Fault detection method and system based on reconstructed residual deep learning

Abstract

The invention discloses a fault detection method and system based on reconstruction residual deep learning, and the method and system achieve the deep feature representation learning of the reconstruction residual of process data, thereby effectively mining the essence of hidden abnormal changes in the process data, and achieving the more effective and reliable fault detection. The fault detection method disclosed by the invention not only discloses a brand new deep learning architecture of the reconstructed residual deep learning model, but also discloses various structural improvement schemes related to each auto-encoder forming the reconstructed residual deep learning model, so as to better represent and learn deep features of the reconstructed residual; and the information comprehensiveness in the deep feature extraction process is improved from multiple angles. Besides, a fault detection system based on reconstruction residual deep learning is established through cooperation of hardware and software execution programs, and the fault detection system comprises a central control module, a data acquisition module, a display module and a fault alarm so as to execute the implementation process of the disclosed fault detection method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data-driven fault detection, and particularly relates to a fault detection method and system based on reconstructed residual deep learning. BACKGROUND

[0002] Due to the characteristics of large scale and complexity, real-time monitoring of the running state of modern industrial process objects is of great significance to ensure production safety and maintain product quality stability. However, it is almost impossible to establish an accurate mechanism model to describe the running state, but with the wide application of advanced measurement instrument technology and industrial internet of things technology, more and more data can be collected and stored in industrial production processes, which lays a solid data foundation for data-driven fault detection. Considering the complexity of the production process, the corresponding process data presents significant nonlinear variation characteristics, and the relationship between different measurement variables is also very complex, which increases the application difficulty of traditional statistical process monitoring methods and technologies based on linear transformation.

[0003] In recent years, deep learning technology has been favored by many fields because of its multi-layer feature abstraction capability to effectively handle nonlinear problems. As an unsupervised feature representation learning method, autoencoder and stack autoencoder have been increasingly applied to fault detection. The Chinese invention patent with patent number 202111041934.6 discloses a fault monitoring method based on autoencoder residual reconstruction, which first reconstructs the process data through the autoencoder, and then detects the fault by using the change of the reconstruction residual. However, this method only involves feature extraction of an autoencoder and cannot represent deep abstract objects of process data.

[0004] In addition, a Chinese invention patent with application number / patent number 202411973333.2 reconstructs process data using an introspective orthogonal autoencoder, thereby utilizing the generated reconstruction residuals for fault detection. A paper titled "Industrial Process Fault Detection Based on Distributed Noise Reduction Orthogonal Autoencoders" published in Chemical Automation and Instrumentation in 2023 proposed a method for distributed fault detection by partitioning variables into modules and using a noise reduction orthogonal autoencoder. In addition to preprocessing the variable blocks, the key technical aspect lies in the reconstruction of process data using the noise reduction autoencoder. A 2024 English paper titled "Comparison of Autoencoder Architectures for Fault Detection in Industrial Processes" published in Digital Chemical Engineering detailed a variety of autoencoder and stacked autoencoder structures that can be used for fault detection.

[0005] It should be noted that the fundamental principle behind applying autoencoders to fault detection is to learn feature representations for the input process data and then reconstruct its representation. Changes in the corresponding reconstructed residuals can reflect anomalies in the process data. Stacked autoencoders, by connecting multiple autoencoder layers in series, can exploit deeper levels of abstract features for feature representation. Their application to fault detection shares the same principle as that of autoencoders. However, learning feature representations for the input process data only focuses on reconstructing the original data. Regardless of the number of autoencoder layers used, changes in the reconstructed residuals remain relative to the original input process data. Because the reconstructed residuals are key to reflecting anomalies in process data and, therefore, fault detection, deep representation learning of the residual features has received little attention. From a fault detection perspective, deep learning of residual changes can better reflect the nature of anomalies in process data and is highly beneficial. Therefore, deep learning of reconstructed residuals, enabling more effective and reliable fault detection, remains a pressing challenge. Summary of the Invention

[0006] The main technical problem to be solved by the present invention is: how to perform deep feature representation learning on the reconstructed residuals, so as to effectively mine the essence of abnormal changes hidden in process data and achieve more effective and reliable fault detection.

[0007] The technical solution adopted by the present invention to solve the above problems is: a fault detection method and system based on reconstructed residual deep learning; wherein, the implementation process of the fault detection method based on reconstructed residual deep learning disclosed by the present invention includes steps 1 to 5 shown below.

[0008] Step 1: Obtain N sets of process data from a historical database of industrial processes that are operating normally. After standardization, each set of process data is arranged in columns to form an input data matrix X.

[0009] Step 2: Use each column vector in X as the input vector, build and train the corresponding autoencoder, and then use the trained parameter set Calculate the reconstruction matrix corresponding to X After that, the initial reconstruction residual matrix is ​​output Among them, V0 and Represent the weight matrices of the encoding layer and decoding layer in the autoencoder, p0 and denote the bias vectors of the encoding layer and decoding layer respectively.

[0010] Step 3: Use each reconstructed residual vector in E0 as the input vector, build and train a reconstructed residual deep learning model consisting of K autoencoders in series, and output the reconstructed residual matrix E of the last autoencoder. K .

[0011] Step 4. Calculate E K The fault index of each reconstructed residual vector in is used to determine the threshold of the fault index.

[0012] Step 5: Repeat the following steps 5-1 to 5-4 according to the sampling time interval to implement online fault detection.

[0013] Step 5-1: Get a set of process data at the latest sampling time, standardize it, and form a column vector x by column t ; Among them, the subscript t represents the latest sampling time.

[0014] Step 5-2, sequentially pass h0=f(V0x t +p0) and Calculate the eigenvector h0 and the reconstruction vector After that, pass Output the initial reconstructed residual vector ε0; where f() and ζ() represent the nonlinear activation functions of the encoding layer and decoding layer, respectively.

[0015] Step 5-3: Using ε0 as the input vector, the reconstructed residual deep learning model trained in step 3 outputs the reconstructed residual vector ε of the last autoencoder. K Then, use ε K Calculate the fault index at the latest sampling time t and judge Is it greater than If not, the industrial process is running normally; if so, execute step 5-4 to determine whether a fault occurs.

[0016] Step 5-4, continue to acquire process data of the latest sampling time until at least three fault indicators of the latest sampling time are obtained, and then determine whether they are all greater than the threshold value If yes, a fault alarm is triggered; if no, it is normal.

[0017] In the above implementation steps, the core technical features of the method of the present application distinguished from the prior art solution are that the reconstruction residual deep learning model composed of K serially connected autoencoders, the number of neurons of the encoding layer of each autoencoder is equal, and the kth autoencoder in series is the reconstruction residual vector ε k-1 As an input vector, sequentially pass through h k = f(V k ε k-1 + p k ) and Calculate the feature vector h k and the reconstruction vector of the kth autoencoder Then pass through Output the corresponding reconstruction residual vector ε k ; wherein V k and respectively represent the weight matrix of the encoding layer and the decoding layer in the kth autoencoder, p k and respectively represent the bias vector of the encoding layer and the decoding layer, the number k = 1, 2, …, K, and when k = 1, ε k-1 represents ε0.

[0018] As can be seen, the reconstruction residual deep learning method of the present application is not a layer-by-layer representation learning of the feature vector of the encoding layer according to the traditional idea, but a layer-by-layer feature representation learning of the reconstruction residual output by each serially connected autoencoder; compared with the traditional stack autoencoder which learns the changes of the encoding layer features layer by layer, the reconstruction residual deep learning model involved in the present application can learn the changes of the reconstruction residual output by the previous autoencoder in a step-by-step progressive manner, thus better analyzing and mining the change features of the reconstruction residual and improving the sensitivity of subsequent fault detection based on the reconstruction residual.

[0019] In addition, the reason why the number of neurons of the encoding layer in each autoencoder in the reconstruction residual deep learning model is required to be equal is that the input vectors of each autoencoder in the reconstruction residual deep learning model are equal in dimension, and setting the same number of neurons not only plays a role similar to the dimension reduction of feature representation learning in the stack autoencoder, but also avoids the information loss problem caused by the layer-by-layer decrease of the number of neurons in the autoencoding layer in the traditional stack autoencoder;

[0020] Further, the reconstruction residual deep learning model does not need to be fine-tuned after sequentially completing the training of the k=1, 2, …, K autoencoders; in other words, the reconstruction residual deep learning model only involves the pre-training phase of the traditional stacked autoencoder, and does not involve the parameter fine-tuning phase; compared with the traditional method, the training process of the model is simplified, and the calculation efficiency is improved.

[0021] As an improvement of the present application, each autoencoder in the reconstruction residual deep learning model further comprises an encoding layer for encoding the process data, and further combines the conversion gate layer to generate the corresponding feature vector h k , so that the calculation method of h k is changed to: , wherein, represents the multiplication of elements at the same position, W k and b k represent the weight matrix and the bias vector of the conversion gate layer respectively, δ() represents the sigmoid function, x represents a column vector composed of process data corresponding to ε k-1 , [ε k-1 ; x] represents the combination of ε k-1 and x into a column vector, Λ k and a k represent the weight matrix and the bias vector for encoding the process data respectively.

[0022] Based on the above improved technical solution, when the reconstruction residual deep learning model is built and trained, the k=1, 2, …, K autoencoders simultaneously take the column vectors at the same column in E k-1 and X as input vectors, and respectively perform encoding conversion through the respective encoding layers, i.e. f(V k ε k-1 +p k ) and f(Λ k x+a k ); since the activation function of the conversion gate layer is the sigmoid function, i.e. δ(u)=1 / (1+e -u ), the range of the output value is 0 to +1, and therefore the rule of multiplying elements at the same position can be used to weight and combine the elements at the same position in f(V k ε k-1 +p k ) and f(Λ k x+a k ).

[0023] Considering that the reconstruction residual deep learning model takes the initial reconstruction residual of the reconstructed original process data as the input vector, and the initial reconstruction residual inherits part of the information in the original process data, therefore, the improved technical solution can further combine the feature representation of the column vector composed of the original process data on the basis of the feature representation learning of the reconstruction residual vector by combining the encoding layer and the conversion gate layer. As can be seen, the combination mode in the improved technical solution can better represent the deep features of the reconstruction residual, and avoid the information loss of the original process data in the deep transformation.

[0024] As another improvement of the present application, in order to better represent the change characteristics of the reconstruction residual in the encoding layer of each autoencoder in the reconstruction residual deep learning model, each autoencoder constituting the reconstruction residual deep learning model further comprises a gate layer in parallel with the encoding layer, so that the calculation mode of the feature vector h k of the encoding layer is changed to Wherein, represents the multiplication of elements at the same position, G k and q k represent the weight matrix and the bias vector of the gate layer respectively, and δ() represents the sigmoid function.

[0025] With the participation of the gate layer, when deep learning of the reconstruction residual is performed, since the activation function of the gate layer is also a sigmoid function with an output value range of 0 to +1, the rule of multiplying elements at the same position can be used to weight each element obtained by the calculation mode f(V k ε k-1 +p k ) of the original feature vector, so as to better perform feature representation learning of the encoding layer.

[0026] As another improvement of the present application, there is a conversion gate layer between the encoding layer and the decoding layer of each autoencoder in the reconstruction residual deep learning model, so that the calculation mode of h k is changed to: Wherein, h k-1 represents the feature vector of the k-1th autoencoder, and when k=1, h k-1 =h0.

[0027] The above improved technical solution actually converts the feature vector of the previous autoencoder and the feature vector f(V k ε k-1 +p k) are weightedly combined to make full use of the component information of the encoding layer feature vector of the previous autoencoder, avoiding the loss of the feature component information of the encoding layer in the shallow autoencoder when extracting the features of the encoding layer of the deeper autoencoder, thereby better performing the feature representation learning of the encoding layer.

[0028] The above three improved technical solutions are all aimed at better learning the feature representation of the coding layer. On the basis of these improved technical solutions, the loss function of each autoencoder of the training and reconstruction residual deep learning model can be further improved to further analyze and extract more useful feature components.

[0029] As a further improvement of the present invention and the various improved technical solutions described above, when building the reconstructed residual deep learning model in step 3, it is preferred to add an output layer in parallel to the decoding layer of each autoencoder for reconstructing the process data, and correspondingly change the loss function L for training each autoencoder to:

[0030] in, and Respectively represent E k-1 and The i-th column vector in X (i) and Represents X and The i-th column vector in , Represents calculation The square value of the length of the regular coefficient λ is in the range of 0.001≤λ≤0.1, E k-1 represents the reconstructed residual matrix output by the k-1th autoencoder, Represents the reconstruction matrix calculated by the kth autoencoder And when k=1, E k-1 and Represent E0 and its corresponding reconstruction matrix respectively

[0031] Typically, the default loss function for training autoencoders is the first half of L. The regularization term in the second half of L guides the analysis of each autoencoder to extract feature vectors directly related to the process data, thereby mining the deep features of the reconstructed residual while taking into account the reconstruction of the original process data to represent more comprehensive feature change information for more effective fault detection.

[0032] Since the loss function of training the autoencoder has a direct guiding role in the generation of feature vectors, in addition to the form defined in the above formula ①, there are two other preferred forms of the loss function, as shown below:

[0033] wherein, and respectively represent the eigenvectors of the i-th column and the j-th column in H k , and H k represents a feature matrix composed of N eigenvectors generated by the k-th autoencoder when training the reconstruction residual deep learning model, and (i) represents a set composed of the column numbers of the C column vectors in X with the minimum distance between X (i) and X (i) . (i) represents a set composed of the column numbers of the C column vectors in X with the closest sampling time corresponding to X k-1 , and the value range of the regularization coefficient λ is 0.001≤λ≤0.1, and C is an integer not less than 5 and not more than 10.

[0034] It can be seen that the regularization terms added in the above formula ② and formula ③ respectively consider the neighbor relationship between the original process data in the spatial distance and the sampling time distribution, which can guide the extracted eigenvectors to maintain the inherent neighbor relationship characteristics of the original process data, so as to mine and analyze the reconstruction residual change characteristics in a deep level, and simultaneously consider the representation learning of the neighbor relationship characteristics of the original process data from two different angles, so as to extract more comprehensive feature components for subsequent fault detection.

[0035] As another improvement of the present application, the input vector of each autoencoder used to compose the reconstruction residual deep learning model is preferably composed of the reconstruction residual vector ε k-1 output by the previous autoencoder and the column vector x corresponding to the process data, and the calculation mode of the corresponding h k is changed to h k =f(V k [ε k-1 ; x]+p k ).

[0036] This improved technical solution further utilizes the original process data when representing the reconstruction residual vector, and since the information contained in the input vector is increased, the feature representation learning of the reconstruction residual vector can be better through h k =f(V k [ε k-1 ; x]+p k ), which can also help to enhance the effect of subsequent fault detection.

[0037] The present application also provides a fault detection system based on reconstruction residual deep learning, which comprises a hardware subsystem composed of a central control module, a data acquisition module, a display module, and a fault alarm, and a medium storing an execution program; the functions of each component are as follows.

[0038] The execution program stored in the medium consists of two parts: offline modeling and online detection. The execution program of the offline modeling part implements steps 1 to 4, while the execution program of the online detection part implements steps 5-2 to 5-3.

[0039] The central control module periodically sends a signal to the data acquisition module according to the set sampling time interval, so that it obtains a set of process data at the latest sampling time t from the industrial process database and sends it to the central control module immediately;

[0040] The central control module will immediately call and implement the execution program of the offline modeling part stored in the medium according to the initialization instructions, and then reconstruct the residual deep learning model and threshold after training. Send to media for storage.

[0041] The central control module can periodically send signals to the data acquisition module according to the set sampling time interval, so that it can obtain a set of process data at the latest sampling time t from the industrial process database and send it to the central control module immediately. After receiving the set of process data, the central control module first performs standardization on it and then forms a column vector x according to the column. t After that, the execution program of the online detection part stored in the medium is immediately called and implemented, and the fault indicators obtained are The data is sent together with the threshold to the display module for real-time display; at the same time, the central control module will determine whether the fault indicators of the latest three sampling moments are all greater than the threshold. If yes, trigger the fault alarm; if no, keep the fault alarm silent

[0042] The display module will display the fault indicators of the latest sampling time t and several sampling times before it in real time, and at the same time set the threshold Displayed as a horizontal line.

[0043] The reconstructed residual deep learning model involved in the execution program has an equal number of neurons in the encoding layer of each autoencoder, and the kth autoencoder in series is the reconstructed residual vector ε output by the previous autoencoder k-1 As input vector, pass h k =f(V k ε k-1 +p k )and Calculate the feature vector h of the kth autoencoder k and reconstruction vector Then pass Output the corresponding reconstructed residual vector ε k .

[0044] Therefore, the fault detection system disclosed in the present invention and the previously disclosed fault detection method are based on the same inventive concept; more specifically, the fault detection system aims to implement the steps included in the above-mentioned fault detection method by combining hardware and software execution programs.

[0045] In addition, based on the above-mentioned fault detection system, the present invention provides an improved fault detection system. The execution program stored in the medium constituting the system, when involved in the reconstruction of the residual deep learning model, each of its autoencoders is preferably in accordance with h k =f(V k ε k-1 +p k )⊙δ(W k [ε k-1 ;x]+b k )+f(Λ k x+a k )⊙(1-δ(W k [ε k-1 ;x]+b k )), or h k =f(V k ε k-1 +p k )⊙δ(G k ε k-1 +q k ), or h k =f(V k ε k-1 +p k )⊙δ(W k [f(V k ε k-1 +p k );h k-1 ]+b k )+h k-1 ⊙(1-δ(W k [f(V k ε k-1 +p k );h k-1 ]+b k ) to calculate the eigenvector h k .

[0046] It is obvious that this improved fault detection system uses the three improved technical solutions of the autoencoder that constitutes the reconstruction residual deep learning model of the above-mentioned fault detection method, which can better analyze and explore the changing characteristics of the reconstruction residual and improve the sensitivity of subsequent fault detection based on the reconstruction residual. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1This is a structural diagram of a typical chemical process object.

[0048] Figure 2 This is a structural diagram of the reconstructed residual deep learning model involved in the present invention.

[0049] Figure 3 The figure is a schematic diagram of the fault detection effect of the fault detection method according to the present invention.

[0050] Figure 4 Schematic diagram of the fault detection effects corresponding to the two fault detection methods based on the traditional stacked autoencoder.

[0051] Figure 5 This is a structural diagram of any autoencoder in the reconstructed residual deep learning model in the second embodiment.

[0052] Figure 6 This is a schematic diagram of the fault detection effect of the second embodiment.

[0053] Figure 7 This is a structural diagram of any autoencoder in the reconstructed residual deep learning model in the third embodiment.

[0054] Figure 8 This is a structural diagram of any autoencoder in the reconstructed residual deep learning model in the fourth embodiment.

[0055] Figure 9 The figure is a schematic diagram of the structure of the fault detection system involved in the present invention. DETAILED DESCRIPTION

[0056] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] In order to clearly and specifically describe the specific implementation of the fault detection method of the present invention, it is applied to Figure 1 In the fault detection task of a typical chemical process object shown in Figure 1, the object mainly consists of five parts: reactor, condenser, stripping tower, compressor, and separator. It can measure the process data of 33 measurement variables such as flow rate (FI), temperature (TI), pressure (PI), liquid level (LI), and valve opening in real time at each sampling moment.

[0058] First, follow steps 1 to 4 below to build a model for implementing fault detection.

[0059] Step 1: Obtain N=960 sets of process data from a historical database of industrial processes that are operating normally. After standardization, each set of process data is arranged in columns to form an input data matrix X.

[0060] Step 2: Use each column vector in X as the input vector, build and train the corresponding autoencoder, and then use the trained parameter set Calculate the reconstruction matrix corresponding to X After that, the initial reconstruction residual matrix is ​​output

[0061] Step 3: Use each reconstructed residual vector in E0 as the input vector to build and train a Figure 2 After the reconstruction residual deep learning model consisting of K=3 autoencoders connected in series, the reconstructed residual matrix E3 of the last autoencoder is output.

[0062] Step 4. Calculate E K The fault index of each reconstructed residual vector in is used to determine the threshold of the fault index.

[0063] In this technical field, E K The i-th reconstructed residual vector in Fault indicators Generally, according to the formula To calculate, that is, use The squared Mahalanobis distance is used as the fault indicator; among them, Indicates E K The average value of all reconstructed residual vectors in E K The covariance matrix of .

[0064] from Figure 2 It can be found that the kth autoencoder in series is the reconstructed residual vector ε output by the previous autoencoder k-1 As input vector, pass h k =f(V k ε k-1 +p k )and Calculate the feature vector h of the kth autoencoder k and reconstruction vector Then pass Output the corresponding reconstructed residual vector ε k ; In addition, in this embodiment, the number of neurons in the encoding layer of each autoencoder is set to be equal, and the nonlinear activation function of the encoding layer and decoding layer of the autoencoder are both selected as tanh function.

[0065] After completing the reconstruction residual deep learning model training and determining the threshold of the fault indicator After that, you can perform online fault detection according to steps 5-1 to 5-4 shown below.

[0066] Step 5-1, a set of process data at the latest sampling time is obtained, and after standardization processing, column vectors x are formed according to columns t .

[0067] Step 5-2, the feature vector h0 of the encoding layer and the reconstruction vector of the decoding layer are calculated in turn through h0 = f(V0x t +p0) and . Then, the initial reconstruction residual vector ε0 is calculated through .

[0068] Step 5-3, the reconstruction residual vector ε of the last autoencoder is outputted through the reconstruction residual deep learning model trained in step 3, with ε0 as the input vector K . Then, the fault index at the latest sampling time t is calculated through K . And it is judged whether is greater than . If not, the industrial process runs normally; if yes, step 5-4 is executed to judge whether a fault occurs.

[0069] Step 5-4, the process data at the latest sampling time is continuously obtained, and after at least three fault indexes at the latest sampling time are obtained, it is judged whether they are all greater than the threshold value . If yes, a fault alarm is triggered; if not, it runs normally.

[0070] According to the above steps 5-1 to 5-4, fault detection is performed on the process data at 960 latest sampling times, and the corresponding fault indexes are plotted in together with the threshold value Figure 3 . From Figure 3 , it can be found that the chemical process object runs normally at the first 160 sampling times, and a fault occurs from the 161st sampling time.

[0071] In order to verify the superiority of the first embodiment of the present application relative to the conventional method technology, the fault detection effect diagrams corresponding to two fault detection methods based on the conventional stack autoencoder are shown in Figure 4 . Compared with Figure 3 , Figure 4 , there is a more obvious fault miss report after the 161st sampling time in the two fault detection diagrams in Figure 3 and Figure 4 . Therefore, through the comparison of and

[0072] , it is fully proved that the fault detection method disclosed in the present application significantly improves the sensitivity of fault detection, and brings significant technical effects. Figure 3The fault detection effect shown in the figure shows that after the fault occurs, there are still some sampling moments where the fault index is lower than the threshold. There is still a certain amount of missed fault reports.

[0073] In order to further improve the fault detection effect, the second embodiment of the present invention is to improve the various autoencoders that constitute the reconstructed residual deep learning model; taking the improvement of the kth autoencoder as an example, the improved structure diagram is as follows Figure 4 As shown in Figure 2, an encoding layer is added to encode the process data, and the corresponding feature vector h is generated by combining the conversion gating layer. k ; correspondingly, h k =f(V k ε k-1 +p k )⊙δ(W k [ε k-1 ;x]+b k )+f(Λ k x+a k )⊙(1-δ(W k [ε k-1 ;x]+b k )).

[0074] based on Figure 5 The reconstruction residual deep learning model composed of the autoencoder structure shown in the figure is used to re-determine the threshold of the fault indicator after training is completed. Then, according to steps 5-1 to 5-4, the same 960 latest sampling moments of process data are used for fault detection. The corresponding fault detection results are as follows: Figure 6 shown.

[0075] Compared to Figure 3 , Figure 6 The fault detection effect has been further improved, which is mainly due to the fact that the improved autoencoder structure can better represent the deep features of the learning and reconstruction residuals, avoiding the information loss of the original process data in the deep transformation.

[0076] Inspired by this, the third embodiment of the present invention also improves the various autoencoders that constitute the reconstructed residual deep learning model. The corresponding structural diagram is shown in FIG. Figure 7 shown; from Figure 7 As can be seen from the figure, the improved autoencoder also includes a gating layer in parallel with the encoding layer, so that h k The calculation method is changed to h k =f(V k ε k-1 +p k )⊙δ(G k ε k-1 +q k ).

[0077] The fourth embodiment of the present application is also an improvement on each autoencoder constituting the reconstruction residual deep learning model, and the corresponding structural diagram is shown in Figure 8 As can be seen from Figure 8 , the improved autoencoder has a conversion gate layer between the encoding layer and the decoding layer, so that the calculation method of the feature vector h k is changed to: h k = f(V k ε k-1 + p k ) o δ(W k [f(V k ε k-1 + p k ); h k-1 ] + b k ) + h k-1 o (1- δ(W k [f(V k ε k-1 + p k ); h k-1 ] + b k ).

[0078] The improvements on each autoencoder constituting the reconstruction residual deep learning model in the third and fourth embodiments of the present application can have almost the same effect as the second embodiment, and the same 960 latest sampling time process data are used for fault detection, and the corresponding fault detection effect can be further improved, which will not be shown here.

[0079] The fifth, sixth, and seventh embodiments of the present application are improvements on the loss function used to train each autoencoder, and the above formulas ①, ②, and ③ are used respectively during training, aiming to capture more and more comprehensive information from different angles to improve the subsequent fault detection effect.

[0080] The eighth embodiment of the present application is an improvement on the input vector of each autoencoder constituting the reconstruction residual deep learning model based on the first embodiment, that is, the input vector of each autoencoder is a column vector x composed of the reconstruction residual vector ε k-1 output by the previous autoencoder and the corresponding process data, and the calculation method of h k is changed to h k = f(V k [ε k-1 ; x] + p k ).

[0081] This improved way can make the extracted feature vector capture more information for fault detection by increasing the information of the input vector.

[0082] The application also provides a fault detection system based on reconstruction residual deep learning, which has the structure as shown in the figure. Figure 9 The system includes a hardware subsystem composed of a central control module, a data acquisition module, a display module, and a fault alarm, and a medium storing an execution program.

[0083] Figure 9 The central control module sends a signal to the data acquisition module periodically according to the set sampling time interval, so that it can obtain a set of process data at the latest sampling time t from the database of the industrial process and send it to the central control module immediately.

[0084] In addition, after receiving the set of process data, the central control module first performs standardization processing, and then forms a column vector x t according to the column, and then calls the execution program stored in the medium to implement steps 5-2 to 5-3, and then sends the obtained fault indicator to the display module for real-time display; at the same time, the central control module judges whether the fault indicators at the three latest sampling times are greater than the threshold If yes, the fault alarm is triggered; if not, the fault alarm remains silent.

[0085] The display module displays the fault indicators of the latest sampling time t and several sampling times before it in real time, and displays the threshold in the form of a horizontal line at the same time, and the specific display method can refer to Figure 3 or Figure 6 .

[0086] Figure 9 The medium stores the execution program for implementing steps 1 to 4 in addition to the execution program for implementing steps 5-2 to 5-3; in addition, the reconstruction residual deep learning model involved in the execution program has equal number of neurons in each encoding layer of each autoencoder, and the kth autoencoder in series is the reconstruction residual vector ε k-1 output by the previous autoencoder as the input vector, which is sequentially passed through h k = f(V k ε k-1 + p k ) and to calculate the feature vector h k and the reconstruction vector of the kth autoencoder, and then output the corresponding reconstruction residual vector ε k through .

[0087] exist Figure 9 In the fault detection system shown in FIG, the execution procedure involving the reconstruction of the residual deep learning model can be further modified to be based on h k =f(V k ε k-1 +p k )⊙δ(W k [ε k-1 ;x]+b k )+f(Λ k x+a k )⊙(1-δ(W k [ε k-1 ;x]+b k )), or h k =f(V k ε k-1 +p k )⊙δ(G k ε k-1 +q k ), or h k =f(V k ε k-1 +p k )⊙δ(W k [f(V k ε k-1 +p k );h k-1 ]+b k )+h k-1 ⊙(1-δ(W k [f(V k ε k-1 +p k );h k-1 ]+b k ) to calculate the eigenvector h k ; These three calculation methods correspond to Figure 5 、 Figure 7 ,and Figure 8 Three improved autoencoders are shown in .

[0088] The above explanation of the technical solution disclosed in the present invention in combination with the accompanying drawings is illustrative rather than restrictive. The technical solution disclosed in the present invention is not limited to the above-mentioned implementation cases. Any modifications made without departing from the purpose of the technical solution disclosed in the present invention and the scope of protection of the claims shall fall within the scope of protection of the present invention.

Claims

1. A fault detection method based on reconstructed residual deep learning, comprising the following steps 1 to 5: Step 1: Obtain N sets of process data from a historical database of industrial processes that are operating normally. After standardizing each set of process data, arrange them in columns to form an input data matrix X. Step 2: Use each column vector in X as the input vector, build and train the corresponding autoencoder, and then use the trained parameter set Calculate the reconstruction matrix corresponding to X After that, the initial reconstruction residual matrix is ​​output in, V0 and Represent the weight matrices of the encoding layer and decoding layer in the autoencoder, p0 and Represent the bias vectors of the encoding layer and decoding layer respectively; Step 3: Use each reconstructed residual vector in E0 as the input vector, build and train a reconstructed residual deep learning model consisting of K autoencoders in series, and output the reconstructed residual matrix E of the last autoencoder. K ; Step 4. Calculate E K The fault index of each reconstructed residual vector in is used to determine the threshold of the fault index. Step 5: Repeat the following steps 5-1 to 5-4 according to the sampling time interval to implement online fault detection; Step 5-1: Get a set of process data at the latest sampling time, standardize it, and form a column vector x by column t ; Step 5-2, sequentially through h0=f(V0x t +p0) and Calculate the eigenvector h0 and the reconstruction vector back, Then pass Output the initial reconstructed residual vector ε0; where f() and ζ() represent the nonlinear activation functions of the encoding layer and the decoding layer respectively; Step 5-3: Using ε0 as the input vector, the reconstructed residual deep learning model trained in step 3 outputs the reconstructed residual vector ε of the last autoencoder. K Then, use ε K Calculate the fault index at the latest sampling time t and judge Is it greater than If not, the industrial process is operating normally; if so, execute step 5-4 to determine whether a fault occurs; Step 5-4: Continue to obtain the process data at the latest sampling time until you get at least 3 fault indicators at the latest sampling time, and then determine whether they are all greater than the threshold. If yes, a fault alarm is triggered; if no, the operation is normal; it is characterized in that the number of neurons in the encoding layer of each autoencoder in the reconstructed residual deep learning model is equal, and the kth autoencoder in series is the reconstructed residual vector ε output by the previous autoencoder k-1 As input vector, pass h k =f(V k ε k-1 +p k )and Calculate the feature vector h of the kth autoencoder k and reconstruction vector Then pass Output the corresponding reconstructed residual vector ε k ; Among them, V k and Represent the weight matrices of the encoding layer and decoding layer in the kth autoencoder, p k and Denotes the bias vectors of the encoding layer and the decoding layer respectively, numbered k = 1, 2, ..., K and when k = 1, ε k-1 Indicates ε0.

2. A fault detection method according to claim 1, characterized in that: Each autoencoder in the reconstructed residual deep learning model also includes an encoding layer for encoding process data, and further combines the conversion gating layer to generate the corresponding feature vector h k , so that the eigenvector h k The calculation method is changed to: in, Indicates the multiplication of elements at the same position, W k and b k Represent the weight matrix and bias vector of the conversion gating layer, δ() represents the sigmoid function, and x represents the value corresponding to ε k-1 The column vector of process data, [ε k-1 ; x] means to convert ε k-1 and x into a column vector, Λ k and a k They represent the weight matrix and bias vector for encoding process data respectively.

3. A fault detection method according to claim 1, characterized in that: Each autoencoder in the reconstructed residual deep learning model also includes a gating layer in parallel with the encoding layer, so that the feature vector h k The calculation method is changed to h k = Among them, G k and q k denote the weight matrix and bias vector of the gating layer respectively.

4. A fault detection method according to claim 1, characterized in that: There is also a conversion gating layer between the encoding layer and the decoding layer of each autoencoder in the reconstructed residual deep learning model, so that h k The calculation method is changed to: Among them, h k-1 represents the feature vector of the previous autoencoder, and when k=1, h k-1 =h0.

5. A fault detection method according to any one of claims 1 to 4, characterized in that: In step 3, when building the reconstructed residual deep learning model, an output layer is added in parallel to the decoding layer of each autoencoder to reconstruct the process data, and the loss function L for training each autoencoder is changed to: in, and Respectively represent E k-1 and The i-th column vector in X (i) and Represents X and The i-th column vector in , Represents calculation The square value of the length of the regular coefficient λ is in the range of 0.001≤λ≤0.1, E k-1 represents the reconstructed residual matrix output by the k-1th autoencoder, Denotes the value corresponding to E calculated by the kth autoencoder k-1 The reconstruction matrix And when k=1, E k-1 and Represent E0 and its corresponding reconstruction matrix respectively 6. A fault detection method according to any one of claims 1 to 4, characterized in that: In step 3, when training each autoencoder in the reconstructed residual deep learning model, the loss function L used is: in, and Respectively represent H k The eigenvectors of the i-th and j-th columns in H k is a feature matrix composed of N feature vectors generated by the k-th autoencoder, Θ{X (i) } indicates that X (i) The set consists of the column numbers of the C column vectors with the smallest distance between them, the value range of the regularization coefficient λ is 0.001≤λ≤0.1, and C is an integer not less than 5 and not greater than 10.

7. A fault detection method according to any one of claims 1 to 4, characterized in that: In step 3, when training the autoencoders constituting the reconstructed residual deep learning model, the loss function L used is: Among them, Φ{X (i) } indicates that X (i) A set of column numbers of the C column vectors closest to the corresponding sampling time.

8. The anomaly detection method according to claim 1, characterized in that: The input vector of each autoencoder in the reconstructed residual deep learning model is the reconstructed residual vector ε output by the previous autoencoder k-1 and its corresponding process data column vector x, the corresponding h k The calculation method is changed to h k =f(V k [ε k-1 ;x]+p k ).

9. A fault detection system based on reconstructed residual deep learning, comprising a hardware subsystem consisting of a central control module, a data acquisition module, a display module, and a fault alarm, and a medium for storing an execution program; wherein, The execution program stored in the medium consists of two parts: offline modeling and online detection. The execution program of the offline modeling part implements steps 1 to 4, while the execution program of the online detection part implements steps 5-2 to 5-3. The central control module will immediately call and implement the execution program of the offline modeling part stored in the medium according to the initialization instructions, and then reconstruct the residual deep learning model and threshold after training. Send to the medium for storage; The central control module can periodically send signals to the data acquisition module according to the set sampling time interval, so that it can obtain a set of process data at the latest sampling time t from the industrial process database and send it to the central control module immediately; After receiving the set of process data, the central control module first performs standardization on it and then forms a column vector x t back, Immediately call and implement the execution program of the online detection part stored in the medium, and then use the obtained fault indicators The data is sent together with the threshold to the display module for real-time display; at the same time, the central control module will determine whether the fault indicators of the latest three sampling moments are all greater than the threshold. If yes, trigger the fault alarm; if no, keep the fault alarm silent; The display module will display the fault indicators of the latest sampling time t and several sampling times before it in real time, and at the same time set the threshold Displayed in horizontal line format; Its characteristic is that the reconstruction residual deep learning model involved in the execution program has the same number of neurons in the encoding layer of each autoencoder, and the kth autoencoder in series is the reconstructed residual vector ε output by the previous autoencoder k-1 As input vector, pass h k =f(V k ε k-1 +p k )and Calculate the feature vector h of the kth autoencoder k and reconstruction vector Then pass Output the corresponding reconstructed residual vector ε k .

10. A fault detection system according to claim 9, characterized in that: When the execution program involves reconstructing the residual deep learning model, each autoencoder calculates the feature vector h according to the calculation method described in any one of claims 2 to 4. k .

Images