A fault detection method and device based on a PLSR autoencoder

Abstract

The application discloses a kind of based on PLSR self-encoder's fault detection method and device, after realizing the construction of PLSR self-encoder by neighborhood variation characteristic analysis stage and PLSR analysis stage, real-time sampling sample data is detected using it.In neighborhood variation analysis stage, the application discloses three technical solutions for each normal sample data to implement neighborhood variation characteristic learning, realize all-around, multi-level learning representation normal data variation characteristics.In addition, the application also discloses the technology for determining the neighborhood matrix corresponding to each normal sample data from different angles, further guaranteeing the all-around of neighborhood variation characteristic analysis.Finally, according to the implementation process of the disclosed fault detection method, a fault detection device is designed correspondingly, which saves and provides real-time key technical parameters to realize state monitoring of real-time sampling sample data, and provides real-time monitoring image.

Description

Technical Field

[0001] This invention relates to the field of operational status monitoring technology, and in particular to a fault detection method and device based on a PLSR self-encoder. Background Technology

[0002] With the widespread use of computer technology and advanced measuring instruments in production operations, chemical process objects are collected and stored in massive amounts of process data in real time. This data can directly or indirectly reflect the operating status of chemical process objects, thereby enabling real-time fault detection. Compared with using the mechanistic model of chemical process objects to implement fault detection tasks, data-driven fault detection only relies on process data, avoiding the need for establishing a high-precision mechanistic model. Its implementation process is simple and easy, and it has a great feasibility advantage in fault detection.

[0003] Data-driven fault detection in chemical processes requires prior feature analysis of sample data under normal operating conditions. This allows the system to learn and represent the normal variation characteristics of the dataset. Fault detection then involves the immediate identification of whether the features corresponding to the online real-time sampled data conform to the predefined normal variation range. Based on this core technology, autoencoders from the neural network field have been successfully used for fault detection. For example, a paper published on February 1, 2023, in the journal *Chemometrics and Intelligent Laboratory Systems*, titled "Nonlocal, local and global preserving stacked autoencoder based faultdetection method for nonlinear process monitoring," discloses a technique for generating reconstruction errors using an autoencoder composed of local, nonlocal, and global features of the data for fault detection in chemical processes. This method uses an autoencoder that focuses on the normal variation characteristics of the distribution of nearest and non-nearest neighbor relationships in the data space.

[0004] Chinese invention patent application number / patent number 202010601588.1 discloses a manufacturing process anomaly monitoring method based on quality data manifold features. This method detects anomalies or faults in operation by learning global and local features of the training dataset and fusing a mechanistic model. Chinese invention patent application number / patent number 201910474174.4 discloses a process monitoring method based on a dynamic orthogonal nearest neighbor preserving embedding model (hereinafter referred to as: the comparison method). This comparison method utilizes the structural preservation features of temporal and spatial nearest neighbors of the normal training dataset to define normal change features. The above three prior art technologies all demonstrate the beneficial role of nearest neighbor distribution in describing and defining normal changes in sample data, which can help achieve more effective anomaly or fault detection.

[0005] However, each set of samples in the dataset used to train the autoencoder or feature representation learning model has its own unique nearest-neighbor normal variation features. In the absence of supervision from abnormal or faulty condition sample data, only comprehensive learning of the nearest-neighbor variation features representing each normal sample data can avoid potential fault underreporting. Furthermore, from the perspective of feature learning through projection transformation, the features obtained from any form of transformation of normal data and its nearest-neighbor sample data are all normal and can represent different normal variation features. Therefore, comprehensive analysis and learning of the nearest-neighbor variation features of the normal sample data used for training, in order to construct an autoencoder describing the differences in various normal variations, is highly beneficial for improving the fault detection performance of the corresponding methods. Summary of the Invention

[0006] The main technical problem to be solved by this invention is: how to comprehensively analyze and learn the nearest neighbor variation features of normal sample data used for training, so as to construct an autoencoder that describes the differences of each normal variation, thereby achieving more effective fault detection.

[0007] The technical solution adopted by the present invention to solve the above problems is: a fault detection method and device based on a PLSR self-encoder; wherein, the fault detection method based on a PLSR self-encoder disclosed in the present invention includes the following steps 1 to 8.

[0008] Step 1: When the chemical process object is operating normally, obtain sample data at N sampling times to form a training data matrix; where N≥500, and each column vector in the training data matrix X is composed of sample data at each sampling time.

[0009] Step 2: Standardize each column vector in the training data matrix X, and then combine the standardized column vectors to form a data matrix.

[0010] Step 3: Data matrix Build a PLSR autoencoder, retaining the encoding matrix E and the decoding matrix D.

[0011] Step 4: Determine the threshold for coding fault indicators and the threshold of decoding fault indicators Then, determine the threshold for the comprehensive fault index.

[0012] Step 5: Obtain a set of sample data of the chemical process object at the latest sampling time t, forming a column vector x. t Then, for the column vector x t Perform the same standardization process as in step 2 to obtain the standardized column vector.

[0013] Step 6: According to the formula and Calculate the encoded feature vector Xs sequentially t and decoding error vector ε t Then, according to the formula and Calculate the coding fault index at the latest sampling time t. and decoding fault index θ t ; where the superscript T denotes the transpose of a matrix or vector.

[0014] Step 7: According to the formula Calculate the comprehensive fault index at the latest sampling time t Then, determine whether the conditions are met. If yes, the chemical process object is operating normally, and the process returns to step 5 to continue fault detection using a set of sample data from the latest sampling time; if no, step 8 is executed to determine whether the chemical process object has malfunctioned.

[0015] Step 8: Return to step 5 and continue to perform fault detection using a set of sample data from the latest sampling time until at least 5 comprehensive fault indicators are obtained. If none of them meet the judgment conditions in step 7, the chemical process object is operating malfunctioning and a fault alarm is triggered; otherwise, the chemical process object is operating normally.

[0016] In the above-mentioned implementation steps of a fault detection method based on a PLSR autoencoder, the core of the technical solution lies in the implementation process of building the PLSR autoencoder in step 3. The preferred implementation process of building the PLSR autoencoder in this invention includes a nearest neighbor change feature analysis stage and a PLSR analysis stage. The nearest neighbor change feature analysis stage includes steps 3-1 to 3-3 as shown below.

[0017] Step 3-1: After setting the number of nearest neighbors to K, the data matrix is... Each column vector in the matrix determines its corresponding nearest neighbor matrix; where K is an integer not less than 5 and not greater than 10.

[0018] Step 3-2: Perform nearest neighbor change feature analysis on each nearest neighbor matrix determined in step 3-1 to obtain the corresponding feature transformation vector.

[0019] Step 3-3: Combine all the feature transformation vectors obtained in step 3-2 to form the feature transformation matrix V.

[0020] After completing the nearest neighbor change feature analysis stage, the Partial Least Squares Regression (PLSR) algorithm is used to implement the PLSR analysis stage. The specific implementation process includes steps 3-4 to 3-5 as described below.

[0021] Steps 3-4: Calculate the feature score matrix using the feature transformation matrix V. Then, the PLSR algorithm is used to establish S and Regression model between Wherein, P, Q and E represent the loading matrix, regression matrix and error matrix in the regression model, respectively, and the superscript T indicates the transpose of the matrix or vector.

[0022] Steps 3-5: According to the formula Λ=P T SS T After calculating the covariance matrix Λ, P then calculates the covariance matrix Λ according to the formula E = VPΛ. -12 and D = I - QP T V T The encoding matrix E and the decoding matrix D are calculated; where I represents the identity matrix.

[0023] As a further improvement of the present invention, in step 3-2 of the nearest neighbor change feature analysis stage, the data matrix is ​​preferably processed according to steps A-1 to A-4 as shown below. Nearest neighbor change feature analysis is performed on the nearest neighbor matrix corresponding to each column vector; among them, and X i Representing data matrices respectively The i-th column vector and its corresponding nearest neighbor matrix are labeled i = 1, 2, ..., N.

[0024] Step A-1: ​​Set the feature vector u to any K. i A non-zero real vector of dimension ×1; where K i Equal to the nearest neighbor matrix X i The total number of column vectors in the middle.

[0025] Step A-2: Solving the generalized eigenvalue problem After finding the eigenvector v corresponding to the largest eigenvalue λ, then based on... The feature vector v is then updated.

[0026] Step A-3: Solve the eigenvalue problem X i T vv T X i After finding the eigenvector u corresponding to the largest eigenvalue η in u = ηu, then based on... The feature vector u is normalized.

[0027] Step A-4: Determine whether the feature vector u has converged; if not, return to step A-2; if yes, set the... The corresponding feature transformation vector is equal to v.

[0028] The iterative execution of steps A-2 and A-3 above aims to determine two feature vectors u and v, thereby... The corresponding nearest neighbor matrix X i The mathematical expression for projective transformation to a location far from the coordinate zero point is as follows: max||v T X i u|| 2 ①

[0029] In equation ① above, the objective function is ||v T X i u|| 2 =v T X i uu T X i T v = u T X i T vv T X i From a geometric perspective, the physical meaning of equation ① above lies in: u and v are used to map the nearest neighbor matrix X. i A projection transformation is performed to move the coordinates away from the zero point as far as possible.

[0030] To constrain the scale of the projection transformation in equation ① above, two scale constraints are added for the eigenvectors u and v: u T u=1 and Among them, the scale constraint condition u T =1 constrains the eigenvector u to a unit-length vector that only represents direction, while The normal sample data used for training can be projected and transformed into the interval [-1, 1].

[0031] Based on the principle of the Lagrange multiplier method, we first introduce the Lagrange multipliers λ and η, and then construct the Lagrange function. Calculate the partial derivatives with respect to u and v respectively, i.e.:

[0032] Setting equations ② and ③ above to 0 respectively, we can obtain... and X i T vv T X i u = ηu, which is the eigenvalue problem in steps A-2 and A-3; furthermore, for X i T vv T X i Multiply both sides of the equation u = ηu by u on the left. T We can obtain: u T X i T vv T X i u=ηu T u = η; therefore, η is equal to the objective function; similarly, it can be deduced that λ is also equal to the objective function; since the objective function needs to be maximized, the maximum eigenvalue needs to be calculated in both steps A-2 and A-3.

[0033] As a further improvement to this invention, and also in order to more comprehensively learn the nearest neighbor variation features of the normal sample data used for training, it is necessary to determine... The corresponding second feature transformation vector is implemented through steps B-1 to B-4 as shown below.

[0034] Step B-1: Set the feature vector Equal to any K i A non-zero real vector of dimension ×1.

[0035] Step B-2: Solve the generalized eigenvalue problem The eigenvector corresponding to the largest eigenvalue α in the vector Then, according to The eigenvector v is updated; where the constraint matrix... X j Representing a data matrix The nearest neighbor matrix corresponding to the j-th column vector in the matrix, j = 1, 2, ..., N and j ≠ i.

[0036] Step B-3: Solve the generalized eigenvalue problem The eigenvector corresponding to the largest eigenvalue β in the vector Then, according to For eigenvectors Perform update processing; where the constraint matrix

[0037] Step B-4: Determine the feature vector Has convergence been achieved? If not, return to step B-2; if yes, set the... The corresponding second feature transformation vector is equal to

[0038] The iterative execution of steps B-2 and B-3 above is to solve the maximization problem shown below:

[0039] Where st is an abbreviation for Subject To, representing the constraint condition; as can be seen from equation ④ above, its maximization objective is also achieved through two eigenvectors. and make Maximize, which is consistent with the objective in equation ① above; the difference from equation ① lies in the constraints. Its function is to divide by X i The remaining nearest neighbor matrices are projected and transformed into the interval [-1,1], and their sum is equal to 1.

[0040] Therefore, this improvement in the present invention focuses more on extracting the nearest neighbor transformation features that differ from other normal sample data in each group of normal sample data; and the solution of equation ④ above can also be achieved by introducing α and β using the Lagrange multiplier method, and then calculating the Lagrange function relative to the two eigenvectors respectively. and By taking the partial derivatives and setting them to 0, we can obtain the two generalized eigenvalue problems in steps B-2 and B-3.

[0041] As a further improvement to the present invention, and also for a deeper and more comprehensive learning of the nearest neighbor variation features of the normal sample data used for training, the following steps C-1 to C-3 can be further determined. The corresponding third and fourth feature transformation vectors.

[0042] Step C-1: Follow the formula sequentially. and Calculate the score vector y and the load vector p, and then apply the formula... and Calculate the score vector and load vector

[0043] Step C-2: Apply the formula V1 = I - vp respectively. T and Calculate transformation matrix V1 and transformation matrix V2.

[0044] Step C-3: Simultaneously execute steps A-5 and B-5 as shown below, thereby simultaneously obtaining the... The corresponding third and fourth feature transformation vectors.

[0045] Step A-5: According to formula X respectively i =V1 T X i and Update the nearest neighbor matrix X i and data matrix Then, repeat steps A-2 and A-3 until the feature vector u converges again, then set... The corresponding third feature transformation vector is equal to V1v.

[0046] Step B-5: According to formula X respectively i =V2 T X i and Update the nearest neighbor matrix X i and data matrix Then, repeat steps B-2 and B-3 until the feature vector is obtained. After convergence again, set The corresponding fourth feature transformation vector is equal to

[0047] As can be seen from this improved technical solution of the present invention, the third feature transformation vector and the first feature transformation vector determined in step A-4 both satisfy the above equation ① and its scale constraint conditions, while the fourth feature transformation vector and the second feature transformation vector satisfy the above equation ④. Therefore, this improved technical solution of the present invention can achieve a more in-depth and comprehensive learning of nearest neighbor change features based on the aforementioned improved solutions.

[0048] Furthermore, the transformation matrix V1 calculated in step C-2 above is used to ensure that the score vectors obtained after the projection transformations of the first and third feature transformation vectors are orthogonal to each other, thus avoiding the information redundancy problem in the nearest neighbor change feature analysis. Similarly, the transformation matrix V2 calculated in step C-2 is used to ensure that the score vectors corresponding to the second and fourth feature transformation vectors are orthogonal to each other, thereby avoiding the information redundancy problem.

[0049] As a further improvement of the present invention, step 3-1 is a data matrix. The i-th column vector in Determine the corresponding nearest neighbor matrix Xi It is preferable to use steps D-1 to D-2 as shown below.

[0050] Step D-1: Calculate the data matrix Each column vector in the middle and The squared Mahalanobis distance between them; where, the data matrix The j-th column vector in and Square Mahalanobis distance between and

[0051] Step D-2: After sorting the squared Mahalanobis distances calculated in Step D-1 in ascending order of their numerical values, the data matrix is... Zhongyu The K column vectors with the smallest square Mahalanobis distance between them are combined to form the... The corresponding nearest neighbor matrix X i .

[0052] As another improvement of the present invention, step 3-1 is a data matrix. The i-th column vector in Determine the corresponding nearest neighbor matrix X i Alternatively, steps E-1 to E-2 can be used as shown below.

[0053] Step E-1: Calculate the data matrix Each column vector in the middle and The cosine value of the angle between; wherein, the data matrix The j-th column vector in and The cosine value of the angle between and

[0054] Step E-2: After sorting the cosine values ​​calculated in Step E-1 in descending order of their numerical values, the data matrix is... Zhongyu The K column vectors with the largest cosine values ​​of the angle between them are combined to form the... The corresponding nearest neighbor matrix X i

[0055] As another improvement of the present invention, step 3-1 is a data matrix. The i-th column vector in Determine the corresponding nearest neighbor matrix X i Alternatively, steps F-1 through F-4 can be used as shown below.

[0056] Step F-1: Calculate the data matrix Each column vector in the middle and The square Mahalanobis distance and the cosine of the angle between them.

[0057] Step F-2: After sorting the squared Mahalanobis distances calculated in Step F-1 in ascending order of their numerical values, record the data matrix. Zhongyu The K column vectors that have the smallest squared Mahalanobis distance between each other.

[0058] Step F-3: After sorting the cosine values ​​of the angles calculated in step F-1 in descending order of their numerical values, record the data matrix. Zhongyu The K column vectors with the largest cosine values ​​of the angle between them.

[0059] Step F-4: After deleting the column vectors that are duplicated in step F-3 and in step F-2, merge the column vectors of all other records to form the final column vector. The corresponding nearest neighbor matrix X i .

[0060] This invention provides three different improved technical solutions for step 3-1. The first improved technical solution uses squared Mahalanobis distance to determine the nearest neighbor matrix, drawing on and utilizing the Mahalanobis distance's ability to consider the spatial distribution pattern of data and its advantage of eliminating interference from correlations between variables. The second improved technical solution uses the cosine value of the angle between vectors to determine the nearest neighbor matrix through the directional correlation of the data spatial distribution, which is more universal than using the distance between vectors. The third improved technical solution combines the nearest neighbor matrix determined by the first two methods and removes duplicate column vectors, which can both leverage the technical advantages of Mahalanobis distance and consider the universality of directional correlation. This is the most preferred implementation process for determining the nearest neighbor matrix disclosed in step 3-1 of this invention.

[0061] As another improvement of the present invention, the implementation process of determining the threshold in step 4 is preferably carried out using steps 4-1 to 4-2 as shown below.

[0062] Step 4-1, according to the formula and Calculate the coding fault index vector respectively and decoding fault indicator vector Then, The average of the ten largest elements in the code is used as the threshold for the coding fault indicator. And The average of the ten largest elements in the dataset is used as the threshold for decoding the fault indicator. Here, diag{} represents the operation of converting the elements on the diagonal of the matrix within the curly braces into column vectors.

[0063] Step 4-2, according to the formula Calculate the comprehensive fault index vector Then, The average of the ten largest elements is determined as the threshold for the comprehensive fault index.

[0064] Steps 4-1 to 4-2 above are further improvements to step 4 of the present invention, providing a technical solution that can objectively determine the threshold without requiring prior statistical distribution assumptions, and also avoiding interference caused by subjective human factors in determining the threshold.

[0065] Finally, based on the above-described fault detection implementation process, this invention also provides a fault detection device based on a PLSR self-encoder, comprising a data acquisition unit, a data processing unit, a monitoring and display unit, and a parameter storage unit; wherein, the data acquisition unit acquires sample data at the latest sampling time in real time according to the sampling time interval of the chemical process object, and transmits it to the data processing unit immediately; after receiving the sample data at the latest sampling time, the data processing unit first performs standardization processing on it, and then calculates the comprehensive fault index. Then, the comprehensive fault indicators will be... The specific values ​​are transmitted to the monitoring and display unit; the monitoring and display unit receives the comprehensive fault index. After obtaining the specific value, the data is displayed in real time with the sampling time as the horizontal axis and the specific value as the vertical axis, while simultaneously setting the threshold. The specific values ​​are displayed in horizontal lines.

[0066] The parameter storage unit is characterized in that the parameters stored specifically include the average and standard deviation required for implementing the standardization process, the encoding matrix E and decoding matrix D retained in step 3, and the threshold of the encoding fault index determined in step 4. Decoding the threshold of fault indicators and the threshold of comprehensive fault indicators The parameter storage unit stores the average value, standard deviation, encoding matrix E, decoding matrix D, and threshold values ​​for encoding fault indicators. and the threshold of decoding fault indicators The data is provided in real time to the data processing unit, and the parameter storage unit stores the threshold values ​​of the comprehensive fault indicators. The comprehensive fault indicators are provided to the monitoring and display unit in real time; the monitoring and display unit receives the comprehensive fault indicators. After obtaining the specific value, the data is displayed in real time with the sampling time as the horizontal axis and the specific value as the vertical axis, while simultaneously setting the threshold. The specific values ​​are displayed in horizontal lines.

[0067] Therefore, the fault detection device provided by the present invention aims to use the parameters determined in step 3 to monitor in real time whether the sample data at the latest sampling time is normal or faulty, and to display the monitoring results in real time through the monitoring display unit, thereby facilitating on-site operators of chemical process objects to monitor whether the operating status of the chemical process is normal. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the control system and measuring instruments of an atmospheric distillation tower for crude oil.

[0069] Figure 2 This is a schematic diagram illustrating the implementation process of the fault detection method involved in this invention.

[0070] Figure 3 This is a schematic diagram illustrating the implementation process of the nearest neighbor change feature analysis involved in this invention.

[0071] Figure 4 This is a schematic diagram illustrating the process of determining the nearest neighbor matrix involved in this invention.

[0072] Figure 5 This is a schematic diagram of the threshold determination process involved in the present invention.

[0073] Figure 6 This is a schematic diagram of the structural composition of the fault detection device involved in the present invention.

[0074] Figure 7 This is a detailed monitoring diagram displayed by the monitoring display unit involved in the present invention.

[0075] Figure 8 This is a detailed monitoring map obtained using the comparison method. Detailed Implementation

[0076] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0077] As a typical chemical process object, such as Figure 1 The crude oil atmospheric distillation tower shown is the first processing step in the crude oil processing production process. After desalting and dehydration, the crude oil enters the crude oil atmospheric distillation tower, where it undergoes effective and thorough fractionation. The crude oil atmospheric distillation tower not only provides semi-finished products such as atmospheric residue oil for a large number of subsequent processing processes, but also directly produces several products, including gasoline, aviation kerosene, light diesel oil, and heavy diesel oil. Since the operation of the crude oil atmospheric distillation tower requires a large amount of energy, timely and reliable detection of malfunctions during operation can provide direct assurance for production quality and energy consumption optimization.

[0078] This embodiment selects Figure 1The crude oil atmospheric distillation tower shown is used as an application object of the fault detection method and device, and should not be regarded as a limitation on the technical solution of the present invention. Within the scope of the technical solution and the protection of the claims of the present invention, other typical chemical process objects can also be selected as application embodiments.

[0079] like Figure 1 The crude oil atmospheric distillation column shown can collect sample data for a total of 20 measurement variables in real time at each sampling moment, namely: column top temperature Tower top return flow Condenser liquid level Side line temperature Side-line circulation flow Side two-line temperature Side-second circulation flow Side three-line temperature Side three-line circulation flow Feed temperature Commonly drawn line temperature Standard stripping tower liquid level Aviation kerosene extraction flow rate Normal second line extraction temperature Level of the second-stage stripping tower Light diesel oil extraction flow rate Normal three-wire extraction temperature Level of the three-stage stripping tower Heavy diesel oil extraction flow rate Bottom liquid level Therefore, the sample data obtained at each sampling time can be organized into a 20×1 dimensional column vector.

[0080] According to such Figure 2 The implementation process of the fault detection method shown is to perform fault detection on the crude oil atmospheric distillation tower, including steps 1 to 8 as shown below.

[0081] Step 1: When the crude oil atmospheric distillation tower is operating normally, obtain sample data at least 500 sampling times to form a training data matrix X.

[0082] Step 2: Standardize each column vector in the training data matrix X, and then combine the standardized column vectors to form a data matrix.

[0083] Step 3: For the data matrix A PLSR autoencoder is constructed, retaining the encoding matrix E and the decoding matrix D. The specific implementation process includes the nearest neighbor change feature analysis stage and the PLSR analysis stage in sequence; among them, the nearest neighbor change feature analysis stage includes steps 3-1 to 3-3 as shown below.

[0084] Step 3-1: After setting the number of nearest neighbors to 5, the data matrix is... Each column vector in the matrix determines its corresponding nearest neighbor matrix.

[0085] Step 3-2: Perform nearest neighbor change feature analysis on each nearest neighbor matrix determined in step 3-1 to obtain the corresponding feature transformation vector.

[0086] Step 3-3: Combine all the feature transformation vectors obtained in step 3-2 to form the feature transformation matrix V.

[0087] After the nearest neighbor change feature analysis phase is completed, the PLSR analysis phase consisting of steps 3-4 to 3-5 as shown below needs to be performed.

[0088] Steps 3-4: Calculate the feature score matrix using the feature transformation matrix V. Then, the PLSR algorithm is used to establish S and Regression model between

[0089] Steps 3-5: According to the formula Λ=P T SS T After calculating the covariance matrix Λ, P then calculates the covariance matrix Λ according to the formula E = VPΛ. -12 and D = I - QP T V T The encoding matrix E and the decoding matrix D are calculated.

[0090] Step 4: Determine the threshold for coding fault indicators and the threshold of decoding fault indicators Then, determine the threshold for the comprehensive fault index.

[0091] Step 5: Obtain a set of sample data from the crude oil atmospheric distillation column at the latest sampling time t, forming a column vector x. t Then, for the column vector x t Perform the same standardization process as in step 2 to obtain the standardized column vector.

[0092] Step 6: According to the formula and Calculate the encoded feature vector s sequentially t and decoding error vector ε t Then, according to the formula and Calculate the coding fault index at the latest sampling time t. and decoding fault index θ t .

[0093] Step 7: According to the formula Calculate the comprehensive fault index at the latest sampling time t Then, determine whether the conditions are met. If yes, the crude oil atmospheric distillation tower is operating normally, and we return to step 5 to continue fault detection using a set of sample data from the latest sampling time; if no, we proceed to step 8 to determine whether a fault has occurred in the crude oil atmospheric distillation tower.

[0094] Step 8: Return to step 5 and continue to use a set of sample data from the latest sampling time to perform fault detection until at least 5 comprehensive fault indicators are obtained. If none of them meet the judgment conditions in step 7, the crude oil atmospheric distillation tower is operating normally and a fault alarm is triggered; otherwise, the crude oil atmospheric distillation tower is operating normally.

[0095] exist Figure 3 Figures A, B, and C respectively disclose the second, third, and fourth embodiments of the present invention, which respectively employ three different nearest-neighbor change feature analysis techniques disclosed in the present invention.

[0096] exist Figure 3 The implementation process of the nearest neighbor change feature analysis selected in Figure A is to set i = 1, 2, ..., N in sequence and execute steps A-1 to A-4 simultaneously, thereby completing the task of performing nearest neighbor change feature analysis on each nearest neighbor matrix determined in step 3-1.

[0097] exist Figure 3 The implementation process of the nearest neighbor change feature analysis selected in Figure B is as follows: Figure 3 Based on the nearest neighbor change feature analysis shown in Figure A, the second feature transformation vector is determined for each nearest neighbor matrix by sequentially setting i = 1, 2, ..., N and simultaneously executing steps B-1 to B-4.

[0098] exist Figure 3 The implementation process of the nearest neighbor change feature analysis selected in the C-plot is as follows: Figure 3 Based on the nearest neighbor change feature analysis shown in Figure B, further steps are taken by sequentially setting i = 1, 2, ..., N and simultaneously executing steps C-1 to C-3, thereby further determining the third and fourth feature transformation vectors for each nearest neighbor matrix.

[0099] like Figure 4 The implementation process shown discloses the improved technical solution of the present invention for determining the nearest neighbor matrix in step 3-1; wherein, Figure 4 In Figure A, the squared Mahalanobis distance is used to determine the nearest neighbor matrix, specifically including steps D-1 to D-2. Figure 4In Figure B, the cosine of the angle is used to determine the nearest neighbor matrix, specifically including steps E-1 to E-2; while Figure 4 The C-graph in the diagram uses both the squared Mahalanobis distance and the cosine of the angle to jointly determine the nearest neighbor matrix, specifically including steps F-1 to F-4.

[0100] like Figure 5 The implementation process shown discloses the threshold determination process preferred by the present invention. A suitable threshold can be objectively determined through steps 4-1 and 4-2 for subsequent identification of whether a fault has occurred.

[0101] Based on the implementation process of steps 5 to 7 above, Figure 6 The invention discloses a fault detection device, including a data acquisition unit, a data processing unit, a monitoring and display unit, and a parameter storage unit; wherein, the data acquisition unit acquires sample data corresponding to 20 measurement variables at the latest sampling time in real time according to the sampling time interval of the chemical process object, and transmits it to the data processing unit in real time.

[0102] like Figure 6 In the fault detection device shown, after the data processing unit receives sample data corresponding to 20 measurement variables, it first performs standardization processing on the data and then calculates the comprehensive fault index. Then, the comprehensive fault indicators will be... The specific values ​​are transmitted to the monitoring and display unit.

[0103] When the data processing unit performs standardization on the received sample data, it needs to immediately access the mean and standard deviation stored in the parameter storage unit for standardization, and then calculate the comprehensive fault index according to the implementation process of step 6. When determining the specific values, it is necessary to immediately call the encoding matrix E and decoding matrix D determined in step 3 above, and the threshold of the encoding fault index in step 4 above, which are stored in the parameter storage unit. and the threshold of decoding fault indicators

[0104] exist Figure 6 In the fault detection device shown, the monitoring and display unit receives comprehensive fault indicators. After obtaining the specific value, the data is displayed in real time with the sampling time as the horizontal axis and the specific value as the vertical axis, while simultaneously setting the threshold. The specific values ​​are displayed in horizontal lines.

[0105] The data acquisition unit acquires 1000 sets of sample data at 1000 recent sampling times from the crude oil atmospheric distillation tower in real time. After processing by the data processing unit, the data is then displayed in the monitoring and display unit. The corresponding display results are as follows: Figure 7 As shown.

[0106] exist Figure 7 In the diagram, the horizontal axis represents the sampling time, numbered sequentially according to the latest sampling time since the fault detection device was officially put into use. The vertical axis displays the specific values ​​of the comprehensive fault index at the corresponding sampling time using a scatter plot method. Figure 5 The horizontal line in the middle represents the threshold. The specific value is provided in real time by the parameter storage unit. Since it has been determined in step 4 above and will not change, the horizontal line displayed in the monitoring display unit is used.

[0107] like Figure 7 As shown in the detailed monitoring chart, when the scatter plot of the comprehensive fault index is located at or below the horizontal line, it indicates that the crude oil atmospheric distillation tower is operating normally. If the scatter plot of the comprehensive fault index is above the horizontal line for multiple consecutive sampling times, it can be determined that the crude oil atmospheric distillation tower has entered a fault state.

[0108] Therefore, from Figure 7 It can be observed that the crude oil atmospheric distillation process entered a faulty operating state at 200 sampling points, and the scatter plot of the comprehensive fault indicators was continuously displayed horizontally until the fault disappeared. However, when fault detection was performed on the same sample data according to the comparative method, the scatter plot of the monitoring indicators used for fault detection was as follows: Figure 8 As shown, from Figure 8 It can be clearly seen that the comparative method has a higher rate of missed fault detections; in contrast, the present invention can further improve fault detection performance and reduce missed fault detections.

[0109] The above explanation of the technical solutions disclosed in this invention in conjunction with the accompanying drawings is illustrative rather than restrictive. The technical solutions disclosed in this invention are not limited to the above-mentioned crude oil atmospheric distillation tower as a chemical process object and its corresponding specific embodiments. Modifications made without departing from the spirit of the disclosed technical solutions and the scope of protection of the claims shall fall within the protection scope of this invention.

Claims

1. A fault detection method based on a PLSR self-encoder, comprising steps 1 to 8 as shown below: Step 1: When the chemical process object is operating normally, obtain... The sample data at each sampling time point form the training data matrix. ;in, The training data matrix Each column vector in the array is composed of sample data from each sampling time. Step 2: Process the training data matrix Each column vector in the matrix is ​​standardized, and then the standardized column vectors are combined to form a data matrix. ; Step 3: For the data matrix Build a PLSR autoencoder and preserve the encoding matrix. and decoding matrix ; Step 4: Determine the threshold for coding fault indicators and the threshold of decoding fault indicators Then, determine the threshold for the comprehensive fault index. ; Step 5: Obtain the chemical process object at the latest sampling time. A set of sample data forms a column vector Then, for the column vector Perform the same standardization process as in step 2 to obtain the standardized column vector. ; Step 6: According to the formula and Calculate the encoded feature vectors sequentially and decoding error vector Then, according to the formula and Calculate the latest sampling time respectively Coding Fault Indicators and decoding fault indicators ; where the superscript T denotes the transpose of a matrix or vector; Step 7: According to the formula Calculate the latest sampling time Comprehensive Fault Indicators Then, determine whether the conditions are met. If yes, the chemical process object is operating normally, and the process returns to step 5 to continue using a set of sample data from the latest sampling time to perform fault detection; if no, step 8 is executed to determine whether the chemical process object has experienced a fault. Step 8: Return to step 5 and continue to use a set of sample data from the latest sampling time to perform fault detection until at least 5 comprehensive fault indicators are obtained. If none of them meet the judgment conditions in step 7, the chemical process object is operating normally and a fault alarm is triggered; otherwise, the chemical process object is operating normally. The characteristic is that the specific implementation process of step 3 includes, in sequence, a nearest neighbor change feature analysis stage and a PLSR analysis stage; wherein, the nearest neighbor change feature analysis stage includes steps 3-1 to 3-3 as shown below: Step 3-1: Set the number of nearest neighbors to equal Then, the data matrix. Each column vector in the matrix determines its corresponding nearest neighbor matrix; where, It is an integer that is not less than 5 and not greater than 10; Step 3-2: Perform nearest neighbor transformation feature analysis on each nearest neighbor matrix determined in Step 3-1 to obtain the corresponding feature transformation vector; wherein, for the data matrix The first in column vector The corresponding nearest neighbor matrix The specific process of performing nearest neighbor change feature analysis includes steps A-1 to A-4 as shown below, where ; Step A-1: ​​Set the feature vector Equal to any one A non-zero real vector of dimension; where, Equal to the nearest neighbor matrix The total number of column vectors in the array; Step A-2: Solving the generalized eigenvalue problem The largest eigenvalue in corresponding feature vector Then, according to For the feature vector Perform the update process; Step A-3: Solving the eigenvalue problem The largest eigenvalue in corresponding feature vector Then, according to For the feature vector Perform normalization processing; Step A-4: Determine the feature vector Has convergence occurred? If not, return to step A-2; if yes, set the... The corresponding feature transformation vector is equal to ; Step 3-3: Combine all the feature transformation vectors obtained in Step 3-2 to form a feature transformation matrix. ; The PLSR analysis phase consists of the following steps 3-4 to 3-5: Steps 3-4: Utilize the aforementioned feature transformation matrix Calculate the feature score matrix Then, the PLSR algorithm is used to establish... and Regression model between ;in, , and These represent the loading matrix, regression matrix, and error matrix in the regression model, respectively, with the superscript T indicating the transpose of the matrix or vector; Steps 3-5: According to the formula Calculate the covariance matrix Then, follow the formulas respectively. and The encoding matrix is ​​calculated. and decoding matrix ;in, Represents the identity matrix.

2. The fault detection method based on a PLSR self-encoder according to claim 1, characterized in that, For the nearest neighbor matrix The specific process of performing nearest neighbor change feature analysis also includes steps B-1 to B-4 as shown below: Step B-1: Set the feature vector Equal to any one A non-zero real vector of dimension; Step B-2: Solve the generalized eigenvalue problem The largest eigenvalue in corresponding feature vector Then, according to For the feature vector Perform update processing; where the constraint matrix , Represents the data matrix The first in The nearest neighbor matrix corresponding to the column vector, and ; Step B-3: Solve the generalized eigenvalue problem The largest eigenvalue in corresponding feature vector Then, according to For the feature vector Perform update processing; where the constraint matrix ; Step B-4: Determine the feature vector Has convergence occurred? If not, return to step B-2; if yes, set the... The corresponding second feature transformation vector is equal to .

3. The fault detection method based on a PLSR self-encoder according to claim 2, characterized in that, For the nearest neighbor matrix The specific process of performing nearest neighbor change feature analysis also includes the following steps C-1 to C-3: Step C-1: Follow the formula sequentially. and Calculate the score vector and load vector And according to the formula and Calculate the score vector and load vector ; Step C-2, follow the formula respectively and Calculate the transformation matrix and transformation matrix ; Step C-3: Simultaneously execute steps A-5 and B-5 as shown below, thereby simultaneously obtaining the... The corresponding third and fourth feature transformation vectors; Step A-5: According to the formula respectively and Update the nearest neighbor matrix and the data matrix Then, repeat steps A-2 and A-3 until the feature vector is obtained. After convergence again, set the... The corresponding third feature transformation vector is equal to ; Step B-5: According to the formula respectively and Update the nearest neighbor matrix and the data matrix Then, repeat steps B-2 and B-3 until the feature vector is obtained. After convergence again, set the... The corresponding fourth feature transformation vector is equal to .

4. A fault detection method based on a PLSR self-encoder according to any one of claims 1 to 3, characterized in that, The data matrix in step 3-1 The first in column vector Determine the corresponding nearest neighbor matrix The specific steps are as follows: Step D-1: Calculate the data matrix Each column vector in the middle and The squared Mahalanobis distance between them; wherein, the data matrix The first in column vector and Square Mahalanobis distance between and , ; Step D-2: After sorting the squared Mahalanobis distances calculated in Step D-1 in ascending order of their numerical values, the data matrix is... Zhongyu The square Mahalanobis distance between them is the smallest Combining column vectors to form The corresponding nearest neighbor matrix .

5. A fault detection method based on a PLSR self-encoder according to any one of claims 1 to 3, characterized in that, The data matrix in step 3-1 The first in column vector Determine the corresponding nearest neighbor matrix The specific steps are as follows: Step E-1: Calculate the data matrix Each column vector in the middle and The cosine value of the angle between; wherein, the data matrix The first in column vector and The cosine value of the angle between and , , ; Step E-2: After sorting the cosine values ​​calculated in Step E-1 in descending order of their numerical values, the data matrix is... Zhongyu The cosine of the angle between them is the largest Combining column vectors to form The corresponding nearest neighbor matrix .

6. A fault detection method based on a PLSR self-encoder according to any one of claims 1 to 3, characterized in that, The data matrix in step 3-1 The first in column vector Determine the corresponding nearest neighbor matrix The specific steps are as follows: Step F-1: Calculate the data matrix Each column vector in the middle and The squared Mahalanobis distance and the cosine of the angle between them; Step F-2: After sorting the squared Mahalanobis distances calculated in Step F-1 in ascending order of their numerical values, record the data matrix. Zhongyu The square Mahalanobis distance between them is the smallest Column vectors; Step F-3: After sorting the cosine values ​​of the angles calculated in step F-1 in descending order of their numerical values, record the data matrix. Zhongyu The cosine of the angle between them is the largest Column vectors; Step F-4: After deleting the column vectors that are duplicated in step F-3 and in step F-2, merge the column vectors of all other records to form the final column vector. The corresponding nearest neighbor matrix .

7. A fault detection method based on a PLSR self-encoder according to any one of claims 1 to 3, characterized in that, The implementation process of step 4 includes the following steps 4-1 to 4-2: Step 4-1, according to the formula and Calculate the coding fault index vector respectively and decoding fault indicator vector Then, The average of the ten largest elements in the code is used as the threshold for the coding fault indicator. and will The average of the ten largest elements in the dataset is used as the threshold for decoding the fault indicator. ;in, This indicates the operation of converting the elements on the diagonal of the matrix within the curly braces into a column vector; Step 4-2, according to the formula Calculate the comprehensive fault index vector Then, The average of the ten largest elements is determined as the threshold for the comprehensive fault index. .

8. A fault detection device based on a PLSR self-encoder, comprising a data acquisition unit, a data processing unit, a monitoring and display unit, and a parameter storage unit; The data acquisition unit acquires the latest sample data in real time according to the sampling time interval of the chemical process object, and transmits it to the data processing unit immediately. After receiving the sample data from the latest sampling time, the data processing unit first performs standardization processing on the data, and then calculates the comprehensive fault index. Then, the comprehensive fault indicators will be... The specific values ​​are transmitted to the monitoring and display unit; The monitoring and display unit receives the comprehensive fault indicators. After obtaining the specific value, the data is displayed in real time with the sampling time as the horizontal axis and the specific value as the vertical axis, while simultaneously setting the threshold. The specific values ​​are displayed as horizontal lines; Its features are, The parameters stored in the parameter storage unit specifically include the average and standard deviation required for implementing the standardization process, and the encoding matrix determined according to steps 1 to 4 of claim 1. Decoding matrix Thresholds for coding fault indicators Decoding the threshold of fault indicators and the threshold of comprehensive fault indicators ; The parameter storage unit stores the average value, standard deviation, and encoding matrix. Decoding matrix Thresholds for coding fault indicators and the threshold of decoding fault indicators The data is provided in real time to the data processing unit, and the parameter storage unit stores the threshold values ​​of the comprehensive fault indicators. It is provided to the monitoring and display unit in real time.

Images