training data fusion features and corresponding output labels

[0036] L total =λ1L1+λ2L2

[0037] Among them, L1 is the initial feature loss function and L2 is the texture feature loss function.

[0038] The present invention also discloses a data-driven fault diagnosis system based on texture feature fusion, comprising:

[0039] The data acquisition and preprocessing module is used to obtain the easily measurable variable data of the monitored process and the corresponding fault category labels to form a training data set;

[0040] The data initial feature extraction module is used to extract the features of the sample rows of the easily measurable variables in the training data set and obtain the initial features of the training data;

[0041] The feature texture analysis module is used to analyze the texture information of the initial features and obtain the texture features of the training data;

[0042] A feature fusion module is used to process the initial features and texture features of the training data and determine the feature fusion scale of the training data;

[0043] The model training module is used to establish a full connection between the training data features and the corresponding sample fault type labels.

[0044] The present invention has the following beneficial effects:

[0045] (1) By extracting data texture features and further mining effective information beyond the data distribution form from the perspective of data curve texture differences, effective distinction can be made between groups of identically distributed data or overlapping distributed data, thereby improving the accuracy of fault mode identification;

[0046] (2) The Glam matrix is ​​used to represent the feature texture information. The initial convolutional neural network feature extraction results can be used without designing a more complex network structure. The difference from the traditional convolutional neural network model is only reflected in the number of parameters. The increase in model complexity is not too large, which is convenient for model upgrade and maintenance.

[0047] (3) The introduction of an adjustable feature texture information requirement vector and a fusion scale ratio coefficient makes the feature fusion process more flexible and can avoid the loss of traditional feature information due to excessive texture information scale or the invalid feature fusion caused by the insignificant top-level texture features. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of the present invention.

[0049] Figure 2 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings.

[0051] Example 1

[0052] like Figure 1 , a data-driven fault diagnosis method based on texture feature fusion, the method comprising:

[0053] First, collect n c The set of easily measurable variables of a discrete process under a mode c (c = 1, 2, ..., C) And the corresponding modal fault category l c =c, where C is the number of fault mode categories, including the fault-free mode under normal working conditions, and dim is the dimension of easy-to-measure variables; after using spline interpolation to supplement missing data values ​​(or eliminate missing values) and using the 3σ method to eliminate data outliers, a rolling window (window length w, w<<n c ) for the discrete process easily measurable variable set d c To improve the efficiency of sample information and reduce sample repetitiveness, the window rolling step can be consistent with the window length, and N c The dimension is w×dim easy-to-measure variable window samples; each easy-to-measure variable window sample is c Constitute a training sample in the training data set;

[0054] Then, feature extraction is performed on the samples of easy-to-measure variables in the training data set, including:c ×C easy-to-measure variables in training samples As input, the convolutional neural network CNN is used to perform L-layer convolution-pooling operations to obtain the initial features of the L layer Among them, N l is the number of convolution kernels, M l It is convolution-pooling after which the number of feature matrix rows is multiplied by the number of columns. The convolution kernel size and the number of channels can be flexibly set according to the sample dimension to ensure that the top-level feature information is sufficient.

[0055] Afterwards, the texture information of the extracted initial features is analyzed, including: setting the initial feature texture information requirement vector

[0056] S=[s L ,s L-1 ,…,s2,s1] T

[0057] Among them, s l Binarized variable representing the texture information analysis requirement of the convolution feature of the lth layer (l = 1, 2, ..., L)

[0058]

[0059] Among them, s l = 0 means that there is no need to perform texture analysis on the l-th layer convolutional features, s l =1 indicates that texture analysis is required for the lth layer convolutional features; when the hidden layer pooling features vary significantly, the texture information requirement vector can be set to S=[s L =1,s L-1 =0,…,s1=0] T , in order to reduce the number of model parameters

[0060] Furthermore, l = 1, the hidden layer convolution feature performs texture analysis of the feature from the perspective of its numerical curve, and uses the Glam matrix to represent the texture feature information. Specifically: select s according to the reference variable vector l = 1 hidden layer initial convolution feature F l , calculate its Glam matrix

[0061]

[0062] in, is the i-row and j-column element corresponding to the Glam matrix in the l-th layer. The Glam matrix is ​​the inner product of the convolutional reconstruction features of the hidden layer;

[0063] Then, the initial features of the training data are fused with the texture features, including: reconstructing the initial features of the top layer of the training data, stacking the features by row stretching, and obtaining a 1-dimensional initial feature vector

[0064] Furthermore, referring to the texture information requirement vector, the selected texture features of the first layer are first stretched and stacked row by row to obtain a 1-dimensional texture feature vector Then stack the selected hidden layer texture features as a whole

[0065]

[0066] Then, determine the feature F * L 、F * Fusion scale coefficients λ1, λ2, where λ1 represents F * L Fusion ratio, λ2 is enough F * Fusion ratio, and λ1+λ2=1; if the recognition accuracy of the feature fusion model is the same as that of the initial feature model, it means that the texture feature does not work, and λ2 can be appropriately increased; if the recognition accuracy of the feature fusion model is significantly lower than that of the initial feature model, it means that λ2 is too high, which has seriously affected the effective recognition of the linearly separable fault mode by the initial feature, and λ2 can be appropriately reduced at this time;

[0067] Finally, the full connection between the training data features and the corresponding sample fault type labels is established, including: * L 、F * Build L f Layer fully connected network, construct fusion feature loss function, and obtain the mapping relationship between training data fusion features and corresponding output labels

[0068] L total =λ1L1+λ2L2

[0069] Among them, L1 is the initial feature loss function, and L2 is the texture feature loss function. If new fault types are added later, the new faults and fully connected parameters can be added from the output dimension, and the original model parameters can be used as the initial value reference to update the parameters.

[0070] Example 2

[0071] like Figure 2 , a data-driven fault diagnosis system based on texture feature fusion, comprising:

[0072] The data acquisition and preprocessing module is used to obtain the easily measurable variable data of the monitored process and the corresponding fault category labels to form a training data set; specifically, collect n c The set of easily measurable variables of a discrete process under a mode c (c = 1, 2, ..., C) And the corresponding modal fault category l c=c, where C is the number of fault mode categories, including the fault-free mode under normal working conditions, and dim is the dimension of easy-to-measure variables; after using spline interpolation to supplement missing data values ​​(or eliminate missing values) and using the 3σ method to eliminate data outliers, a rolling window (window length w, w<<n c ) for the discrete process easily measurable variable set d c To improve the efficiency of sample information and reduce sample repetitiveness, the window rolling step can be consistent with the window length, and N c The dimension is w×dim easy-to-measure variable window samples; each easy-to-measure variable window sample is c Constitute a training sample in the training data set;

[0073] The data initial feature extraction module is used to extract the features of the sample rows of the easily measurable variables in the training data set to obtain the initial features of the training data; specifically, N c ×C easy-to-measure variables in training samples As input, a convolutional neural network is used to perform L-layer convolution-pooling operations to obtain the initial features of the L layer Among them, N l is the number of convolution kernels, M l It is convolution-pooling after which the number of feature matrix rows is multiplied by the number of columns. The convolution kernel size and the number of channels can be flexibly set according to the sample dimension to ensure that the top-level feature information is sufficient.

[0074] The feature texture analysis module is used to analyze the texture information of the initial features and obtain the texture features of the training data; specifically, the texture information analysis of the extracted initial features includes: setting the initial feature texture information requirement vector

[0075] S=[s L ,s L-1 ,…,s2,s1] T

[0076] Among them, s l Binarized variable representing the texture information analysis requirement of the convolution feature of the lth layer (l = 1, 2, ..., L)

[0077]

[0078] Among them, s l = 0 means that there is no need to perform texture analysis on the l-th layer convolutional features, s l =1 indicates that texture analysis is required for the lth layer convolutional features; when the hidden layer pooling features vary significantly, the texture information requirement vector can be set to S=[s L =1,s L-1 =0,…,s1=0] T , in order to reduce the number of model parameters

[0079] Furthermore, l = 1, the hidden layer convolution feature performs texture analysis of the feature from the perspective of its numerical curve, and uses the Glam matrix to represent the texture feature information. Specifically: select s according to the reference variable vector l = 1 hidden layer initial convolution feature F l , calculate its Glam matrix

[0080]

[0081] in, is the i-row and j-column element corresponding to the Glam matrix in the l-th layer. The Glam matrix is ​​the inner product of the convolutional reconstruction features of the hidden layer;

[0082] The feature fusion module is used to process the initial features and texture features of the training data and determine the feature fusion scale of the training data; specifically, the initial features of the top layer of the training data are reconstructed, and the features are stacked by row stretching to obtain a 1-dimensional initial feature vector.

[0083] Furthermore, referring to the texture information requirement vector, the selected texture features of the first layer are first stretched and stacked row by row to obtain a 1-dimensional texture feature vector Then stack the selected hidden layer texture features as a whole

[0084]

[0085] Then, determine the feature F * L 、F * Fusion scale coefficients λ1, λ2, where λ1 represents F * L Fusion ratio, λ2 is enough F * Fusion ratio, and λ1+λ2=1; if the recognition accuracy of the feature fusion model is the same as that of the initial feature model, it means that the texture feature does not work, and λ2 can be appropriately increased; if the recognition accuracy of the feature fusion model is significantly lower than that of the initial feature model, it means that λ2 is too high, which has seriously affected the effective recognition of the linearly separable fault mode by the initial feature, and λ2 can be appropriately reduced at this time;

[0086] The model training module is used to establish a full connection between the training data features and the corresponding sample fault type labels; specifically, F * L 、F * Build L f Layer fully connected network, construct fusion feature loss function, and obtain the mapping relationship between training data fusion features and corresponding output labels

[0087] Ltotal =λ1L1+λ2L2

[0088] Among them, L1 is the initial feature loss function, and L2 is the texture feature loss function. If new fault types are added later, the new faults and fully connected parameters can be added from the output dimension, and the original model parameters can be used as the initial value reference to update the parameters.

[0089] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. A data-driven fault diagnosis method based on texture feature fusion, characterized in that: include: S1: Obtain the easily measurable variable data and corresponding fault type labels in the monitored process to form a training data set; S2: Extract features from samples of easily measurable variables in the training data set to obtain initial features of the training data; S3: Analyze the texture information of the extracted initial features to obtain the texture features of the training data; S4: Fusing the initial features of the training data with the texture features to determine the feature fusion scale of the training data; S5: Establish a full connection between the training data features and the corresponding sample fault type labels; S6: By incorporating data texture information, we can further explore the deep features of the training data and realize the separability identification of data between overlapping or coincident distribution groups to solve the misclassification of fault types.

2. The data-driven fault diagnosis method based on texture feature fusion according to claim 1, characterized in that: In S1, collect n c The set of easily measurable variables of a discrete process under mode c Corresponding modal fault type l c =c; The spline difference is used to supplement or eliminate missing data values. After eliminating data outliers through the 3σ method, the rolling window is used to calculate the discrete process easy-to-measure variable set d c Split the sample and get N c The dimension is w×dim easy-to-measure variable window samples; Each easily measurable variable window sample is c Constitute a training sample in the training data set; Wherein, c = 1, 2, 3...C, C is the number of fault mode types, including the fault-free mode under normal working conditions, dim is the dimension of the easily measurable variable; and w is the window length.

3. The data-driven fault diagnosis method based on texture feature fusion according to claim 2, characterized in that: In S2, N c ×C easy-to-measure variables in training samples As input, a convolutional neural network is used to perform L-layer convolution pooling operations to obtain the initial features of the L layer Among them, N l is the number of convolution kernels, M l It is the number of rows and columns of the feature matrix after convolution and pooling.

4. The data-driven fault diagnosis method based on texture feature fusion according to claim 3, characterized in that: In S3, set the initial feature texture information requirement vector S=[s L ,s L-1 ,…,s2,s1] T Among them, s l Binarized variable representing the texture information analysis requirement of the convolution feature of the lth layer (l = 1, 2, ..., L) Among them, s l = 0 means that there is no need to perform texture analysis on the l-th layer convolutional features, s l =1 means that texture analysis needs to be performed on the lth layer of convolutional features.

5. The data-driven fault diagnosis method based on texture feature fusion according to claim 4, characterized in that: In S3, l = 1, the hidden layer convolution feature performs texture analysis from the perspective of its numerical curve, and uses the Glam matrix to represent the texture feature information, specifically: Select s according to the reference variable vector l = 1 hidden layer initial convolution feature F l , calculate its Glam matrix in, is the i-th row and j-th column element corresponding to the Glam matrix in the l-th layer.

6. The data-driven fault diagnosis method based on texture feature fusion according to claim 5, characterized in that: In S4, the fusion of the initial features of the training data and the texture features includes: Stretch and stack the top initial features of the training data to obtain a 1-dimensional initial feature vector Stretch and stack the selected l-layer texture features to obtain a 1-dimensional texture feature vector And define the overall texture features according to the texture information requirement vector Determine the characteristics F * Fusion scale coefficients λ1, λ2, where λ1 represents Fusion ratio, λ2 represents F * Fusion ratio, and λ1+λ2=1.

7. The data-driven fault diagnosis method based on texture feature fusion according to claim 6, characterized in that: In S5, F * Build L f Layer fully connected network, construct fusion feature loss function, and obtain the mapping relationship between training data fusion features and corresponding output labels L total =λ1L1+λ2L2 Among them, L1 is the initial feature loss function and L2 is the texture feature loss function.

8. A data-driven fault diagnosis system based on texture feature fusion, characterized in that: include The data acquisition and preprocessing module is used to obtain the easily measurable variable data of the monitored process and the corresponding fault category labels to form a training data set; The data initial feature extraction module is used to extract the features of the sample rows of the easily measurable variables in the training data set and obtain the initial features of the training data; The feature texture analysis module is used to analyze the texture information of the initial features and obtain the texture features of the training data; A feature fusion module is used to process the initial features and texture features of the training data and determine the feature fusion scale of the training data; The model training module is used to establish a full connection between the training data features and the corresponding sample fault type labels.

Images