A process fault diagnosis method based on two-stage recurrent network

Abstract

The present invention belongs to the technical field of process fault diagnosis, and relates to a process fault diagnosis method based on a two-stage recurrent network, which is used for high-accuracy real-time identification of fault types, and its main steps include: establishing a first-stage fault classifier, fault grouping, establishing a second-stage fault classifier and identifying fault types; extracting long-term trend features by using an independent recurrent neural network, and extracting short-term dynamic features by using an attention full convolutional network based on compression excitation and global time attention; establishing a diagnostic difficulty evaluation index, and dividing faults into difficult-to-diagnose faults and easy-to-diagnose faults; the second-stage classifier performs fine processing on difficult-to-diagnose faults, and uses principal component analysis statistics to augment measurement vectors, and uses robust normalization to improve the generalization ability of the model, which can effectively enhance the classification feature extraction ability and improve the process fault diagnosis accuracy. Its overall process is simple, the principle is reliable, the fault diagnosis effect is good, the adaptability is wide, the logic is strong, and the environment is friendly.

Description

Technical Field

[0001] The present invention belongs to the technical field of process fault diagnosis and relates to a process fault diagnosis method based on a two-stage cyclic network. By adopting a two-stage dynamic independent cyclic neural network to extract long-term trend characteristics and short-term dynamic characteristics of the fault process, it is possible to identify the fault type under abnormal process conditions. Background Art

[0002] With the development of automation and artificial intelligence technology, industrial process fault diagnosis technology has received more and more attention. Online fault diagnosis technology can provide important information about the fault to the maintenance team, which helps to eliminate the fault as soon as possible and restore normal production operation. Complex industrial processes generally contain a large number of measurement variables and have nonlinear characteristics. Due to large devices and long processes, industrial processes generally have dynamic characteristics, that is, short-term time series have time correlation. The process of production operation deterioration caused by faults often has trend characteristics, that is, long-term time series have time correlation. In order to accurately diagnose the type of fault, it is necessary to fully consider the above characteristics of the industrial process. The fault diagnosis method based on deep learning achieves powerful feature extraction and learning capabilities through a deeper and more complex network structure, and can more effectively process the fault data of complex processes. The existing deep learning method generally uses a single-level classifier, which does not simultaneously consider the long-term trend characteristics and short-term dynamic characteristics of the process fault data, resulting in a high overall accuracy of fault diagnosis, and the diagnosis accuracy of some faults is always low. Therefore, the invention of a process fault diagnosis method based on a two-level recurrent network can improve the shortcomings of the existing technology, can accurately diagnose the type of process faults in real time, reduce fault losses, and increase the generalization ability of the overall method. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art. Based on the improvement of the prior art, a high-efficiency nonlinear multivariate fault diagnosis method is designed to utilize a dynamic independent recurrent neural network and can extract long-term trend characteristics and short-term dynamic characteristics in process fault data, in particular, a process fault diagnosis method based on a two-stage recurrent network is designed to solve the problems of the prior art that rely on prior knowledge, have poor generalization performance, and have low diagnostic accuracy.

[0004] In order to achieve the above objectives, the present invention provides a process fault diagnosis method based on a two-stage recurrent network. The method first uses an independent recurrent neural network to extract long-term trend features, uses an attention full convolutional network to extract short-term dynamic features, establishes a first-level classifier to identify all fault types, and then combines principal component analysis and robust normalization to establish a second-level classifier to finely identify difficult-to-diagnose faults, which can improve the fault feature extraction capability and thus improve the process fault diagnosis results.

[0005] The present invention provides a process fault diagnosis method based on a two-stage cyclic network, comprising the following steps:

[0006] S1: Establish the first-level fault classifier: collect historical fault data and construct a training set , validation set and test set , training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting; Dynamic independent recurrent neural network uses independent recurrent neural network to extract trend characteristics of fault process, uses attention full convolution network to extract dynamic characteristics of fault process, and uses softmax layer to obtain fault type prediction value;

[0007] S2: Fault grouping: checksum The data is input into the first-level dynamic independent recurrent neural network to obtain the fault type prediction value, calculate the confusion rate, and classify all fault types into easy-to-diagnose faults and difficult-to-diagnose faults;

[0008] S3: Establish the second-level fault classifier: collect normal operation data, establish the principal component analysis model and , Statistics; collect training data for difficult-to-diagnose faults and calculate and Statistics, build augmented vectors, perform robust normalization, and train the second-level dynamic independent recurrent neural network;

[0009] S4: Identify fault type: collect real-time process data, perform standardization processing, input it into the first-level dynamic independent recurrent neural network, and obtain the fault type prediction value; when the predicted type belongs to a difficult-to-diagnose fault, construct an augmented vector, perform robust standardization processing, and input it into the second-level dynamic independent recurrent neural network to obtain the final fault type prediction value.

[0010] The present invention relates to a process fault diagnosis method based on a two-stage cyclic network, wherein step S1 specifically includes the following steps:

[0011] S101: Collect historical fault data and construct training sets , validation set and test set ; Training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting;

[0012] S102: Constructing measurement vector using process measurement data ,in represents the sampling time, Indicates the number of variables; for the measurement vector Standardization is performed so that the mean and variance of each variable are distributed to 0 and 1, and a standardized vector is obtained. , construct a time series window ,in Indicates the time series window length;

[0013] S103: The first feature extraction subprocess includes an independent recurrent neural network and a discard layer; the matrix Input independent recurrent neural network; in the first layer of the network, The time corresponding to The output of a neuron is ,

[0014] Formula (1)

[0015] in , is the number of hidden layer neurons, represents the rectified linear unit activation function, Represents the input weight vector, coefficient represents the cycle weight, represents the bias term; layer, with The time corresponding to The output of a neuron is ,

[0016] Formula (2)

[0017] in , is the number of layers of independent recurrent neural network, represents the input weight vector, The output vector of the layer It is The input of the layer of neurons; through deep feature mining, the output vector of the last layer of the independent recurrent neural network is obtained ;

[0018] S104: In order to avoid overfitting of the model, the vector Input the discard layer to obtain long-term time series features ,

[0019] Formula (3)

[0020] where the vector ,element , represents the Bernoulli distribution, express The probability of, operator represents the Hadamard product;

[0021] S105: The second feature extraction subprocess is an attention full convolutional network, including an ACBRD module and a global average pooling layer; the ACBRD module includes an attention module, a one-dimensional temporal convolution layer, a batch normalization layer, a rectified linear unit activation layer, and a discard layer; the attention module includes a compression excitation submodule and a global temporal attention submodule; the matrix Input compression excitation submodule, calculate the mean vector ,

[0022] Formula (4)

[0023] and weight vector ,

[0024] Formula (5)

[0025] in and is the parameter matrix to be learned, Represents the dimensionality reduction rate; using the weight vector Weight the input data to get the output matrix ,

[0026] Formula (6)

[0027] in Represents a diagonal matrix constructor;

[0028] S106: compress the output matrix of the excitation submodule Enter the global time attention submodule and calculate the weight vector ,

[0029] Formula (7)

[0030] in , and is the parameter matrix to be learned; using the weight vector Weight the data to get the output matrix ,

[0031] Formula (8)

[0032] S107: The weighted time series window The output of the first ACBRD module is obtained by sequentially inputting the one-dimensional temporal convolution layer, batch normalization layer, rectified linear unit activation layer and discard layer for processing. ; In order to extract deep features, ACBRD modules are connected in series to Process and output features ; The feature Input the global average pooling layer for processing to obtain short-term dynamic features ;

[0033] S108: Trend characteristics and dynamic features Concatenate and input into the softmax layer for processing to obtain the fault type prediction value .

[0034] The present invention relates to a process fault diagnosis method based on a two-stage cyclic network, wherein step S2 specifically includes the following steps:

[0035] S201: Verify the set The data of the first-level fault classifier is input to obtain the fault type prediction value, and the predicted fault type is compared with the actual fault type to calculate the Confusion rate of class faults ,

[0036] Formula (9)

[0037] in represents the number of false negative samples, represents the number of false positive samples, Indicates the number of true samples;

[0038] S202: Confusion rate Make a judgment, when When the control limit is exceeded, it is considered Type 1 fault is a difficult-to-diagnose fault; otherwise, it is considered This type of fault is an easy-to-diagnose fault; establish an easy-to-diagnose fault set and difficult to diagnose fault sets .

[0039] The present invention relates to a process fault diagnosis method based on a two-stage cyclic network, wherein step S3 specifically includes the following steps:

[0040] S301: Constructing a matrix using normal operation data , establish the principal component model; principal component analysis aims to find

[0041] Formula (10)

[0042] The projection vector ; Calculate the projection matrix by singular value decomposition and the standard deviation matrix ; Select The largest standard deviation, construct the standard deviation matrix , using the projection vectors corresponding to these standard deviations to construct the principal element space projection matrix ;

[0043] S302: For the normalized measurement vector ,calculate and Statistics,

[0044] Formula (11)

[0045] Formula (12)

[0046] in represents the identity matrix;

[0047] S303: Exploit and Statistics augment the measurement vector and construct the vector ; In order to reduce the impact of outliers and noise and enhance the generalization ability of the model, the vector Perform robust normalization; use express No. variables, and calculate their standardized values ,

[0048] Formula (13)

[0049] in and Respectively represent The median and interquartile range of each variable;

[0050] S304: Constructing vectors using data from difficult-to-diagnose faults and time series window ,Will As input, the fault type is output; using the training set Establish a new training set based on the difficult-to-diagnose fault samples in , using the validation set A new verification set is established based on the difficult-to-diagnose fault samples in ; Using the training set Train the second-level dynamic independent recurrent neural network, using the validation set Adjust network parameters.

[0051] The present invention relates to a process fault diagnosis method based on a two-stage cyclic network, wherein step S4 specifically includes the following steps:

[0052] S401: Collect real-time measurement data, perform standardization processing, and construct a time series window ;

[0053] S402: Independent recurrent neural network input into the first level classifier to calculate trend features ;Will Input the attention full convolutional network of the first-level classifier to calculate dynamic features ;

[0054] S403: Features and Concatenate and input into the softmax layer to obtain the fault type prediction value ;

[0055] S404: Yes Make a judgment, if , the fault to be diagnosed is considered to be an easy-to-diagnose fault, and the fault type prediction value is output ; Otherwise, the fault to be diagnosed is considered to be a difficult-to-diagnose fault, and steps S405 and S406 are executed for further diagnosis;

[0056] S405: Calculation and Statistics, construct augmented vectors , perform robust standardization and construct a time series window ;

[0057] S406: Input the dynamic independent recurrent neural network of the second-level classifier to obtain the final fault type prediction value .

[0058] Compared with the prior art, the present invention has the following advantages: (1) The process fault diagnosis method provided by the present invention identifies the fault type by analyzing the process data collected in real time, is applicable to nonlinear multivariable continuous industrial processes, and can meet the process safety monitoring requirements; (2) The process faults are divided into easy-to-diagnose faults and difficult-to-diagnose faults according to the confusion rate, and the easy-to-diagnose faults are identified by the first-level classifier, and the difficult-to-diagnose faults are finely identified by the second-level classifier, which can improve the fault diagnosis accuracy; (3) The long-term trend features are extracted by using an independent recurrent neural network, and the short-term dynamic features are extracted by using an attention full convolutional network, thereby realizing the simultaneous extraction of trend features and dynamic features; the key process variables are highlighted by using a compression excitation submodule, and the important sampling moments are highlighted by using a global time attention submodule, thereby realizing the simultaneous strengthening of time and space attention; (4) The principal component analysis is used to extract the long-term trend features and the short-term dynamic features are extracted by using an attention full convolutional network. and The statistical augmentation of the measurement vector enhances the sensitivity to faults, improves the generalization ability of the model through robust normalization, and enhances the ability to accurately identify difficult-to-diagnose faults; the overall process is simple, the principle is reliable, the fault type diagnosis is accurate, the real-time safety monitoring effect is good, the adaptability is wide, the logic is strong, and the environment is friendly. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The present invention relates to a flow chart of the working principle of the fault diagnosis method.

[0060] Figure 2 The present invention is a process flow chart of the fault diagnosis method.

[0061] Figure 3 This is the overall structural diagram of the dynamic independent recurrent neural network involved in the present invention.

[0062] Figure 4 This is a local structural diagram of the independent recurrent neural network involved in the present invention.

[0063] Figure 5 This is a local structure diagram of the attention full convolutional network involved in the present invention.

[0064] Figure 6 The present invention provides a structural diagram of the Tennessee-Eastman process, a nonlinear chemical engineering model.

[0065] Figure 7 is the fault diagnosis accuracy of the first-level classifier involved in the embodiment of the present invention

[0066] Figure 8 is the fault confusion rate of the first-level classifier involved in the embodiment of the present invention.

[0067] Fig. 9 The figure is a comparison chart of the accuracy of six fault diagnosis methods involved in the embodiments of the present invention.

[0068] Fig.10 This is a comparison chart of the accuracies of four fault diagnosis methods in the ablation experiment involved in the embodiment of the present invention. DETAILED DESCRIPTION

[0069] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0070] Example 1

[0071] like Figure 1 As shown, this embodiment relates to a process fault diagnosis method based on a two-stage cyclic network, comprising the following steps:

[0072] S1: Establish the first-level fault classifier: collect historical fault data and construct a training set , validation set and test set , training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting; Dynamic independent recurrent neural network uses independent recurrent neural network to extract trend characteristics of fault process, uses attention full convolution network to extract dynamic characteristics of fault process, and uses softmax layer to obtain fault type prediction value;

[0073] S2: Fault grouping: checksum The data is input into the first-level dynamic independent recurrent neural network to obtain the fault type prediction value, calculate the confusion rate, and classify all fault types into easy-to-diagnose faults and difficult-to-diagnose faults;

[0074] S3: Establish the second-level fault classifier: collect normal operation data, establish the principal component analysis model and , Statistics; collect training data for difficult-to-diagnose faults and calculate and Statistics, build augmented vectors, perform robust normalization, and train the second-level dynamic independent recurrent neural network;

[0075] S4: Identify fault type: collect real-time process data, perform standardization processing, input it into the first-level dynamic independent recurrent neural network, and obtain the fault type prediction value; when the predicted type belongs to a difficult-to-diagnose fault, construct an augmented vector, perform robust standardization processing, and input it into the second-level dynamic independent recurrent neural network to obtain the final fault type prediction value.

[0076] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S1 specifically includes the following steps:

[0077] S101: Collect historical fault data and construct training sets , validation set and test set ; Training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting;

[0078] S102: Constructing measurement vector using process measurement data ,in represents the sampling time, Indicates the number of variables; for the measurement vector Standardization is performed so that the mean and variance of each variable are distributed to 0 and 1, and a standardized vector is obtained. , construct a time series window ,in Indicates the time series window length;

[0079] S103: The first feature extraction subprocess includes an independent recurrent neural network and a discard layer; the matrix Input independent recurrent neural network; in the first layer of the network, The time corresponding to The output of a neuron is ,

[0080] Formula (1)

[0081] in , is the number of hidden layer neurons, represents the rectified linear unit activation function, Represents the input weight vector, coefficient represents the cycle weight, represents the bias term; layer, with The time corresponding to The output of a neuron is ,

[0082] Formula (2)

[0083] in , is the number of layers of independent recurrent neural network, represents the input weight vector, The output vector of the layer It is The input of the layer of neurons; through deep feature mining, the output vector of the last layer of the independent recurrent neural network is obtained ;

[0084] S104: In order to avoid overfitting of the model, the vector Input the discard layer to obtain long-term time series features ,

[0085] Formula (3)

[0086] where the vector ,element , represents the Bernoulli distribution, express The probability of, operator represents the Hadamard product;

[0087] S105: The second feature extraction subprocess is an attention full convolutional network, including an ACBRD module and a global average pooling layer; the ACBRD module includes an attention module, a one-dimensional temporal convolution layer, a batch normalization layer, a rectified linear unit activation layer, and a discard layer; the attention module includes a compression excitation submodule and a global temporal attention submodule; the matrix Input compression excitation submodule, calculate the mean vector ,

[0088] Formula (4)

[0089] and weight vector ,

[0090] Formula (5)

[0091] in and is the parameter matrix to be learned, Represents the dimensionality reduction rate; using the weight vector Weight the input data to get the output matrix ,

[0092] Formula (6)

[0093] in Represents a diagonal matrix constructor;

[0094] S106: compress the output matrix of the excitation submodule Enter the global time attention submodule and calculate the weight vector ,

[0095] Formula (7)

[0096] in , and is the parameter matrix to be learned; using the weight vector Weight the data to get the output matrix ,

[0097] Formula (8)

[0098] S107: The weighted time series window The output of the first ACBRD module is obtained by sequentially inputting the one-dimensional temporal convolution layer, batch normalization layer, rectified linear unit activation layer and discard layer for processing. ; In order to extract deep features, ACBRD modules are connected in series to Process and output features ; The feature Input the global average pooling layer for processing to obtain short-term dynamic features ;

[0099] S108: Trend characteristics and dynamic features Concatenate and input into the softmax layer for processing to obtain the fault type prediction value .

[0100] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S2 specifically includes the following steps:

[0101] S201: Verify the set The data of the first-level fault classifier is input to obtain the fault type prediction value, and the predicted fault type is compared with the actual fault type to calculate the Confusion rate of class faults ,

[0102] Formula (9)

[0103] in represents the number of false negative samples, represents the number of false positive samples, Indicates the number of true samples;

[0104] S202: Confusion rate Make a judgment, when When the control limit is exceeded, it is considered Type 1 fault is a difficult-to-diagnose fault; otherwise, it is considered This type of fault is an easy-to-diagnose fault; establish an easy-to-diagnose fault set and difficult to diagnose fault sets .

[0105] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S3 specifically includes the following steps:

[0106] S301: Constructing a matrix using normal operation data , establish the principal component model; principal component analysis aims to find

[0107] Formula (10)

[0108] The projection vector ; Calculate the projection matrix by singular value decomposition and the standard deviation matrix ; Select The largest standard deviation, construct the standard deviation matrix , using the projection vectors corresponding to these standard deviations to construct the principal element space projection matrix ;

[0109] S302: For the normalized measurement vector ,calculate and Statistics,

[0110] Formula (11)

[0111] Formula (12)

[0112] in represents the identity matrix;

[0113] S303: Exploit and Statistics augment the measurement vector and construct the vector ; In order to reduce the impact of outliers and noise and enhance the generalization ability of the model, the vector Perform robust normalization; use express No. variables, and calculate their standardized values ,

[0114] Formula (13)

[0115] in and Respectively represent The median and interquartile range of the variables;

[0116] S304: Constructing vectors using data from difficult-to-diagnose faults and time series window ,Will As input, the fault type is output; using the training set Establish a new training set based on the difficult-to-diagnose fault samples in , using the validation set A new verification set is established based on the difficult-to-diagnose fault samples in ; Using the training set Train the second-level dynamic independent recurrent neural network, using the validation set Adjust network parameters.

[0117] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S4 specifically includes the following steps:

[0118] S401: Collect real-time measurement data, perform standardization processing, and construct a time series window ;

[0119] S402: Independent recurrent neural network input into the first level classifier to calculate trend features ;Will Input the attention full convolutional network of the first-level classifier to calculate dynamic features ;

[0120] S403: Features and Concatenate and input into the softmax layer to obtain the fault type prediction value ;

[0121] S404: Yes Make a judgment, if , the fault to be diagnosed is considered to be an easy-to-diagnose fault, and the fault type prediction value is output ; Otherwise, the fault to be diagnosed is considered to be a difficult-to-diagnose fault, and steps S405 and S406 are executed for further diagnosis;

[0122] S405: Calculation and Statistics, construct augmented vectors , perform robust standardization and construct a time series window ;

[0123] S406: Input the dynamic independent recurrent neural network of the second-level classifier to obtain the final fault type prediction value .

[0124] like Figure 2 As shown, in the above method, steps S1 to S3 are the offline modeling stage, and step S4 is the online diagnosis stage.

[0125] Example 2

[0126] like Figure 3-Figure 8 As shown in FIG. 1 , this embodiment applies the process fault diagnosis method based on the two-stage cycle network to the Tennessee-Eastman process. Downs and Vogel of the American Chemical Company developed the Tennessee-Eastman process based on a real chemical process. The process is widely used to test and evaluate process control methods and fault diagnosis methods. The Tennessee-Eastman process mainly includes five typical chemical units, such as Figure 6 As shown, they are the reactor, condenser, stripper, compressor and separator. The Tennessee-Eastman process involves 8 components from A to H; Figure 6The numbers 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 represent a fluid, which is referred to as stream 1, stream 2, stream 3, stream 4, stream 5, stream 6, stream 7, stream 8, stream 9, stream 10, stream 11, stream 12, stream 13. The Tennessee-Eastman process of this embodiment includes 52 measurement variables (as shown in Table 1), where variables 1-11 are control variables and variables 12-52 are process variables. The Tennessee-Eastman process of this embodiment simulates 20 faults (as shown in Table 2), where faults 1-7 are step changes in process variables, faults 8-12 are increased volatility of process variables, fault 13 is slow offset of reaction kinetics, faults 14-15 are valve sticking, and faults 15-20 are unknown faults.

[0127] Table 1 Measurement variables of the Tennessee-Eastman process

[0128]

[0129] Table 2 Fault description of the Tennessee-Eastman process

[0130]

[0131] The specific implementation steps of the process fault diagnosis method based on two-stage cyclic network for the Tennessee-Eastman process fault diagnosis are as follows:

[0132] S1: Establish the first-level fault classifier: collect historical fault data and construct a training set , validation set and test set , training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting; Dynamic independent recurrent neural network uses independent recurrent neural network to extract trend characteristics of fault process, uses attention full convolution network to extract dynamic characteristics of fault process, and uses softmax layer to obtain fault type prediction value;

[0133] S2: Fault grouping: checksum The data is input into the first-level dynamic independent recurrent neural network to obtain the fault type prediction value, calculate the confusion rate, and classify all fault types into easy-to-diagnose faults and difficult-to-diagnose faults;

[0134] S3: Establish the second-level fault classifier: collect normal operation data, establish the principal component analysis model and , Statistics; collect training data for difficult-to-diagnose faults and calculate and Statistics, build augmented vectors, perform robust normalization, and train the second-level dynamic independent recurrent neural network;

[0135] S4: Identify fault type: collect real-time process data, perform standardization processing, input it into the first-level dynamic independent recurrent neural network, and obtain the fault type prediction value; when the predicted type belongs to a difficult-to-diagnose fault, construct an augmented vector, perform robust standardization processing, and input it into the second-level dynamic independent recurrent neural network to obtain the final fault type prediction value.

[0136] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S1 specifically includes the following steps:

[0137] S101: Collect historical fault data and construct training sets , validation set and test set ; Training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting;

[0138] S102: Constructing measurement vector using process measurement data ,in Represents the sampling time; for the measurement vector Standardization is performed so that the mean and variance of each variable are distributed to 0 and 1, and a standardized vector is obtained. , construct a time series window ;

[0139] S103: The first feature extraction subprocess includes an independent recurrent neural network and a discard layer; the matrix Input independent recurrent neural network; in the first layer of the network, The time corresponding to The output of a neuron is ,

[0140] Formula (1)

[0141] in , represents the rectified linear unit activation function, Represents the input weight vector, coefficient represents the cycle weight, represents the bias term; layer, with The time corresponding to The output of a neuron is ,

[0142] Formula (2)

[0143] in , represents the input weight vector, The output vector of the layer It is The input of the layer of neurons; through deep feature mining, the output vector of the last layer of the independent recurrent neural network is obtained ;

[0144] S104: In order to avoid overfitting of the model, the vector Input the discard layer to obtain long-term time series features ,

[0145] Formula (3)

[0146] where the vector ,element , represents Bernoulli distribution, 0.2 represents The probability of, operator represents the Hadamard product;

[0147] S105: The second feature extraction sub-process is an attention full convolutional network, including 3 ACBRD modules and a global average pooling layer; the ACBRD module includes an attention module, a one-dimensional temporal convolution layer, a batch normalization layer, a rectified linear unit activation layer, and a discard layer; the attention module includes a compression excitation submodule and a global temporal attention submodule; the matrix Input compression excitation submodule, calculate the mean vector ,

[0148] Formula (4)

[0149] and weight vector ,

[0150] Formula (5)

[0151] in and is the parameter matrix to be learned, 16 represents the dimensionality reduction rate; using the weight vector Weight the input data to get the output matrix ,

[0152] Formula (6)

[0153] in Represents a diagonal matrix constructor;

[0154] S106: compress the output matrix of the excitation submodule Enter the global time attention submodule and calculate the weight vector ,

[0155] Formula (7)

[0156] in , and is the parameter matrix to be learned; using the weight vector Weight the data to get the output matrix ,

[0157] Formula (8)

[0158] S107: The weighted time series window The output of the first ACBRD module is obtained by sequentially inputting the one-dimensional temporal convolution layer, batch normalization layer, rectified linear unit activation layer and discard layer for processing. ; In order to extract deep features, three ACBRD modules are connected in series to the time series window Process and output features , where the number of convolution kernels in each module is 128, 256, and 128 respectively; Input the global average pooling layer for processing to obtain short-term dynamic features ;

[0159] S108: Trend characteristics and dynamic features Concatenate and input into the softmax layer for processing to obtain the fault type prediction value .

[0160] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S2 specifically includes the following steps:

[0161] S201: Verify the set The data of the first-level fault classifier is input to obtain the fault type prediction value, and the predicted fault type is compared with the actual fault type to calculate the Confusion rate of class faults ,

[0162] Formula (9)

[0163] in represents the number of false negative samples, represents the number of false positive samples, Indicates the number of true samples;

[0164] S202: Confusion rate Make a judgment, when When the control limit is exceeded by 5%, it is considered Type 1 fault is a difficult-to-diagnose fault; otherwise, it is considered This type of fault is an easy-to-diagnose fault; establish an easy-to-diagnose fault set and difficult to diagnose fault sets .

[0165] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S3 specifically includes the following steps:

[0166] S301: Constructing a matrix using normal operation data , establish the principal component model; principal component analysis aims to find

[0167] Formula (10)

[0168] The projection vector ; Calculate the projection matrix by singular value decomposition and the standard deviation matrix ; Select the 27 largest standard deviations and construct the standard deviation matrix , using the projection vectors corresponding to these standard deviations to construct the principal element space projection matrix ;

[0169] S302: For the normalized measurement vector ,calculate and Statistics,

[0170] Formula (11)

[0171] Formula (12)

[0172] in represents the identity matrix;

[0173] S303: Exploit and Statistics augment the measurement vector and construct the vector ; In order to reduce the impact of outliers and noise and enhance the generalization ability of the model, the vector Perform robust normalization; use express No. variables, and calculate their standardized values ,

[0174] Formula (13)

[0175] in and Respectively represent The median and interquartile range of the variables;

[0176] S304: Constructing vectors using data from difficult-to-diagnose faults and time series window ,Will As input, the fault type is output; using the training set The difficult-to-diagnose fault samples in the , using the validation set A new verification set is established based on the difficult-to-diagnose fault samples in ; Using the training set Train the second-level dynamic independent recurrent neural network, using the validation set Adjust network parameters.

[0177] The process fault diagnosis method based on the two-stage cyclic network involved in this embodiment, wherein step S4 specifically includes the following steps:

[0178] S401: Collect real-time measurement data, perform standardization processing, and construct a time series window ;

[0179] S402: Independent recurrent neural network input into the first level classifier to calculate trend features ;Will Input the attention full convolutional network of the first-level classifier to calculate dynamic features ;

[0180] S403: Features and Concatenate and input into the softmax layer to obtain the fault type prediction value ;

[0181] S404: Yes Make a judgment, if , the fault to be diagnosed is considered to be an easy-to-diagnose fault, and the fault type prediction value is output ; Otherwise, the fault to be diagnosed is considered to be a difficult-to-diagnose fault, and steps S405 and S406 are executed for further diagnosis;

[0182] S405: Calculation and Statistics, construct augmented vectors , perform robust standardization and construct a time series window ;

[0183] S406: Input the dynamic independent recurrent neural network of the second-level classifier to obtain the final fault type prediction value .

[0184] Embodiment 3:

[0185] In this embodiment, when establishing the first-level fault classifier, the fault diagnosis accuracy of the verification set is as follows: Figure 7 It can be seen that the diagnostic accuracy of faults 10, 15 and 18 is relatively low, which are 65.3%, 12% and 46.5% respectively, while the diagnostic accuracy of the remaining faults is higher than 90%.

[0186] Example 4

[0187] In the fault grouping stage of this embodiment, it can be seen from the confusion matrix of the check set that for fault 10, 9.5% of the samples are misidentified as fault 9, and 25.1% of the samples are misidentified as fault 16. For fault 15, 87.8% of the samples are misidentified as fault 9. For fault 18, 53.5% of the samples are misidentified as fault 16. All faults are sorted according to the confusion rate values, as shown in Figure 1. Figure 8 As shown in the figure, it can be seen that although the diagnostic accuracy of fault 9 is higher than 90%, its confusion rate is the largest. This is because many samples of faults 10 and 15 are misdiagnosed as fault 9. Finally, faults 9, 15, 16, 18, 10 and 12 are classified as difficult-to-diagnose faults, and the remaining 14 faults are easy-to-diagnose faults.

[0188] Example 5

[0189] In the fault type identification stage, this embodiment uses long short-term memory (LSTM), independent recurrent neural network (IndRNN), long short-term memory-full convolutional network (LSTM-FCN), attention long short-term memory-full convolutional network (ALSTM-FCN) and single-level dynamic independent recurrent neural network (single DIndRNN) to compare with the method in Example 2. The diagnostic accuracy of 20 faults in the Tennessee-Eastman process is shown in Table 3. Fault 15 is a sticking fault of the condenser cooling water regulating valve. The diagnostic accuracy of the single-level classifier single DIndRNN is 48.3%, and the diagnostic accuracy of the method in Example 2 is 67.3%. By adding a second-level classifier, the diagnostic accuracy is improved by 19%. For fault 18, the diagnostic accuracy of the single DIndRNN method and the method in Example 2 are 53.7% and 92.4%, respectively, and the diagnostic accuracy is improved by 38.7%. Compared with the other 5 methods, the method in Example 2 achieves the best or tied for the best fault diagnosis performance in 15 fault conditions. The bar graph of the average diagnostic accuracy of the six methods is shown in Fig. 9As shown, it can be seen that the method in Example 2 has the highest average diagnostic accuracy of 90%. The implementation results show that the method involved in Example 2 can effectively improve the diagnostic accuracy of process faults by finely identifying difficult-to-diagnose faults through a two-level classifier based on a dynamic independent recurrent neural network, and provide accurate fault information to the maintenance team in a timely manner, so as to accelerate the fault recovery process and reduce fault losses.

[0190] Table 3 Diagnostic accuracy of 20 faults in the Tennessee-Eastman process (%)

[0191]

[0192] Example 6

[0193] In order to compare the fault feature extraction performance of the six methods, t-distributed random neighbor embedding (t-SNE) is used to visualize the feature distribution. Referring to the t-SNE scatter plot of the difficult-to-diagnose fault feature, it can be seen from the scatter plot of the process measurement data that the samples of faults 9, 10, 15 and 16 are mixed together and difficult to distinguish. From the scatter plot of the LSTM feature, it can be seen that the sample separation effect of faults 12 and 18 is improved. From the scatter plot of the IndRNN feature, it can be seen that the sample range of faults 10 and 16 is clearer, but still inseparable. From the scatter plot of the LSTM-FCN feature, it can be seen that the samples of faults 10 and 16 are completely separated, but the sample of fault 10 is clustered into 2 clusters. From the scatter plot of the ALSTM-FCN feature, it can be seen that the sample separation effect of faults 9 and 15 is improved. From the scatter plot of the single DIndRNN feature, it can be seen that some samples of faults 12 and 18 are mixed together. Using the method in Example 2 for feature extraction, the sample range of faults 10, 12, 16 and 18 is clearer and separable. Although some samples of faults 9 and 15 are mixed together, the discrimination ability of this method is better than that of the other 5 methods. The implementation results show that by extracting long-term trend features and short-term dynamic features, introducing variable and time attention mechanisms, integrating principal component analysis statistics and robust normalization, the method involved in Example 2 has a higher classification feature extraction ability.

[0194] Example 7

[0195] In order to evaluate the contribution of key modules in the single-stage dynamic independent recurrent neural network involved in Example 2, an ablation experiment was performed using the Tennessee-Eastman process data. The independent recurrent neural network (IndRNN), attention fully convolutional network (AFCN) and attention mechanisms in the dynamic independent recurrent neural network were removed respectively to analyze the impact of each module on the fault diagnosis performance; Dynamic IndRNN is a dynamic independent recurrent neural network, and w / o means removal (without). The ablation experiment results of the Tennessee-Eastman process are shown in Figure 2. Fig.10 As shown. It can be seen that when the method involved in Example 2 includes all modules, the average fault diagnosis accuracy is the highest, which is 88.52%. (1) Independent recurrent neural network: The dynamic independent recurrent neural network uses the independent recurrent neural network to extract long-term trend features. Removing this module causes the average fault diagnosis accuracy to decrease from 88.52% to 87.36%. The reduction in diagnostic accuracy verifies the effectiveness of the independent recurrent neural network in improving fault diagnosis capabilities. (2) Attention full convolutional network: The dynamic independent recurrent neural network uses the attention full convolutional network to extract short-term dynamic features. After removing this module, the average fault diagnosis accuracy decreases from 88.52% to 76.6%. The diagnostic accuracy decreases by 11.92%, reflecting the importance of the attention full convolutional network in extracting classification features. (3) Attention mechanism: The attention mechanism module uses the compression excitation submodule and the global time attention submodule to enhance key variables and key time points. The introduction of the attention mechanism module increases the average fault diagnosis accuracy from 80.36% to 88.52%, an increase of 8.16%. It can be seen that the attention mechanism module can effectively highlight the key information of multivariate time series. The ablation experiment results show that the three modules of independent recurrent neural network, attention fully convolutional network and attention mechanism can effectively process multivariate process fault data and play a key role in enhancing the classification feature extraction capability and improving fault diagnosis performance.

Claims

1. A process fault diagnosis method based on a two-stage cyclic network, characterized in that: The following steps are involved: S1: Establish the first-level fault classifier: collect historical fault data and construct a training set Verification Set and the test set Ω t , training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting; Dynamic independent recurrent neural network uses independent recurrent neural network to extract trend characteristics of fault process, uses attention full convolution network to extract dynamic characteristics of fault process, and uses softmax layer to obtain fault type prediction value; S2: Fault grouping: checksum The data is input into the first-level dynamic independent recurrent neural network to obtain the fault type prediction value, calculate the confusion rate, and classify all fault types into easy-to-diagnose faults and difficult-to-diagnose faults; specifically, the following steps are included: S201: Verify the set The data of is input into the first-level fault classifier to obtain the fault type prediction value, and the predicted fault type is compared with the actual fault type to calculate the confusion rate CR of the p-th type of fault p , where FN p Indicates the number of false negative samples, FP p Indicates the number of false positive samples, TP p Indicates the number of true samples; S3: Establish the second-level fault classifier: collect normal operation data, establish the principal component analysis model and T 2 , SPE statistics; collect training data of difficult-to-diagnose faults and calculate T 2 and SPE statistics, establish augmented vectors, perform robust normalization, and train the second-level dynamic independent recurrent neural network; specifically, the following steps are included: S301: Construct matrix X using normal operation data nor , establish the principal component model; principal component analysis aims to find The projection vector v of the image is obtained; the projection matrix V and the standard deviation matrix ∑ are calculated by singular value decomposition; a largest standard deviations are selected to construct the standard deviation matrix ∑ a , construct the principal element space projection matrix P using the projection vectors corresponding to these standard deviations; S302: For the normalized measurement vector x t , calculate T 2 and the SPE statistic, in represents the identity matrix; S303: Using T 2 And SPE statistics to augment the measurement vector and construct the vector Pair Vector Perform robust normalization; use express The e-th variable is calculated as its standardized value in and represent the median and interquartile range of the e-th variable, respectively; S304: Constructing vectors using data from difficult-to-diagnose faults and time series window Will as input and the fault type as output; S4: Identify fault type: collect real-time process data, perform standardization processing, input it into the first-level dynamic independent recurrent neural network, and obtain the fault type prediction value; when the predicted type belongs to a difficult-to-diagnose fault, construct an augmented vector, perform robust standardization processing, and input it into the second-level dynamic independent recurrent neural network to obtain the final fault type prediction value.

2. The process fault diagnosis method based on a two-stage cyclic network according to claim 1 is characterized in that: The step S1 specifically includes the following steps: S101: Collect historical fault data and construct training sets Verification Set and the test set Ω t ; Training set Used to train the first level dynamic independent recurrent neural network, verification set Used to adjust network parameters to avoid overfitting; S102: Constructing measurement vector using process measurement data Where t represents the sampling time, M represents the number of variables; for the measurement vector x′ t Standardization is performed so that the mean and variance of each variable are distributed to 0 and 1, and a standardized vector x is obtained. t =[x t,1 ,x t,2 ,…,x t,M ] T , construct a time series window Where L represents the time series window length; S103: The first feature extraction subprocess includes an independent recurrent neural network and a discard layer; the matrix X t Input independent recurrent neural network; in the first layer of the network, the output of the jth neuron corresponding to time t is Where j = 1, ..., N, N is the number of hidden layer neurons, σ(·) represents the rectified linear unit activation function, Represents the input weight vector, coefficient represents the cycle weight, represents the bias term; in the i-th layer of the network, the output of the j-th neuron corresponding to time t is Where i = 2, ..., K, K is the number of layers of independent recurrent neural network, Represents the input weight vector, the output vector of the (i-1)th layer is the input of the i-th layer of neurons; through deep feature mining, the output vector of the last layer of the independent recurrent neural network is obtained S104: In order to avoid overfitting of the model, the vector Input the discard layer to get trend features where the vector Element r i ~B(p1), B represents Bernoulli distribution, p1 represents r i =1, the operator ⊙ represents the Hadamard product; S105: The second feature extraction subprocess is an attention full convolutional network, including an ACBRD module and a global average pooling layer; the ACBRD module includes an attention module, a one-dimensional temporal convolution layer, a batch normalization layer, a rectified linear unit activation layer, and a discard layer; the attention module includes a compression excitation submodule and a global temporal attention submodule; the matrix X t Input compression excitation submodule, calculate the mean vector x m , and the weight vector w v , w v =τ(W s2 σ(W s1 x m )) in and is the parameter matrix to be learned, ρ represents the dimensionality reduction rate; using the weight vector w v Weight the input data to get the output matrix S t , S t =X t diag(w v ) Where diag(·) represents the diagonal matrix constructor; S106: compress the output matrix S of the excitation submodule t Enter the global time attention submodule and calculate the weight vector w t , in and is the parameter matrix to be learned; using the weight vector w t Weight the data to get the output matrix S107: The weighted time series window The output of the first ACBRD module is obtained by sequentially inputting the one-dimensional temporal convolution layer, batch normalization layer, rectified linear unit activation layer and discard layer for processing. In order to extract deep features, Q ACBRD modules are connected in series to the time series window X t Process and output features The characteristics Input the global average pooling layer for processing to obtain short-term dynamic features S108: Trend characteristics and dynamic features Concatenate and input into the softmax layer for processing to obtain the fault type prediction value y t .

3. The process fault diagnosis method based on a two-stage cyclic network according to claim 1 is characterized in that: The step S2 also includes the following steps: p Make a judgment, when CR p When it exceeds the control limit, the p-th fault is considered to be a difficult-to-diagnose fault; otherwise, the p-th fault is considered to be an easy-to-diagnose fault; Establishing a set of easily diagnosable faults and difficult to diagnose fault sets 4. The process fault diagnosis method based on a two-stage cyclic network according to claim 1 is characterized in that: The step S3 also includes using the training set Establish a new training set based on the difficult-to-diagnose fault samples in Using the validation set A new verification set is established based on the difficult-to-diagnose fault samples in Using the training set Train the second-level dynamic independent recurrent neural network, using the validation set Adjust network parameters.

5. The process fault diagnosis method based on a two-stage cyclic network according to claim 1 is characterized in that: The step S4 specifically includes the following steps: S401: Collect real-time measurement data, perform standardization processing, and construct a time series window X t ; S402: X t Independent recurrent neural network input to the first level classifier to calculate trend features X t Input the attention full convolutional network of the first-level classifier to calculate dynamic features S403: Features and Concatenate and input into the softmax layer to obtain the fault type prediction value y t ; S404: Yes t Make a judgment, if y t ∈E, the fault to be diagnosed is considered to be an easy-to-diagnose fault, and the fault type prediction value y is output t ; Otherwise, the fault to be diagnosed is considered to be a difficult-to-diagnose fault, and steps S405-S406 are executed for further diagnosis; S405: Calculate T 2 and SPE statistics, construct augmented vector Perform robust standardization and construct time series windows S406: Input the dynamic independent recurrent neural network of the second-level classifier to obtain the final fault type prediction value y′ t .

6. The process fault diagnosis method based on a two-stage cyclic network according to claim 3 is characterized in that: In step S202, CR p The control limit is 5%.

Images