Fault monitoring method and system based on RNN complementary double residual generator

Abstract

The present invention relates to a fault monitoring method and system based on a complementary double residual generator of RNN. The monitoring method comprises the following steps: inputting the fault detection process into a front residual generator; obtaining a residual vector through the front residual generator; dividing the residual into a Gaussian residual part and a non-Gaussian residual part through a Gaussianity checker; correcting the non-Gaussian residual part into a Gaussian residual through a back residual generator; inputting the aggregated Gaussian residual into a result diagnostic device, and 2 Statistics are used to diagnose faults. The monitoring system includes a process input module, a front Gaussian residual generation module, a back Gaussian residual generation module and a result output module. The method provided by the present invention can output Gaussian residuals in the case of unknown noise, and finally can be used to diagnose faults through T 2 Statistics are used to diagnose faults, which improves the fault detection rate while ensuring a low fault alarm rate.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault monitoring method and system of a complementary double residual generator based on RNN. Background Art

[0002] A recurrent neural network (RNN) is a type of recursive neural network that takes sequence data as input, performs recursion in the direction of sequence evolution, and all nodes (recurrent units) are connected in a chain.

[0003] As a deep neural network with a special structure, the recurrent neural network (RNN) associates the output of the current moment of the sequence with the previous information, similar to how people's existing cognition is influenced by past knowledge, experience and memory. It effectively extracts temporal and spatial features from the fault signal's temporal correlation and deep features, further improving the accuracy and reliability of intelligent fault diagnosis.

[0004] The problems and defects of the prior art are:

[0005] 1. Most methods assume that the noise is known, but the generated residuals do not necessarily conform to the Gaussian distribution. Does the Gaussian distribution have any effect on T? 2 Statistics and determining corresponding thresholds are very important.

[0006] 2. Using T based on non-Gaussian residuals 2 When using statistics for fault detection, even with kernel density estimation, some faults still cannot be detected.

[0007] 3. RNN-based framework monitoring Due to the complex internal structure and activation function, it is difficult to generate Gaussian residuals during minimum regression error learning.

[0008] 4. The RNN-based method can improve the fault detection rate, but its fault alarm rate is also higher than other data-driven process monitoring methods. Summary of the invention

[0009] In view of some or all of the problems in the prior art, the present invention provides a fault monitoring method for a complementary double residual generator based on RNN, the method comprising the following steps:

[0010] The process of fault detection is input into the front residual generator;

[0011] In the front residual generator, the process output prediction value is obtained through the RNN network, and the residual vector is obtained by comparing it with the true value;

[0012] Inputting the residual vector into a Gaussianity checker, and dividing the residual into a Gaussian residual part and a non-Gaussian residual part by the Gaussianity checker;

[0013] Retaining the Gaussian residual part as the previous Gaussian residual;

[0014] Extracting the corresponding process input index of the non-Gaussian residual part, inputting the index and the non-Gaussian residual part into a residual generator, and correcting the non-Gaussian residual part into a post-Gaussian residual through the post-residual generator;

[0015] Aggregating the front Gaussian residual and the back Gaussian residual to obtain a total residual;

[0016] The total residual is input into the result diagnostics unit, and the 2 Statistics for fault diagnosis.

[0017] Furthermore, the residual r generated by the previous residual generator pre It is expressed as,

[0018]

[0019] Among them, y pre is the process output of the previous residual generator, is the real process output;

[0020] The loss function L used to train the front residual generator pre (y) is defined as the sum of the generated residual vectors,

[0021]

[0022] Where K B is the batch size, K T is the length of a batch of time series data.

[0023] Furthermore, the Gaussianity checker sets a Gaussian threshold and uses a residual Gaussian index to determine Gaussianity; the Gaussian index G(z) used to calculate the residual is expressed as,

[0024] G(z)≈||E{H(z)}-E{H(z nor )}|| 2

[0025] Among them, z nor is a vector of zero-mean unit-variance Gaussian variables, and the negative entropy H(·) of the residual vector is expressed as,

[0026]

[0027] Among them, 1≤h1≤2;

[0028] The previous residual r pre Contains Gaussian part and the non-Gaussian part The previous residual r pre It can be expressed as,

[0029]

[0030] Gaussian part is the former Gaussian residual;

[0031] The Gaussianity of the residual is measured by setting the threshold Gth.

[0032]

[0033]

[0034] in, represents the Gaussian residual, represents a non-Gaussian residual.

[0035] Furthermore, the residual r generated by the post-residual generator pOS It is expressed as,

[0036]

[0037] Among them, y pos is the process output of the post-residual generator, is the real process output;

[0038] r pos is the post-Gaussian residual Non-Gaussian part of the front residual generator use Correction;

[0039] The loss function L used to train the residual generator pos (y) by the residual Gaussian index get,

[0040]

[0041] Where K B is the batch size.

[0042] Furthermore, the total residual r represents,

[0043]

[0044] in, is the pre-Gaussian residual, is the post-Gaussian residual.

[0045] Furthermore, the T 2 Statistics It is expressed as,

[0046]

[0047] Where r is the total residual, ∑ r is the covariance matrix of the total residual;

[0048] The threshold J used for judgment th It is expressed as,

[0049]

[0050] Among them, m and Nm are degrees of freedom, α is the significance level, and F 1-α (m, Nm) is the F distribution with degrees of freedom m and Nm respectively;

[0051] If T 2 Statistics Greater than or equal to threshold J th , then it is considered that there is a fault in the process, that is, the standard for fault diagnosis is,

[0052]

[0053] If T 2 Statistics Less than the threshold J th , then it is considered that there is no fault in the process.

[0054] The present invention also provides a fault monitoring system for a complementary double residual generator based on RNN, the system comprising the following modules:

[0055] A process input module, used for inputting the process of fault detection into the front residual generator;

[0056] A pre-Gaussian residual generation module is used to generate pre-Gaussian residuals;

[0057] A post-Gaussian residual generation module is used to generate a post-Gaussian residual;

[0058] The result output module is used to summarize the front Gaussian residual and the back Gaussian residual and pass T 2 Statistics for fault diagnosis.

[0059] Furthermore, the front Gaussian residual generation module includes a front residual generator and a Gaussianity checker, and the rear Gaussian residual generation module includes a rear residual generator.

[0060] The present invention also provides a computer system, characterized in that it comprises:

[0061] a processor configured to execute machine-readable instructions; and

[0062] The memory stores machine-readable instructions, which, when executed by the processor, perform the steps of the diagnostic method.

[0063] The present invention also provides a computer-readable storage medium, characterized in that machine-readable instructions are stored thereon, and the machine-readable instructions execute the steps of the diagnostic method when executed by a processor.

[0064] Compared with the prior art, the present invention has the following advantages:

[0065] 1. A new fault monitoring method is provided by integrating complementary dual residual generator and negative entropy loss function, which has never appeared before;

[0066] 2. Even when the noise statistics of the dynamic process are unknown, the generated residuals satisfy the Gaussian distribution; the Gaussian distribution result makes the dynamic process simply through T 2 Statistics to monitor, and the threshold can be easily determined by F distribution without kernel density estimation;

[0067] 3. The proposed fault diagnosis method can replace RNN with other network structures, such as LSTM-based and GRU-based dual residual generators. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] To further illustrate the above and other advantages and features of various embodiments of the present invention, a more specific description of various embodiments of the present invention will be presented with reference to the accompanying drawings. It will be understood that these drawings only depict typical embodiments of the present invention and are therefore not to be considered as limiting the scope thereof. In the accompanying drawings, for clarity, identical or corresponding parts will be represented by identical or similar reference numerals.

[0069] Figure 1 A schematic flow chart of a fault monitoring method of a complementary double residual generator based on an RNN according to an embodiment of the present invention;

[0070] Figure 2 A schematic diagram of a network structure of a residual generator based on RNN according to an embodiment of the present invention;

[0071] Figure 3 FIG. 1 is a schematic diagram of a fault monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In the following description, the present invention is described with reference to various embodiments. However, those skilled in the art will recognize that various embodiments may be implemented without one or more specific details or with other replacement and / or additional methods or components. In other cases, well-known structures or operations are not shown or described in detail to avoid obscuring the inventive point of the present invention. Similarly, for the purpose of explanation, specific numbers and configurations are set forth to provide a comprehensive understanding of embodiments of the present invention. However, the present invention is not limited to these specific details.

[0073] In this specification, reference to "one embodiment" or "the embodiment" means that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least one embodiment of the present invention. The phrase "in one embodiment" appearing in various places in this specification does not necessarily all refer to the same embodiment.

[0074] It should be noted that the embodiments of the present invention describe the method steps in a specific order, but this is only for the purpose of illustrating the specific embodiment, rather than limiting the order of the steps. On the contrary, in different embodiments of the present invention, the order of the steps can be adjusted according to actual needs.

[0075] In the present invention, each module of the system according to the present invention can be implemented using software, hardware, firmware or a combination thereof. When the module is implemented using software, the function of the module can be implemented by a computer program flow, for example, the module can be implemented by a code segment (such as a code segment of a language such as C, C++) stored in a storage device (such as a hard disk, a memory, etc.), wherein the corresponding function of the module can be implemented when the code segment is executed by a processor. When the module is implemented using hardware, the function of the module can be implemented by setting a corresponding hardware structure, for example, the function of the module can be implemented by hardware programming a programmable device such as a field programmable gate array (FPGA), or the function of the module can be implemented by designing an application-specific integrated circuit (ASIC) including electronic devices such as a plurality of transistors, resistors and capacitors. When the module is implemented using firmware, the function of the module can be written into a read-only memory such as an EPROM or EEPROM of the device in the form of a program code, and the corresponding function of the module can be implemented when the program code is executed by a processor. In addition, certain functions of the module may need to be implemented by separate hardware or by collaboration with the hardware, for example, the detection function is implemented by corresponding sensors (such as proximity sensors, acceleration sensors, gyroscopes, etc.), the signal transmission function is implemented by corresponding communication devices (such as Bluetooth devices, infrared communication devices, baseband communication devices, Wi-Fi communication devices, etc.), the output function is implemented by corresponding output devices (such as displays, speakers, etc.), and so on.

[0076] The technical solutions in the embodiments of the present invention are described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Figure 1 The flowchart of the fault monitoring method of the complementary double residual generator based on RNN according to an embodiment of the present invention is shown in the figure. As shown in the figure, the diagnostic method includes the following steps:

[0077] First, the fault detection process is input into the front residual generator and executed through the process input module. Process input refers to the input of the collected process data, such as the data collected by the system when running the process, or other process data. Figure 1 As shown, the data indexes of the process input are 1, 2, ..., p, where p is a positive integer greater than 2, and the size of p depends on the data size of the process input.

[0078] Next, in the front residual generator, the process output prediction value is obtained through the RNN network, and compared with the true value to obtain the residual vector. Figure 2 FIG. 1 is a schematic diagram of a network structure of a residual generator based on RNN according to an embodiment of the present invention. Figure 2 As shown, the data index of the process input is 1, 2, ..., n, where n is a positive integer greater than 2, and the size of n depends on the size of the process input data. Figure 2 As shown, the process output prediction value is obtained through the RNN network, and the residual vector is obtained by comparing it with the true value. The residual r generated by the previous residual generator pre It is expressed as,

[0079]

[0080] Among them, y pre is the process output of the previous residual generator, is the actual process output. Figure 1 As shown, the residual r generated by the previous residual generator pre The index is 1, 2, ..., q, where q is a positive integer greater than 2, and the size of q depends on the data size of the residual generated by the previous residual generator.

[0081] The loss function L used to train the front residual generator pre (y) is defined as the sum of the generated residual vectors,

[0082]

[0083] Where K B is the batch size, which indicates the data size of each batch, that is, how much data is selected as a batch; K T is the length of a batch of time series data, each sequence point is k, and the total length is K T .

[0084] Next, the residual vector is input into the Gaussianity checker, and the Gaussianity checker is used to divide the residual into a Gaussian residual part and a non-Gaussian residual part. The pre-residual generator captures the underlying dynamics and nonlinear relationships, and the generated residual is sent to the Gaussianity checker to obtain Gaussian residuals and non-Gaussian residuals. Negative entropy is a quantitative measure that can be used to estimate the Gaussianity of a random variable. The closer the negative entropy is to 0, the closer the distribution is to a Gaussian distribution. The Gaussianity checker sets a Gaussian threshold and uses the residual Gaussian indicator to judge Gaussianity; the Gaussian indicator G(Z) used to calculate the residual is expressed as,

[0085] G(z)≈||E{H(z)}-E{H(z nor )}|| 2

[0086] Among them, z nor is a vector of zero-mean unit-variance Gaussian variables, and the negative entropy H(·) of the residual vector is expressed as,

[0087]

[0088] Among them, 1≤h1≤2;

[0089] The previous residual r pre Contains Gaussian part and the non-Gaussian part The previous residual r pre It can be expressed as,

[0090]

[0091] The Gaussianity of the residual is measured by setting the threshold Gth.

[0092]

[0093]

[0094] in, represents the Gaussian residual, represents the non-Gaussian residual. Figure 1 As shown, the index of the Gaussian residual is 1, ..., m, where m is a positive integer less than q and greater than 1. Figure 1 As shown, the indexes of the non-Gaussian residuals are m+1, ..., q.

[0095] Next, retain the Gaussian residual part as the pre-Gaussian residual, that is, the Gaussian part is the pre-Gaussian residual.

[0096] Next, the process input index corresponding to the non-Gaussian residual part is extracted, and the index and the non-Gaussian residual part are input into the post-residual generator, and the non-Gaussian residual part is corrected to the post-Gaussian residual by the post-residual generator. The post-residual generator is trained by the negative entropy loss function so that the output residual is a Gaussian residual, which replaces the non-Gaussian residual part of the front residual generator. The residual r generated by the post-residual generator pos It is expressed as,

[0097]

[0098] Among them, y pos is the process output of the post-residual generator, is the actual process output.

[0099] r pos is the post-Gaussian residual Non-Gaussian part of the front residual generator use Correction. Figure 1 As shown, the index of the posterior residual is the index of the non-Gaussian residual, that is, m+1, ..., q.

[0100] The loss function L used to train the residual generator pos (y) by the residual Gaussian index get,

[0101]

[0102] Where K B is the batch size.

[0103] Next, the front Gaussian residual and the back Gaussian residual are summed up to obtain the total residual. The total residual r represents,

[0104]

[0105] in, is the pre-Gaussian residual, is the post-Gaussian residual. Figure 1 As shown, the index of the total residual r is 1, 2, ..., q, where q is a positive integer greater than 2, and the size of q depends on the data size of the residual generated by the previous residual generator.

[0106] Due to the existence of the double residual generator, Gaussian residuals can still be output even when the noise statistics are unknown.

[0107] Finally, the total residual is input into the result diagnostics unit through T 2 Statistics are used to diagnose faults. Producing Gaussian residuals means that the dynamic process can be simply explained by T 2Statistics can be used to monitor, and the threshold can be determined by F distribution without kernel density estimation.

[0108] In the present invention, the residual that conforms to Gaussian distribution is referred to as Gaussian residual, and the other is non-Gaussian residual. Gaussian distribution is also called normal distribution. If the random variable X obeys a normal distribution with a mathematical expectation of μ and a variance of σ, it is recorded as X~N(μ,σ).

[0109] For random variables X1, X2, ...X n ~N(0,1), let Then X is called a chi-square distribution with n degrees of freedom. 2 Recorded as

[0110] For random variables And X and Y are independent, then the F variable with degrees of freedom m and n respectively is, Its distribution is called F distribution with degrees of freedom m and n respectively, denoted by F~F m，n .

[0111] Hotelling T 2 Statistics of the distribution It is expressed as,

[0112]

[0113] Where r is the total residual, ∑ r is the covariance matrix of the total residual;

[0114] The threshold J used for judgment th It is expressed as,

[0115]

[0116] Among them, m and Nm are degrees of freedom, α is the significance level, and F 1-α (m, Nm) is the F distribution with m and Nm degrees of freedom respectively.

[0117] If T 2 Statistics Greater than or equal to threshold J th , then it is considered that there is a fault in the process, that is, the standard for fault diagnosis is,

[0118]

[0119] If T 2 Statistics Less than the threshold J th , then it is considered that there is no fault in the process.

[0120] The following example illustrates the impact of the fault monitoring method based on the RNN complementary double residual generator provided by the present invention on fault diagnosis.

[0121] The method proposed in the present invention can replace RNN with other network structures, such as being generalized to LSTM-based and GRU-based dual residual generators. Long Short-Term Memory Network LSTM (Long Short-Term Memory) is a time recurrent neural network specially designed to solve the long-term dependency problem existing in general RNN. Compared with ordinary RNN, LSTM can perform better in longer sequences. Gated Recurrent Unit GRU (Gate Recurrent Unit) is a type of RNN. Like LSTM, it is also proposed to solve problems such as gradients in long-term memory and back propagation. Compared with LSTM, using GRU can achieve comparable results and is easier to train. The method of the complementary dual residual generator based on RNN of the present invention is extended to LSTM-based and GRU-based methods, respectively represented as D-RNN, D-LSTM, and D-GRU, where D is the abbreviation of dual, referring to the dual residual generator.

[0122] Taking the Tennessee-Eastman process (TE process) data set as an example, the TE process data set contains 1 normal data set and 21 fault data sets. Each fault data set contains 960 samples, and a fault is introduced at the 161st sample. 80% of the fault-free data sets are used for model training, and the rest are used for model verification. Therefore, the number of samples used for training, verification, and testing are 768, 192, and 20160, respectively. Due to limited samples, only two hidden layers of 50 neurons are used, and the dropout technology with a probability of 0.2 is used to avoid overfitting. Dropout refers to the temporary discarding of neural network units from the network according to a certain probability during the training process of the neural network. The probability of 0.2 here means that each dropout retains 20% of the neural network units and no longer passes them back. Overfitting refers to the fact that the machine learning model or deep learning model performs too well in the training samples, resulting in poor performance in the verification data set and the test data set. Overfitting is generally caused by excessive feature extraction, so using dropout to reduce neural network units can prevent overfitting to a certain extent. The learning rates of the front residual generator and the back residual generator are set to 0.1 and 0.01 respectively. The hyperparameters are optimized by back propagation, and the loss function L pre and L posThrough Adam optimization. The learning rate is a tuning parameter in the optimization algorithm, which determines the step size in each iteration so that the loss function converges to the minimum value. If the learning rate is too large, the parameters to be optimized will fluctuate around the minimum value. If the learning rate is too small, the parameters to be optimized will converge slowly. Hyperparameters refer to tuning parameters in machine learning algorithms and need to be set manually. Adam optimization is a stochastic optimization method with adaptive momentum. It is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process. It can iteratively update the weights of the neural network based on the training data. Other parameters are set as follows: batch size KB = 50, sequence length KT = 50, total number of training epochs is set to 200, number of batches in each epoch is set to 19, Gaussian threshold Gth = 0.015, significance level α = 0.99.

[0123] Table 1 compares different methods on 21 fault datasets in the TE process. These methods include: DPCA (dynamic principal component analysis), DKPCA (dynamic kernel PCA), SAM (subspace-aided method), LQ-SAM (combined SAM with LQdecomposition-based), RNN, D-RNN, D-LSTM and D-GRU. D is the abbreviation of dual, which refers to the dual residual generator. D-RNN, D-LSTM and D-GRU are complementary dual residual generators based on RNN, LSTM and GRU. Table 1 compares the FDR and FAR differences of different methods on 21 fault datasets in the TE process. Among them, the best results of each fault dataset are marked in bold. FDR (fault detection rate) is the fault detection rate, and FAR (fault alarm rate / false alarm rate) is the fault false alarm rate (false alarm rate).

[0124] Table 1 Comparison of different methods on 21 fault datasets in TE process

[0125]

[0126] As shown in Table 1, the following conclusions can be drawn: the methods proposed in the present invention include D-RNN, and the extended D-LSTM and D-GRU. These methods can achieve better performance under the conditions of high FDR and acceptable FAR; the FDR value of D-RNN is 16.41%, 17.63%, 7.19%, 6.09% and 7.02% higher than that of DPCA, DKPCA, SAM, LQ-SAM and RNN, respectively; in most of the selected faults, the D-RNN proposed in the present invention is superior to the extended D-LSTM and D-GRU. RU, because D-LSTM and D-GRU introduce gating structures when modeling the TE process, which increases the number of network parameters (close to 3 times and 4 times that of D-RNN, respectively). However, the data set of the TE process has relatively few variables and data, which makes D-LSTM and D-GRU overfitted, resulting in inferior performance to D-RNN. Although the FAR value of D-RNN is higher than that of DPCA and DKPCA, it is acceptable under the condition of significance level α=0.99. Compared with RNN, the FAR value of D-RNN is much smaller.

[0127] The present invention proposes a new RNN-based fault monitoring method through a complementary double residual generator, which is used to process nonlinear dynamic processes with unknown noise statistics. The front residual generator captures the potential dynamics and nonlinear relationships, and sends the generated residuals to the Gaussianity checker. The Gaussianity checker sets a Gaussian threshold and uses the negative entropy of the residual vector to approximately judge the Gaussianity, thereby obtaining Gaussian residuals and non-Gaussian residual parts. The Gaussian residual part is retained, the index of the non-Gaussian residual part is extracted, and the non-Gaussian residual part is corrected to a Gaussian residual through the rear residual generator. The rear residual generator is trained using a negative entropy loss function. The RNN-based complementary double residual generator proposed in the present invention solves the problem that when the noise statistics are unknown, it is difficult for traditional RNNs to output residuals that satisfy the Gaussian distribution. The method provided by the present invention can output Gaussian residuals in the case of unknown noise, and finally can be used through T 2 Statistics are used to diagnose faults. The fault detection rate is improved while ensuring a low fault alarm rate. Finally, the invention can be extended to a dual residual generator based on LSTM and GRU.

[0128] The present invention also provides a fault monitoring system of a complementary double residual generator based on RNN. Figure 3 FIG. 1 is a schematic diagram of a fault monitoring system according to an embodiment of the present invention. Figure 3 As shown, the system includes the following modules:

[0129] A process input module, used for inputting the process of fault detection into the front residual generator;

[0130] A pre-Gaussian residual generation module, used for generating a pre-Gaussian residual; the pre-Gaussian residual generation module comprises a pre-residual generator and a Gaussianity checker;

[0131] A post-Gaussian residual generation module, used for generating a post-Gaussian residual; the post-Gaussian residual generation module includes a post-residual generator;

[0132] The result output module is used to summarize the front Gaussian residual and the back Gaussian residual and pass T 2 Statistics for fault diagnosis.

[0133] In one embodiment, the present invention further provides a computer system, including a memory and a processor, wherein the memory stores machine-readable instructions, and the processor implements the following processing steps when executing the machine-readable instructions: inputting the fault detection process into a front residual generator; in the front residual generator, obtaining a process output prediction value through an RNN network, and comparing it with the true value to obtain a residual vector; inputting the residual vector into a Gaussianity checker, and dividing the residual into a Gaussian residual part and a non-Gaussian residual part through the Gaussianity checker; retaining the Gaussian residual part as a front Gaussian residual; extracting the process input index corresponding to the non-Gaussian residual part, inputting the index and the non-Gaussian residual part into a rear residual generator, and correcting the non-Gaussian residual part into a rear Gaussian residual through the rear residual generator; summing up the front Gaussian residual and the rear Gaussian residual to obtain a total residual; inputting the total residual into a result diagnoser, and 2 Statistics for fault diagnosis.

[0134] It can be understood that in addition to the memory and processor mentioned above, the above-mentioned computer system also includes other software and hardware components not listed in this specification. The specific components can be determined according to the model of specific data processing equipment in different application scenarios, and this specification will not list them one by one in detail.

[0135] In one embodiment, a computer-readable storage medium is also provided, on which machine-readable instructions are stored. When the machine-readable instructions are executed by a processor, the following processing steps are implemented: inputting the fault detection process into a front residual generator; in the front residual generator, obtaining a process output prediction value through an RNN network, and comparing it with the true value to obtain a residual vector; inputting the residual vector into a Gaussianity checker, and dividing the residual into a Gaussian residual part and a non-Gaussian residual part through the Gaussianity checker; retaining the Gaussian residual part as a front Gaussian residual; extracting the process input index corresponding to the non-Gaussian residual part, inputting the index and the non-Gaussian residual part into a rear residual generator, and correcting the non-Gaussian residual part into a rear Gaussian residual through the rear residual generator; summing up the front Gaussian residual and the rear Gaussian residual to obtain a total residual; inputting the total residual into a result diagnostic device, and 2Statistics for fault diagnosis.

[0136] Although various embodiments of the present invention are described above, it should be understood that they are presented as examples only and not as limitations. It is obvious to those skilled in the relevant art that various combinations, modifications and changes can be made thereto without departing from the spirit and scope of the present invention. Therefore, the breadth and scope of the present invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should only be defined according to the attached claims and their equivalents.

Claims

1. A fault monitoring method based on RNN complementary double residual generator, characterized in that: The following steps are involved: The process of fault detection is input into the front residual generator; In the front residual generator, the process output prediction value is obtained through the RNN network, and the residual vector is obtained by comparing it with the true value; Inputting the residual vector into a Gaussianity checker, and dividing the residual into a Gaussian residual part and a non-Gaussian residual part by the Gaussianity checker; Retaining the Gaussian residual part as the previous Gaussian residual; Extract the corresponding process input index of the non-Gaussian residual part, input the index and the non-Gaussian residual part into the post-residual generator, correct the non-Gaussian residual part into the post-Gaussian residual by the post-residual generator, and train the post-residual generator through the negative entropy loss function so that the output residual is a Gaussian residual, replacing the non-Gaussian residual part of the front residual generator; Aggregating the front Gaussian residual and the back Gaussian residual to obtain a total residual; The total residual is input into the result diagnostics unit, and the 2 Statistics for fault diagnosis.

2. The method according to claim 1, characterized in that The residual r generated by the previous residual generator pre It is expressed as, Among them, y pre is the process output of the previous residual generator, is the real process output; The loss function L used to train the front residual generator pre (y) is defined as the sum of the generated residual vectors, Where K B is the batch size, K T is the length of a batch of time series data.

3. The method according to claim 1, characterized in that The Gaussianity checker sets a Gaussian threshold and uses a residual Gaussian index to determine Gaussianity; the Gaussian index G(z) used to calculate the residual is expressed as, G(z)≈||E{H(z)}-E{H(z nor )}|| 2 Among them, z nor is a vector of zero-mean unit-variance Gaussian variables, and the negative entropy H(·) of the residual vector is expressed as, Among them, 1≤h1≤2; The previous residual r pre Contains Gaussian part Non-Gaussian part The previous residual r pre It can be expressed as, Gaussian part is the former Gaussian residual; The Gaussianity of the residual is measured by setting the threshold Gth. in, represents the Gaussian residual, represents a non-Gaussian residual.

4. The method according to claim 1, characterized in that: The residual r generated by the post-residual generator pos It is expressed as, Among them, y pos is the process output of the post-residual generator, is the real process output; r pos is the post-Gaussian residual Non-Gaussian part of the front residual generator use Correction; The loss function L used to train the residual generator pos (y) by the residual Gaussian index get, Where K B is the batch size.

5. The method according to claim 1, characterized in that The total residual r express, in, is the pre-Gaussian residual, is the post-Gaussian residual.

6. The method according to claim 1, characterized in that The T 2 Statistics It is expressed as, Where r is the total residual, ∑ r is the covariance matrix of the total residual; The threshold J used for judgment th It is expressed as, Among them, m and Nm are degrees of freedom, α is the significance level, and F 1-α (m, Nm) is the F distribution with degrees of freedom m and Nm respectively; If T 2 Statistics Greater than or equal to threshold J th , then it is considered that there is a fault in the process, that is, the standard for fault diagnosis is, If T 2 Statistics Less than the threshold J th , then it is considered that there is no fault in the process.

7. A system for the fault monitoring method of the RNN-based complementary double residual generator according to any one of claims 1 to 6, characterized in that: Includes the following modules: A process input module, used for inputting the process of fault detection into the front residual generator; A pre-Gaussian residual generation module is used to generate pre-Gaussian residuals; A post-Gaussian residual generation module is used to generate a post-Gaussian residual; The result output module is used to summarize the front Gaussian residual and the back Gaussian residual and pass T 2 Statistics for fault diagnosis.

8. The system according to claim 7, characterized in that The front Gaussian residual generation module includes a front residual generator and a Gaussianity checker, and the back Gaussian residual generation module includes a back residual generator.

9. A computer system, characterized in that: include: a processor configured to execute machine-readable instructions; and A memory storing machine-readable instructions, wherein the machine-readable instructions, when executed by a processor, perform the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: Machine-readable instructions are stored thereon, and when the machine-readable instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are performed.

Images