A method and system for detecting faults in non-stationary industrial processes
By introducing differential cyclic autoencoder with differential input into long and short-term memory neural networks, the non-stationary characteristics of industrial data are extracted, and the problem of difficulty in extracting non-stationary characteristics in the existing methods is solved, early fault detection is realized, and production safety is improved and economic benefits.
Patent Information
- Application Number
- CN202211124746.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-15
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-09-15
AI Technical Summary
Existing deep learning-based process monitoring methods are difficult to effectively extract non-stationary features in industrial data, resulting in high false alarm rates or long fault detection delays.
A differential cyclic autoencoder method based on long and short time memory neural network is proposed. By introducing differential inputs, non-stationary features are extracted, fault detection models are constructed, and reconstruction errors are used as monitoring statistics to determine the monitoring statistics threshold to judge the fault.
Effectively extract the non-stationary characteristics of data, realize early and accurate fault identification, and improve production safety and economic benefits.
Smart Images

Figure CN115469619B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of fault detection, and in particular relates to a method and system for detecting faults in a non-stationary industrial process. Background Art
[0002] The rapid development of information technologies such as artificial intelligence and big data around the world has promoted the development of industrial manufacturing towards digitalization and intelligence. In industrial production, a large amount of production data can be collected, corresponding to the status of each device at every moment. The application of artificial intelligence technology to process and analyze industrial big data can improve production efficiency for factories and achieve high-quality economic development.
[0003] In industrial production, the stable operation of the process is an important prerequisite for ensuring the quality of factory products and production safety. The dense configuration of sensor networks in modern factories has greatly promoted the development of process monitoring technology. At the same time, distributed control systems are widely used in the collection and storage of industrial data, prompting data-driven fault detection methods to become a research hotspot. A fault is usually defined as at least one measured variable deviating from the normal operating condition. The distributed control system is also the main basis for fault detection in the factory. When the variable value in actual production exceeds the alarm value set by the operator in the system in advance, an alarm will be given. Then, when the variable exceeds the set value range, the fault has often occurred for a long time. When it is detected by the distributed control system, it may have caused irreversible impact on product quality or production. Since the historical production data contains a lot of process information, data-driven fault detection methods can reduce the impact of faults by extracting this information.
[0004] Commonly used multivariate statistical methods, such as principal component analysis and partial least squares, usually model the process by extracting the correlation between process variables. In principal component analysis, the data under normal conditions are projected into a principal space that maximizes the variance of the latent variables and a residual space, and fault detection can be achieved by constructing the corresponding process monitoring statistics in each space and their thresholds under normal conditions. However, nonlinear correlations cannot be extracted by linear transformation in principal component analysis. et al. first applied kernel principal component analysis to nonlinear feature extraction, in which the data is first mapped to a higher-dimensional space by a kernel function before being transformed by traditional principal component analysis. Later, Lee et al. applied kernel principal component analysis to nonlinear process monitoring, but this method still has difficulty in selecting appropriate kernels and their parameters to extract complex nonlinearities in the process.
[0005] Another disadvantage of principal component analysis is that it does not consider the extraction of dynamic information because the autocorrelation of the data is not taken into account. Ku et al. applied dynamic principal component analysis to extract autocorrelation by expanding the variables with their time-lagged measurements. However, dynamic principal component analysis has several obvious limitations, for example, the dimension of the augmented matrix will be high as the time lag increases, and the method still only considers variance rather than the dynamic information inherent in the data. Another option is the dynamic latent variable method proposed by Li et al. This method first performs autoregressive principal component analysis to obtain the latent variable with the largest autovariance, and then builds an internal model for the latent variable to extract the dynamic information inherent in the data. The dynamic latent variable method has attracted great attention in statistical process monitoring, and several variants have been proposed, such as dynamic internal partial least squares, dynamic internal principal component analysis, and autoregressive dynamic latent variables. Although the above latent variable methods solve some problems of feature extraction, these models are still not sufficient to extract the intrinsic characteristics of industrial data.
[0006] Recently, the application of deep learning to extract latent variables for process monitoring has become a research hotspot. Deep learning process monitoring methods are mainly based on autoencoders, which are special neural network structures that use the back-propagation algorithm to make the output value equal to the input value. Autoencoders have better performance in feature extraction because almost any type of nonlinearity can be fitted through the nonlinear activation function of the neural network. Jang et al. proposed an adversarial autoencoder model that can make the extracted latent variables obey a special distribution, thereby better representing the original data manifold. Yu and Zhao proposed a denoising autoencoder for feature extraction from noisy input data to achieve robust process monitoring. In order to extract the dynamic information of the data, Zhang and Qiu proposed a dynamic internal autoencoder to extract process dynamics by developing a vector autoregression model of the latent variables. Further research on deep learning-based process monitoring is to use recurrent neural networks to extract serial correlations. Cheng et al. proposed a variational recurrent autoencoder to handle process nonlinearity and dynamics through a model structure with gated recurrent units. Yu et al. proposed a fault detection method based on convolutional long short-term memory autoencoders, which can extract long-term temporal dependencies of data through forgetting gates.
[0007] However, existing deep learning-based process monitoring methods still cannot well consider non-stationary features in high-dimensional data. Non-stationarity can be understood as the statistical characteristics of data during the operation of industrial processes, such as mean, covariance, etc., which change over time due to various factors. Fault information can be hidden in these non-stationary trends, resulting in high false alarm rates or long fault detection delays. A study showed that most forget gates in LSTM prediction models are saturated, which means that the transition of memory states is almost time-invariant. Specifically, the features extracted from the information of previous signals are constant, corresponding to long-term stationary features, while short-term non-stationary information is difficult to obtain. Summary of the invention
[0008] In order to solve the above technical problems, the present invention proposes a technical solution of a non-stationary industrial process fault detection method to solve the above technical problems.
[0009] A first aspect of the present invention discloses a method for detecting faults in a non-stationary industrial process, the method comprising:
[0010] Step S1, obtaining historical data under normal working conditions from the factory historical database, and standardizing the historical data;
[0011] Step S2, introducing differential input into the recurrent neural network based on the long short-term memory neural network, constructing a differential recurrent autoencoder, and decoding the encoding layer output of the differential recurrent autoencoder with a decoder layer to obtain a fault detection model;
[0012] Step S3, dividing the standardized historical data into a training data set and a verification data set, and using the training data set and the verification data set to train the fault detection model;
[0013] Step S4: Use the trained fault detection model to calculate the reconstruction error of the verification data set as a monitoring statistic, and determine the monitoring statistic threshold;
[0014] Step S5: acquiring real-time data from sensors / instruments in the factory in real time, and performing standardization processing on the real-time data;
[0015] Step S6: input the standardized real-time data into the trained fault detection model to reconstruct the data and calculate the real-time monitoring statistics;
[0016] Step S7: compare the real-time monitoring statistic with the monitoring statistic threshold. If the real-time monitoring statistic exceeds the monitoring statistic threshold, it is determined to be a fault; otherwise, it is determined to be a normal operating condition.
[0017] According to the method of the first aspect of the present invention, in step S2, the method of introducing differential input into a recurrent neural network based on a long short-term memory neural network to construct a differential recurrent autoencoder includes:
[0018] The differential recurrent autoencoder includes four layers of long short-term memory neural networks. The difference between the hidden layer output of the first layer of long short-term memory neural network at the current moment and the hidden layer output of the first layer of long short-term memory neural network at the previous moment is used as the input of the second layer of long short-term memory neural network to extract non-stationary features. The non-stationary features are input into the third layer of long short-term memory neural network to extract the hidden state of the non-stationary features. The hidden state of the non-stationary features is input into the fourth layer of long short-term memory neural network to obtain the hidden features of the differential recurrent autoencoder.
[0019] According to the method of the first aspect of the present invention, in step S2, the hidden layer output of the first layer of the long short-term memory neural network is: the hidden layer output of the last layer of the first layer of the long short-term memory neural network.
[0020] According to the method of the first aspect of the present invention, in step S2, the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment is used as the input of the second layer of the long short-term memory neural network, and the method for extracting non-stationary features includes:
[0021] i t1 =σ(W i1 ·[h t -h t-1 , N t-1 ]+b i1 )
[0022] f t1 =σ(W f1 ·[h t -h t-1 , N t-1 ]+b f1 )
[0023] g t1 =tanh(W C1 ·[h t -h t-1 , N t-1 ]+b C1 )
[0024] N t =f t1 ·N t-1 +i t1 ·g t1
[0025] o t1 =σ(WO1 ·[h t -h t-1 , N t ]+b o1 )
[0026] D t =o t1 ·tanh(N t )
[0027] Among them, h t and h t-1 are the hidden layer output of the first layer of the LSTM neural network at the current moment and the hidden layer output of the first layer of the LSTM neural network at the previous moment, N t and N t-1 are the cell states of the layer at the current moment and the previous moment, respectively, t represents the extracted non-stationary features, σ and tanh are activation functions, W i1 and b i1 are the weights and biases of the first layer of the second layer of the long short-term memory neural network; W f1 and b F1 are the weights and biases of the second layer of the second layer of the long short-term memory neural network; W C1 and b C1 are the weights and biases of the third layer of the second layer of the long short-term memory neural network; W O1 and b O1 They are the weights and biases of the fourth layer of the second-layer long short-term memory neural network.
[0028] According to the method of the first aspect of the present invention, in step S3, the loss function for training the fault detection model is:
[0029] The mean square error between the reconstructed output and the actual value is used as the loss function of the fault detection model;
[0030] The reconstructed output is the output of the decoder layer.
[0031] According to the method of the first aspect of the present invention, in step S4, the monitoring statistic is a mean square error between the reconstructed output and the actual value.
[0032] According to the method of the first aspect of the present invention, in step S4, the monitoring statistic threshold is the mean square error monitoring statistic threshold under a 99% confidence interval; the mean square error monitoring statistic threshold under a 99% confidence interval is determined by a kernel density estimation method.
[0033] A second aspect of the present invention discloses a non-stationary industrial process fault detection system; the system comprises:
[0034] The first processing module is configured to obtain historical data under normal working conditions from a factory historical database and perform standardization processing on the historical data;
[0035] The second processing module is configured to introduce differential input into the recurrent neural network based on the long short-term memory neural network, construct a differential recurrent autoencoder, and decode the output of the encoding layer of the differential recurrent autoencoder with a decoder layer to obtain a fault detection model;
[0036] A third processing module is configured to divide the standardized historical data into a training data set and a verification data set, and use the training data set and the verification data set to train the fault detection model;
[0037] A fourth processing module is configured to calculate a reconstruction error of a validation data set using the trained fault detection model as a monitoring statistic, and determine a monitoring statistic threshold;
[0038] The fifth processing module is configured to obtain real-time data from sensors / instruments of the factory in real time and perform standardized processing on the real-time data;
[0039] A sixth processing module is configured to input the standardized real-time data into the trained fault detection model for data reconstruction and calculate real-time monitoring statistics;
[0040] The seventh processing module is configured to compare the real-time monitoring statistic with the monitoring statistic threshold, and if the real-time monitoring statistic exceeds the monitoring statistic threshold, determine it as a fault, otherwise determine it as a normal operating condition.
[0041] According to the system of the second aspect of the present invention, the second processing module is configured as follows: introducing differential input into the recurrent neural network based on the long short-term memory neural network to construct a differential recurrent autoencoder includes:
[0042] The differential recurrent autoencoder includes four layers of long short-term memory neural networks. The difference between the hidden layer output of the first layer of long short-term memory neural network at the current moment and the hidden layer output of the first layer of long short-term memory neural network at the previous moment is used as the input of the second layer of long short-term memory neural network to extract non-stationary features. The non-stationary features are input into the third layer of long short-term memory neural network to extract the hidden state of the non-stationary features. The hidden state of the non-stationary features is input into the fourth layer of long short-term memory neural network to obtain the hidden features of the differential recurrent autoencoder.
[0043] According to the system of the second aspect of the present invention, the second processing module is configured such that the hidden layer output of the first layer of the long short-term memory neural network is: the hidden layer output of the last layer of the first layer of the long short-term memory neural network.
[0044] According to the system of the second aspect of the present invention, the second processing module is configured to use the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment as the input of the second layer of the long short-term memory neural network, and the extracted non-stationary features include:
[0045] i t1 =σ(W i1 ·[h t -h t-1 , N t-1 ]+b i1 )
[0046] f t1 =σ(W f1 ·[h t -h t-1 , N t-1 ]+b f1 )
[0047] g t1 =tanh(W C1 ·[h t -h t-1 , N t-1 ]+b C1 )
[0048] N t =f t1 ·N t-1 +i t1 ·g t1
[0049] o t1 =σ(W O1 ·[h t -h t-1 , N t ]+b o1 )
[0050] D t =o t1 ·tanh(N t )
[0051] Among them, h t and h t-1 are the hidden layer output of the first layer of the LSTM neural network at the current moment and the hidden layer output of the first layer of the LSTM neural network at the previous moment, N t and N t-1 are the cell states of the layer at the current moment and the previous moment, respectively, t represents the extracted non-stationary features, σ and tanh are activation functions, W i1 and b i1are the weights and biases of the first layer of the second layer of the long short-term memory neural network; W f1 and b f1 are the weights and biases of the second layer of the second layer of the long short-term memory neural network; W C1 and b C1 are the weights and biases of the third layer of the second layer of the long short-term memory neural network; W O1 and b O1 They are the weights and biases of the fourth layer of the second-layer long short-term memory neural network.
[0052] According to the system of the second aspect of the present invention, the third processing module is configured to train the loss function of the fault detection model as follows:
[0053] The mean square error between the reconstructed output and the actual value is used as the loss function of the fault detection model;
[0054] The reconstructed output is the output of the decoder layer.
[0055] According to the system of the second aspect of the present invention, the fourth processing module is configured such that the monitoring statistic is a mean square error between the reconstructed output and the actual value.
[0056] According to the system of the second aspect of the present invention, the fourth processing module is configured such that the monitoring statistic threshold is the mean square error monitoring statistic threshold under a 99% confidence interval; the mean square error monitoring statistic threshold under a 99% confidence interval is determined by a kernel density estimation method.
[0057] The third aspect of the present invention discloses an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of any one of the non-stationary industrial process fault detection methods in the first aspect of the present disclosure are implemented.
[0058] The fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the non-stationary industrial process fault detection methods in the first aspect of the present disclosure.
[0059] The solution proposed in the present invention can extract the non-stationary characteristics of data, so as to effectively distinguish faults from normal non-stationary trends, realize early and accurate identification of process faults, and thus ensure the production safety and economic benefits of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0061] Figure 1 is a flow chart of a non-stationary industrial process fault detection method according to an embodiment of the present invention;
[0062] Figure 2 A logic flow chart of a non-stationary industrial process fault detection method according to an embodiment of the present invention;
[0063] Figure 3 A structural diagram of a differential cyclic autoencoder according to an embodiment of the present invention;
[0064] Figure 4 The single-input single-output process original data according to an embodiment of the present invention;
[0065] Figure 5a The visualization results of principal component analysis of normal samples and fault samples according to an embodiment of the present invention;
[0066] Figure 5b The visualization results of the kernel principal component analysis of normal samples and fault samples according to the embodiment of the present invention;
[0067] Figure 5c Visualization results of the automatic encoders for normal samples and fault samples according to an embodiment of the present invention;
[0068] Figure 5d The visualization results of the variational recurrent autoencoder for normal samples and fault samples according to an embodiment of the present invention;
[0069] Figure 5e The visualization results of the variational recurrent autoencoder for normal samples and fault samples according to an embodiment of the present invention;
[0070] Figure 5f The visualization result of the differential cyclic autoencoder of normal samples and fault samples according to an embodiment of the present invention;
[0071] Figure 6a is a fault detection result of a principal component analysis of a single-input single-output non-stationary process according to an embodiment of the present invention;
[0072] Figure 6b is a fault detection result of a kernel principal component analysis of a single-input single-output non-stationary process according to an embodiment of the present invention;
[0073] Figure 6c is a fault detection result of an automatic encoder for a single-input single-output non-stationary process according to an embodiment of the present invention;
[0074] Figure 6d is a fault detection result of a variational cyclic autoencoder for a single-input single-output non-stationary process according to an embodiment of the present invention;
[0075] Figure 6e is a fault detection result of a variational cyclic autoencoder for a single-input single-output non-stationary process according to an embodiment of the present invention;
[0076] Figure 6f is a fault detection result of a differential cyclic autoencoder for a single-input single-output non-stationary process according to an embodiment of the present invention;
[0077] Figure 7a is a fault detection result of the principal component analysis T2 of the industrial continuous catalytic reforming heat exchange unit according to an embodiment of the present invention;
[0078] Figure 7b The fault detection result of the principal component analysis SPE of the industrial continuous catalytic reforming heat exchange unit according to the embodiment of the present invention;
[0079] Figure 7c is a fault detection result of the nuclear principal component analysis T2 of the industrial continuous catalytic reforming heat exchange unit according to an embodiment of the present invention;
[0080] Figure 7d The fault detection result of the core principal component analysis SPE of the industrial continuous catalytic reforming heat exchange unit according to the embodiment of the present invention;
[0081] Figure 7e is a fault detection result of a dynamic principal component analysis T2 of an industrial continuous catalytic reforming heat exchange unit according to an embodiment of the present invention;
[0082] Figure 7f The fault detection result of the dynamic principal component analysis SPE of the industrial continuous catalytic reforming heat exchange unit according to the embodiment of the present invention;
[0083] Figure 7g is a fault detection result of an automatic encoder of an industrial continuous catalytic reforming heat exchange unit according to an embodiment of the present invention;
[0084] Figure 7h is a fault detection result of a variational cycle automatic encoder of an industrial continuous catalytic reforming heat exchange unit according to an embodiment of the present invention;
[0085] Figure 7i is a fault detection result of a differential cycle autoencoder of an industrial continuous catalytic reforming heat exchange unit according to an embodiment of the present invention;
[0086] Figure 8 is a structural diagram of a non-stationary industrial process fault detection system according to an embodiment of the present invention;
[0087] Fig. 9 The figure is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0088] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0089] A first aspect of the present invention discloses a method for detecting faults in a non-stationary industrial process. Figure 1 FIG. 1 is a flow chart of a non-stationary industrial process fault detection method according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method includes:
[0090] Step S1, obtaining historical data under normal working conditions from the factory historical database, and standardizing the historical data;
[0091] Step S2, introducing differential input into the recurrent neural network based on the long short-term memory neural network, constructing a differential recurrent autoencoder, and decoding the encoding layer output of the differential recurrent autoencoder with a decoder layer to obtain a fault detection model;
[0092] Step S3, dividing the standardized historical data into a training data set and a verification data set, and using the training data set and the verification data set to train the fault detection model;
[0093] Step S4: Use the trained fault detection model to calculate the reconstruction error of the verification data set as a monitoring statistic, and determine the monitoring statistic threshold;
[0094] Step S5: acquiring real-time data from sensors / instruments in the factory in real time, and performing standardization processing on the real-time data;
[0095] Step S6: input the standardized real-time data into the trained fault detection model to reconstruct the data and calculate the real-time monitoring statistics;
[0096] Step S7: compare the real-time monitoring statistic with the monitoring statistic threshold. If the real-time monitoring statistic exceeds the monitoring statistic threshold, it is determined to be a fault; otherwise, it is determined to be a normal operating condition.
[0097] In step S1, historical data under normal working conditions are obtained from the factory historical database, and the historical data are standardized.
[0098] Specifically, historical data under normal working conditions are obtained from the factory history database, and the historical data are standardized to have a mean of 0 and a variance of 1.
[0099] In step S2, differential input is introduced into the recurrent neural network based on the long short-term memory neural network, a differential recurrent autoencoder is constructed, and the output of the encoding layer of the differential recurrent autoencoder is decoded by a decoder layer to obtain a fault detection model.
[0100] like Figure 3 As shown, in some embodiments, in step S2, the method of introducing differential input into a recurrent neural network based on a long short-term memory neural network to construct a differential recurrent autoencoder includes:
[0101] The differential recurrent autoencoder includes four layers of long short-term memory neural networks. The difference between the hidden layer output of the first layer of long short-term memory neural network at the current moment and the hidden layer output of the first layer of long short-term memory neural network at the previous moment is used as the input of the second layer of long short-term memory neural network to extract non-stationary features. The non-stationary features are input into the third layer of long short-term memory neural network to extract the hidden state of the non-stationary features. The hidden state of the non-stationary features is input into the fourth layer of long short-term memory neural network to obtain the hidden features of the differential recurrent autoencoder.
[0102] The hidden layer output of the first layer of long short-term memory neural network is: the hidden layer output of the last layer of the first layer of long short-term memory neural network.
[0103] The method of extracting non-stationary features by taking the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment as the input of the second layer of the long short-term memory neural network includes:
[0104] i t1 =σ(W i1 ·[h t -h t-1 , N t-1 ]+b i1 )
[0105] f t1 =σ(W f1 ·[h t-h t-1 , N t-1 ]+b f1 )
[0106] g t1 =tanh(W C1 ·[h t -h t-1 , N t-1 ]+b C1 )
[0107] N t =f t1 ·N t-1 +i t1 ·g t1
[0108] o t1 =σ(W O1 ·[h t -h t-1 , N t ]+b o1 )
[0109] D t =o t1 ·tanh(N t )
[0110] Among them, h t and h t-1 are the hidden layer output of the first layer of the LSTM neural network at the current moment and the hidden layer output of the first layer of the LSTM neural network at the previous moment, N t and N t-1 are the cell states of the layer at the current moment and the previous moment, respectively, t represents the extracted non-stationary features, σ and tanh are activation functions, W i1 and b i1 are the weights and biases of the first layer of the second layer of the long short-term memory neural network; W f1 and b f1 are the weights and biases of the second layer of the second layer of the long short-term memory neural network; W C1 and b C1 are the weights and biases of the third layer of the second layer of the long short-term memory neural network; W O1 and b O1 They are the weights and biases of the fourth layer of the second layer of the long short-term memory neural network. Specifically, as shown in Figure 3, the model introduces the idea of difference based on the commonly used recurrent neural network and long short-term memory neural network, so that it can capture the non-stationary features in the data while ensuring the nonlinearity and dynamics of the extraction process. The specific explanation of the model structure is as follows.
[0111] Neural networks have been widely used in feature extraction of industrial data in recent years due to their powerful nonlinear fitting capabilities. The mechanism of artificial neural networks is to process inputs at different time points separately, and most processes are dynamic, which means that the input at the current time is related to the previous input. For this reason, recurrent neural networks were proposed to extract the serial correlation of the process by connecting nodes between hidden layers. However, traditional recurrent neural networks only have a single transmission state and perform poorly in feature extraction on long-term time series. Long short-term memory neural networks, as a special type of recurrent neural network, have been proposed and widely used to process long-term time dependencies. It uses a cell state C t The idea of memory transfer is introduced into the network structure. The transmission of information is controlled by the gate structure and the calculation is as follows:
[0112] i t =σ(W i ·[h t-1 , x t ]+b i ) (1)
[0113] f t =σ(W f ·[h t-1 , x t ]+b f ) (2)
[0114] g t =tanh(W C ·[h t-1 , x t ]+b C ) (3)
[0115] o t =σ(W O ·[h t-1 , x t ]+b o ) (4)
[0116] C t =f t ·C t-1 +i t ·g t (5)
[0117] h t =o t tanh(C t ) (6) where i t , f t , o t They are input gate, forget gate and output gate, x t , C t , g t,h t is the current input, cell state, candidate state, and hidden state. t-1 , C t-1 is the hidden state and cell state of the previous moment, σ and tanh are activation functions, W i , W f , W C , W O is the weight of the neural network, b i , b f , b C , b O is the bias. Normally, the cell state C t Changes slowly, most of which are transferred to the cell state of the previous moment C t-1 Therefore, the information transfer between cell states is time-invariant, which means that LSTM neural networks can only transmit long-term steady-state information but cannot capture local non-stationary information. The Differential Recurrent Autoencoder (DRAE) will solve this problem to achieve better performance in fault detection in non-stationary processes.
[0118] The differential recurrent autoencoder includes four layers of long short-term memory neural networks. The difference between the hidden layer output of the first layer of long short-term memory neural network at the current moment and the hidden layer output of the first layer of long short-term memory neural network at the previous moment is used as the input of the second layer of long short-term memory neural network to extract non-stationary features. The non-stationary features are input into the third layer of long short-term memory neural network to extract the hidden state of the non-stationary features. The hidden state of the non-stationary features is input into the fourth layer of long short-term memory neural network to obtain the hidden features of the differential recurrent autoencoder.
[0119] The hidden layer output of the first layer of long short-term memory neural network is: the hidden layer output of the last layer of the first layer of long short-term memory neural network.
[0120] The method of extracting non-stationary features by taking the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment as the input of the second layer of the long short-term memory neural network includes:
[0121] i t1 =σ(W i1 ·[h t -h t-1 , N t-1 ]+b i1 ) (7)
[0122] f t1 =σ(W f1 ·[h t -h t-1 , N t-1 ]+bf1 ) (8)
[0123] g t1 =tanh(W C1 ·[h t -h t-1 , N t-1 ]+b C1 ) (9)
[0124] N t =f t1 ·N t-1 +i t1 ·g t1 (1)
[0125] o t1 =σ(W O1 ·[h t -h t-1 , N t ]+b o1 ) (11)
[0126] D t =o t1 ·tanh(N t ) (12)
[0127] Among them, i t1 , f t1 , o tl They are input gate, forget gate and output gate, g tl is a candidate state, h t and h t-1 are the hidden layer output of the first layer of the LSTM neural network at the current moment and the hidden layer output of the first layer of the LSTM neural network at the previous moment, N t and N t-1 are the cell states of the layer at the current moment and the previous moment, respectively, t represents the extracted non-stationary features, σ and tanh are activation functions, W i1 , W f1 , W C1 , W O1 is the weight of the neural network, b i1 , b f1 , b C1 , b O1 is bias.
[0128] Then use another memory state S t-1 It is used to determine how much non-stationary information should be retained. It can be calculated by formula 13-18:
[0129] i t2 =σ(W i2 ·[Dt , C t-1 ]+b i2 ) (13)
[0130] f t2 =σ(W f2 ·[D t , C t-1 ]+b f2 ) (14)
[0131] g t2 =tanh(W C2 ·[D t , C′ t-1 ]+b C2 ) (15)
[0132] S t =f t2 ·S t-1 +i t2 ·g t2 (16)
[0133] o t2 =σ(W O2 ·[D t , C′ t-1 , S t ]+b o2 ) (17)
[0134] T t =o t2 ·tanh(S t ) (18)
[0135] where i t2 , f t2 , o t2 They are input gate, forget gate and output gate, g t2 is a candidate state, C' t-1 is the hidden state at the previous moment, T t is the hidden state output of the layer at the current moment, indicating the non-stationary features extracted by the memory unit, W i2 , W f2 , W C2 , W O2 is the weight of the neural network, b i2 , b f2 , b C2 , b O2 is the bias, S t is the memory state at the current moment, T t is the non-stationary feature that is retained.
[0136] Finally, another LSTM neural network layer is used to obtain the hidden features of the encoder of the DRAE method:
[0137] i t3 =σ(W i3 ·[H t-1 ,h t ]+b i3 ) (19)
[0138] g t3 =tanh(W C3 ·[H t-1 ,h t ]+b C3 ) (20)
[0139] C′ t =T t +i t3 ·g t3 (twenty one)
[0140] o t3 =σ(W O3 ·[H t-1 ,h t ]+b o3 ) (twenty two)
[0141] H t =o t3 tanh(C′ t ) (twenty three)
[0142] where i t3 , o t3 are the input gate and output gate respectively, g t3 is a candidate state, C' t is the cell state, H t and H t-1 are the hidden state of the previous moment and the hidden features of the DRAE encoder, W i3 , W C3 , W O3 is the weight of the neural network, b i3 , b C3 , b O3 is the bias. Therefore, based on the proposed structure, the nonlinear, dynamic and time-varying non-stationary features of the input data can be extracted simultaneously. In order to achieve the fault detection task, a decoder layer can be used to obtain the reconstruction of the input data based on the extracted hidden features:
[0143] x′ t = decoder(H t ) (twenty four)
[0144] where x' tTo reconstruct the output, decoder(·) represents the decoder.
[0145] The DRAE method can consider both the stationary information and the non-stationary information of industrial process data, and will not lose the useful information of the original data during the difference process.
[0146] In step S3, the standardized historical data is divided into a training data set and a verification data set, and the training data set and the verification data set are used to train the fault detection model.
[0147] In some embodiments, in step S3, the loss function for training the fault detection model is:
[0148] The mean square error between the reconstructed output and the actual value is used as the loss function of the fault detection model;
[0149] The reconstructed output is the output of the decoder layer.
[0150] Specifically, the reconstruction error, that is, x' t and the actual value x t The mean squared error between them is used as the loss function of the DRAE model.
[0151] In step S4, the trained fault detection model is used to calculate the reconstruction error of the verification data set as a monitoring statistic, and a monitoring statistic threshold is determined.
[0152] In some embodiments, in step S4, the monitoring statistic is a mean square error between the reconstructed output and the actual value.
[0153] The monitoring statistic threshold is a mean square error monitoring statistic threshold under a 99% confidence interval; the mean square error monitoring statistic threshold under a 99% confidence interval is determined by a kernel density estimation method.
[0154] Specifically, the statistic used to quantify the fault detection results is selected as the mean square prediction error (SPE) statistic in Equation 25, which is the reconstruction error between the input and output of the DRAE model.
[0155]
[0156] The statistical threshold at the 99% confidence interval is determined by the kernel density estimation method as follows:
[0157]
[0158]
[0159] Where MSE iis the mean square error calculated under normal conditions, K(·) is the kernel function, d is the window length, ρ(SPE) is the kernel density estimate, MSE is, and threshold is the statistical threshold at 99% confidence level.
[0160] In order to demonstrate the application effect of the present application in various non-stationary industrial processes, it is verified in a non-stationary single-input single-output process, the Tennessee Eastman benchmark chemical process and a continuous catalytic reforming heat exchange unit of a Chinese chemical company, and compared with some traditional methods and current advanced methods. In order to compare the fault detection performance between different methods, quantitative indicators commonly used in industry, fault detection rate (FDR) and false alarm rate (FAR) are applied in this invention. FDR represents the proportion of faults correctly detected, while FAR is the proportion of normal samples that are mistakenly identified as faults. Excellent fault detection models usually show the highest FDR based on low FAR. Although FDR is used in most related works to illustrate the fault detection ability of the model, the sensitivity of the model to early fault detection is more important. Especially for actual industrial processes, the key to process monitoring technology is to detect faults in the shortest time, rather than focusing on whether the model still alarms long after the fault occurs. Therefore, the third indicator, fault detection delay, is also used to evaluate the fault detection performance of the proposed model.
[0161] Example 1: Single-input single-output non-stationary process
[0162] x i =sin(0.01i)+α i (28)
[0163] u i+1 =0.1u i +2x i (29)
[0164] y i =u i 2 +β i +f i (30)
[0165]
[0166] where u i is the state variable, x i is a non-stationary input signal, y i is the nonlinear output signal, α i and β i Is in accordance with N(0,0.1 2 )’s measurement noise, f iis a fault, introduced between the 1000th sample and the 1280th sample. The simulation data is as follows Figure 4 As shown, it can be seen that the fault is hidden in the non-stationary trend and is not easy to identify.
[0167] The first 800 fault-free samples are used to train the DRAE model, of which 70% are training data and the remaining 30% are validation data. The activation function of the model is tanh and sigmoid functions according to the model structure introduced earlier, and the number of nodes in the hidden state is selected as 16. The same parameters are applied in other comparison methods using neural networks to ensure fairness of comparison.
[0168] x i and i As the input of this process, x' is obtained by the operation of formula 1-24. i and y' i , and then calculate the mean square error monitoring statistic and its threshold through the operation of equations 25-27 to achieve fault detection. The visualization results of each method in the principal component space or residual space are as follows: Figures 5a-5f As shown. Figure 5f As shown, the reconstruction error obtained by the method of the present application is more concentrated at the origin coordinates, and faulty samples can be clearly distinguished from normal samples, indicating that it has better performance in non-stationary feature extraction.
[0169] like Figure 6a-6f It can be seen from the fault detection diagram results shown that the fault cannot be effectively and quickly identified in any of the other compared methods. Only the monitoring statistic in the method of the present invention exceeds the control limit immediately after the fault is introduced, and an alarm is issued for the fault.
[0170] Sequence length is an important parameter for dynamic and non-stationary feature extraction in neural networks, which depends on how many previous samples the current sample is associated with. If the sequence length is selected as 1, the performance of the proposed DRAE model will be similar to that of a normal autoencoder. Therefore, a sensitivity analysis is performed to select an appropriate sequence length parameter. As shown in Table 1, when the sequence length is selected as 3, the proposed method achieves the best performance with an FDR of 99.3% and a FAR of 0.65%. In contrast, the FDR of the autoencoder is less than 80% regardless of the sequence length selected. The results show that the proposed method has better performance in monitoring non-stationary signals compared to other methods.
[0171] Table 1. Fault detection performance of autoencoder and DRAE at different sequence lengths
[0172]
[0173] Example 2: Tennessee Eastman Benchmark Process
[0174] The Tennessee Eastman process, a chemical process benchmark platform commonly used to compare fault detection effects in the field of process monitoring, is used to further demonstrate the performance of the proposed DRAE-based method. The Tennessee Eastman process is a chemical simulation process consisting of 5 units and 53 variables. The process is a highly nonlinear dynamic system, and the variables related to the stripper are mostly non-stationary. A total of 20 process disturbances can be introduced into the process, of which the types of the first 15 disturbances are known. Among these 15 disturbances, disturbances 3, 9, and 15 are difficult to detect because the amplitudes of these disturbances are small, so they are not considered in this section. Referring to the literature, we used 22 measured variables and 11 manipulated variables with a sampling frequency of 3 minutes. The normal data set contains 500 samples, and all 15 fault data sets contain 960 samples. The fault is introduced in the 161st sample in each fault data set.
[0175] The parameters of different methods are set as follows. The time lag is 2 in dynamic principal component analysis, and the cumulative variance contribution method is used in principal component analysis, kernel principal component analysis, and dynamic principal component analysis to select the number of principal components so that 85% of the variance of the data can be explained. For all neural network-based methods, the number of hidden states is set to 256; the batch size is 25; the learning rate is 0.01, the sequence length is 5, and the Adam optimizer is selected. In order to avoid overfitting, an early stopping strategy is used in the validation data set. As mentioned above, FDR, FAR, and fault detection delay are used to evaluate the performance of different methods in fault detection. The FDR and FAR results of each method are shown in Tables 2 and 4. It can be seen that the average FDR of the DRAE method of the present invention is the highest except for the dynamic principal component analysis method, but the average FAR of the dynamic principal component analysis method is as high as 7.42%, indicating that the process monitoring results are unreliable. The average FAR of the proposed DRAE-based method is 2.63%, which is only about 1% higher than that of the principal component analysis-based method, while the average FDR of the proposed method is nearly 14% higher than that of the principal component analysis, which is enough to illustrate the superiority of the proposed method. For autoencoder and DRAE, the FARs of both methods are similar, but our DRAE-based method performs better in faults 4, 5, 8, 10, 11, and 13.
[0176] Although the average FDR of the proposed DRAE-based method is only 2.3% higher than that of the autoencoder-based method, their performance can be further compared on the fault detection delay. As shown in Table 3, the average fault detection delay in the proposed DRAE-based method is the shortest. Among these faults, faults 8 and 10 best illustrate the performance of the proposed method in monitoring non-stationary processes. According to the process flow of the process and the description of faults 8 and 10, these two faults are randomly varied, introduced in stream 4, and directly transported to the stripper. Therefore, it can be considered that these faults have an impact on the variables in the stripper, most of which are non-stationary. These faults are hidden in the non-stationary trend of the process in the early stage, so they cannot be detected early by traditional process monitoring methods. In contrast, the proposed DRAE-based method can extract non-stationary features more effectively and distinguish faults from normal non-stationary trends in the early stage of the fault. It can be concluded that fault 8 can be detected by the proposed method 18 to 54 minutes earlier than other methods. The superiority of the proposed method can be more clearly reflected in fault 10. The fault can be detected 21 to 150 minutes in advance by the proposed method. This case study shows that the proposed DRAE-based method also shows better performance in monitoring larger non-stationary systems.
[0177] Table 2 FDR results of each method on different faults
[0178]
[0179] Table 3 Fault detection delay results of each method on different faults
[0180]
[0181] Table 4 FAR results of each method on different faults
[0182]
[0183] Example 3: Industrial continuous catalytic reforming heat exchange unit
[0184] The method of this application was applied to an industrial continuous catalytic reforming heat exchanger of a chemical company in China to demonstrate the performance of the proposed DRAE-based method in actual industrial process applications. The monitoring variables used in this study were collected from the distributed control system of the plant. There are four types of variables: temperature, pressure, differential pressure, and flow, involving 4 reactors, 4 heating furnaces, and a plate heat exchanger, of which the heat exchanger is the most core and expensive equipment. The detailed information of the process variables is shown in Table 5. The process is a highly nonlinear and non-stationary system, in which more than half of the 28 process variables used in this study are non-stationary. Therefore, online monitoring of the process is required to ensure that the heat exchanger maintains stable operation. Due to various complex factors in the actual operation process, the non-stationary characteristics are more significant in the actual industrial process, so fault detection of the process is more challenging.
[0185] Table 5 Variable information table of industrial continuous catalytic reforming heat exchange unit
[0186]
[0187]
[0188] A dataset consisting of 3,000 samples with a sampling frequency of 1 minute under normal operating conditions is used to train the proposed DRAE fault detection model, of which 20% is used as validation data. The test dataset contains 1,000 samples, of which the 660th sample is faulty. The fault manifests itself as an abnormal increase in the hot end pressure drop of the heat exchanger, which is a slowly changing cumulative process. Since the alarm limit of the distributed control system is not reached, the operator cannot detect it in time. This fault will affect the heat exchange efficiency of the heat exchanger, which will lead to an increase in the heating furnace load and gas consumption. Therefore, early detection of this fault is crucial.
[0189] Since the parameter sequence length has a great influence on the effect of the model, the fault detection performance of the autoencoder and DRAE-based methods under different sequence lengths is first evaluated in Table 6. It can be seen that the results of the autoencoder are not affected by the parameter sequence length, while the results of DRAE are greatly affected by the sequence length due to its model structure. Considering the fault detection delay and FAR, the sequence length is selected as 4 in the autoencoder and 10 in the DRAE.
[0190] Table 6 Fault detection performance of autoencoder and DRAE at different sequence lengths
[0191]
[0192]
[0193] Figure 7a-7iThe fault detection results of each method show that the DRAE-based method proposed in the present invention performs better than other methods. The fault detection delay and FAR of different methods are shown in Table 7. The fault detection delay of the DRAE method of the present invention is only 6 minutes, which means that the fault can be detected at least 21 minutes in advance by the proposed method. At the same time, the FAR of this method is also relatively low, and it can be fully applied in industrial processes. Through the proposed method, fault information can be quickly provided to the operator after the fault occurs. After receiving the fault information, the operator can take timely measures to troubleshoot the fault, thereby controlling the abnormal increase in pressure drop and ensuring the profitability and stable operation of the plant. The case study illustrates the industrial application prospects of the method of the present invention in non-stationary industrial process monitoring.
[0194] Table 7
[0195] method Fault detection delay (minutes) FAR(%) Principal component analysis T2 198 5.76 Principal Component Analysis (SPE) 65 6.36 Kernel principal component analysis T2 201 0.61 Kernel principal component analysis SPE 82 12.7 Autoencoder 41 5.49 Variational Recurrent Autoencoder 41 3.23 Dynamic principal component analysis T2 106 7.73 Dynamic principal component analysis SPE 27 7.58 DRAE 6 3.85
[0196] In summary, the scheme proposed in the present invention can propose a DRAE-based process monitoring method for non-stationary industrial processes. A new neural network structure is constructed, in which differential operations are introduced on hidden states to capture the non-stationary characteristics of process data. Combined with the potential of long short-term memory neural networks in nonlinear and dynamic feature extraction, the proposed neural network model can effectively capture the stationary and non-stationary characteristics of the process. And further construct an autoencoder structure to reconstruct the model input with the hidden features extracted by the model, and use the reconstruction error as a monitoring statistic to perform fault detection. First, a non-stationary but single-input process case is used to illustrate the performance of the method in non-stationary process monitoring, and then it is applied to the Tennessee Eastman benchmark industrial process and an actual industrial process in a chemical enterprise, proving the prospect of the inventive method in large-scale systems and industrial applications. Compared with the traditional principal component analysis, kernel principal component analysis, dynamic principal component analysis, and the current advanced autoencoder-based and variational cyclic autoencoder-based methods in this field, the proposed DRAE-based method shows significantly better performance in terms of FDR and fault detection delay. This technology can distinguish slow-changing faults in actual industrial production from non-stationary trends under normal operating conditions by extracting non-stationary features of process data, so as to assist operators in identifying process faults early and quickly and taking corresponding measures to eliminate the faults in a timely manner, thereby improving the economic benefits of industrial production enterprises and reducing the probability of production accidents.
[0197] A second aspect of the present invention discloses a non-stationary industrial process fault detection system. Figure 8 is a structural diagram of a non-stationary industrial process fault detection system according to an embodiment of the present invention; Figure 8 As shown, the system 100 includes:
[0198] The first processing module 101 is configured to obtain historical data under normal working conditions from a factory historical database and perform standardization processing on the historical data;
[0199] The second processing module 102 is configured to introduce differential input into the recurrent neural network based on the long short-term memory neural network, construct a differential recurrent autoencoder, and decode the output of the encoding layer of the differential recurrent autoencoder with a decoder layer to obtain a fault detection model;
[0200] The third processing module 103 is configured to divide the standardized historical data into a training data set and a verification data set, and use the training data set and the verification data set to train the fault detection model;
[0201] The fourth processing module 104 is configured to calculate the reconstruction error of the validation data set using the trained fault detection model as a monitoring statistic, and determine a monitoring statistic threshold;
[0202] The fifth processing module 105 is configured to obtain real-time data from sensors / instruments of the factory in real time and perform standardized processing on the real-time data;
[0203] The sixth processing module 106 is configured to input the standardized real-time data into the trained fault detection model for data reconstruction and calculate the real-time monitoring statistics;
[0204] The seventh processing module 107 is configured to compare the real-time monitoring statistic with the monitoring statistic threshold, and if the real-time monitoring statistic exceeds the monitoring statistic threshold, determine it as a fault, otherwise determine it as a normal operating condition.
[0205] According to the system of the second aspect of the present invention, the second processing module 102 is configured as follows: introducing differential input into the recurrent neural network based on the long short-term memory neural network to construct a differential recurrent autoencoder includes:
[0206] The differential recurrent autoencoder includes four layers of long short-term memory neural networks. The difference between the hidden layer output of the first layer of long short-term memory neural network at the current moment and the hidden layer output of the first layer of long short-term memory neural network at the previous moment is used as the input of the second layer of long short-term memory neural network to extract non-stationary features. The non-stationary features are input into the third layer of long short-term memory neural network to extract the hidden state of the non-stationary features. The hidden state of the non-stationary features is input into the fourth layer of long short-term memory neural network to obtain the hidden features of the differential recurrent autoencoder.
[0207] According to the system of the second aspect of the present invention, the second processing module 102 is configured such that the hidden layer output of the first layer of the long short-term memory neural network is: the hidden layer output of the last layer of the first layer of the long short-term memory neural network.
[0208] According to the system of the second aspect of the present invention, the second processing module 102 is configured to use the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment as the input of the second layer of the long short-term memory neural network, and the extracted non-stationary features include:
[0209] i t1 =σ(W i1 ·[h t -h t-1 , N t-1 ]+b i1 )
[0210] f t1 =σ(W f1 ·[h t -h t-1 , N t-1 ]+b f1 )
[0211] g t1 =tanh(W C1 ·[h t -h t-1 , N t-1 ]+b C1 )
[0212] N t =f t1 ·N t-1 +i t1 ·g t1
[0213] o t1 =σ(W O1 ·[h t -h t-1 , N t ]+b o1 )
[0214] D t =o t1 ·tanh(N t )
[0215] Among them, h t and h t-1 are the hidden layer output of the first layer of the LSTM neural network at the current moment and the hidden layer output of the first layer of the LSTM neural network at the previous moment, N t and N t-1are the cell states of the layer at the current moment and the previous moment, respectively, t represents the extracted non-stationary features, σ and tanh are activation functions, W i1 and b i1 are the weights and biases of the first layer of the second layer of the long short-term memory neural network; W f1 and b f1 are the weights and biases of the second layer of the second layer of the long short-term memory neural network; W C1 and b C1 are the weights and biases of the third layer of the second layer of the long short-term memory neural network; W O1 and b O1 They are the weights and biases of the fourth layer of the second-layer long short-term memory neural network.
[0216] According to the system of the second aspect of the present invention, the third processing module 103 is configured to train the fault detection model with a loss function of:
[0217] The mean square error between the reconstructed output and the actual value is used as the loss function of the fault detection model;
[0218] The reconstructed output is the output of the decoder layer.
[0219] According to the system of the second aspect of the present invention, the fourth processing module 104 is configured such that the monitoring statistic is a mean square error between the reconstructed output and the actual value.
[0220] According to the system of the second aspect of the present invention, the fourth processing module 104 is configured such that the monitoring statistic threshold is the mean square error monitoring statistic threshold under a 99% confidence interval; the mean square error monitoring statistic threshold under a 99% confidence interval is determined by a kernel density estimation method.
[0221] The third aspect of the present invention discloses an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of any one of the non-stationary industrial process fault detection methods disclosed in the first aspect of the present invention are implemented.
[0222] Fig. 9 is a structural diagram of an electronic device according to an embodiment of the present invention, such as Fig. 9As shown, the electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, near field communication (NFC) or other technologies. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covered on the display screen, or a button, a trackball or a touch pad set on the housing of the electronic device, or an external keyboard, touch pad or mouse, etc.
[0223] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a structural diagram of the part related to the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the technical solution of the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0224] The fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in any one of the steps of a non-stationary industrial process fault detection method in the first aspect of the present invention.
[0225] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present application, several variations and improvements can be made, which all belong to the scope of protection of the present application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.
Claims
1. A method for detecting faults in a non-stationary industrial process, characterized in that: The method comprises: Step S1, obtaining historical data under normal working conditions from the factory historical database, and standardizing the historical data; Step S2, introducing differential input into the recurrent neural network based on the long short-term memory neural network, constructing a differential recurrent autoencoder, and decoding the output of the encoding layer of the differential recurrent autoencoder with a decoder layer to obtain a fault detection model; Step S3, dividing the standardized historical data into a training data set and a verification data set, and using the training data set and the verification data set to train the fault detection model; Step S4: Use the trained fault detection model to calculate the reconstruction error of the verification data set as a monitoring statistic, and determine the monitoring statistic threshold; Step S5: acquiring real-time data from sensors / instruments in the factory in real time, and performing standardization processing on the real-time data; Step S6: input the standardized real-time data into the trained fault detection model to reconstruct the data and calculate the real-time monitoring statistics; Step S7: compare the real-time monitoring statistic with the monitoring statistic threshold, and if the real-time monitoring statistic exceeds the monitoring statistic threshold, determine that it is a fault, otherwise determine that it is a normal operating condition; In step S2, the method of introducing differential input into a recurrent neural network based on a long short-term memory neural network to construct a differential recurrent autoencoder includes: The differential cyclic autoencoder includes a four-layer long short-term memory neural network, takes the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment as the input of the second layer of the long short-term memory neural network, extracts non-stationary features, inputs the non-stationary features into the third layer of the long short-term memory neural network, extracts the hidden state of the non-stationary features, inputs the hidden state of the non-stationary features into the fourth layer of the long short-term memory neural network, and obtains the hidden features of the differential cyclic autoencoder; In the step S2, the hidden layer output of the first layer of the long short-term memory neural network is: the hidden layer output of the last layer of the first layer of the long short-term memory neural network; In step S2, the method of extracting non-stationary features by using the difference between the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment as the input of the second layer of the long short-term memory neural network includes: in, h t and h t-1 are the hidden layer output of the first layer of the long short-term memory neural network at the current moment and the hidden layer output of the first layer of the long short-term memory neural network at the previous moment, N t and N t-1 are the cell states of the layer at the current moment and the previous moment respectively, D t represents the extracted non-stationary features, σ and tanh are activation functions, W i1 and b i1 They are the weights and biases of the first layer of the second layer of the long short-term memory neural network; W f1 and b f1 They are the weights and biases of the second layer of the second long short-term memory neural network; W C1 and b C1 They are the weights and biases of the third layer of the second layer of the long short-term memory neural network; W O1 and b O1 They are the weights and biases of the fourth layer of the second-layer long short-term memory neural network.
2. A non-stationary industrial process fault detection method according to claim 1, characterized in that: In step S3, the loss function of training the fault detection model is: The mean square error between the reconstructed output and the actual value is used as the loss function of the fault detection model; The reconstructed output is the output of the decoder layer.
3. A non-stationary industrial process fault detection method according to claim 1, characterized in that: In step S4, the monitoring statistic is the mean square error between the reconstructed output and the actual value.
4. A non-stationary industrial process fault detection method according to claim 3, characterized in that: In step S4, the monitoring statistic threshold is the mean square error monitoring statistic threshold under a 99% confidence interval; the mean square error monitoring statistic threshold under a 99% confidence interval is determined by a kernel density estimation method.
5. A system for detecting faults in non-stationary industrial processes, the system adopting the method according to any one of claims 1 to 4, characterized in that: The system comprises: The first processing module is configured to obtain historical data under normal working conditions from a factory historical database and perform standardization processing on the historical data; The second processing module is configured to introduce differential input into the recurrent neural network based on the long short-term memory neural network, construct a differential recurrent autoencoder, and decode the output of the encoding layer of the differential recurrent autoencoder with a decoder layer to obtain a fault detection model; A third processing module is configured to divide the standardized historical data into a training data set and a verification data set, and use the training data set and the verification data set to train the fault detection model; A fourth processing module is configured to calculate a reconstruction error of a validation data set using the trained fault detection model as a monitoring statistic, and determine a monitoring statistic threshold; The fifth processing module is configured to obtain real-time data from sensors / instruments of the factory in real time and perform standardized processing on the real-time data; A sixth processing module is configured to input the standardized real-time data into the trained fault detection model for data reconstruction and calculate real-time monitoring statistics; The seventh processing module is configured to compare the real-time monitoring statistic with the monitoring statistic threshold, and if the real-time monitoring statistic exceeds the monitoring statistic threshold, determine it as a fault, otherwise determine it as a normal operating condition.
6. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps in the non-stationary industrial process fault detection method described in any one of claims 1 to 4 are implemented.
7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the non-stationary industrial process fault detection method described in any one of claims 1 to 4 are implemented.
Citation Information
Patent Citations
Power load prediction method based on empirical mode decomposition and long-short-term memory network
CN112488415A
Diesel engine air system fault detection method
CN113700558A
Multi-working-condition process industrial fault detection and diagnosis method based on deep transfer learning
CN114065613A
