Fault mode identification diagnosis method for castor oil hydrogenation reactor

Abstract

The application provides a castor oil hydrogenation reaction kettle fault mode identification diagnosis method, the method comprises collecting the reaction kettle operation state monitoring parameter, constructing a multidimensional time sequence characteristic matrix, and dividing the main dimension and the auxiliary dimension according to the synergistic correlation between each dimension; the cross covariance matrix is constructed according to the main and auxiliary dimension data and singular value decomposition is carried out, and the main dimension data is mapped to a low-dimensional subspace; based on the composite folding scaling factor, the manifold topology is reconstructed, a two-dimensional folding characteristic vector is obtained, and the weighted Chebyshev distance and the dynamic threshold are combined to carry out fault mode matching. Through the main and auxiliary division, low-dimensional mapping and manifold reconstruction of the multidimensional monitoring parameter, the stability and reliability of the fault mode identification can be improved.

Description

Technical Field

[0001] This application belongs to the field of identification and diagnosis, and in particular relates to a method for identifying and diagnosing fault modes in a castor oil hydrogenation reactor. Background Technology

[0002] During the monitoring of reactor operation, it is typically necessary to collect real-time data from multiple sensors, including temperature, pressure, stirring speed, and hydrogen flow rate. These monitoring parameters collectively constitute a temporal feature space reflecting the reactor's operating status. To extract effective fault characterization information from the multidimensional monitoring data, the industry usually requires dimensionality reduction and feature extraction of the high-dimensional time-series data.

[0003] However, existing data preprocessing methods, when faced with high-dimensional feature matrices, often employ a globally uniform dimensionality reduction strategy, lacking in-depth analysis of the synergistic relationships between different monitoring dimensions. This makes it difficult to prioritize each dimension based on its contribution to the fault state representation. Consequently, redundant data and environmental noise can easily be mixed into the dimensionality-reduced feature set, weakening core fault information that reflects early, minor equipment defects and increasing the computational burden of subsequent pattern matching, leading to a decrease in the recognizability of early fault features.

[0004] In the process of mapping high-dimensional features to a low-dimensional subspace, conventional mapping algorithms easily disrupt the original local manifold topology of the data, distorting the geometric distance relationships between data points before and after dimensionality reduction and affecting the expression of fault evolution patterns. In the fault mode determination stage, traditional methods often use conventional distance algorithms for similarity calculation, failing to incorporate differentiated measurement based on the importance of different feature directions after dimensionality reduction; their determination thresholds also rely heavily on empirically set fixed values, lacking an adaptive adjustment mechanism based on the data topology mapping state. Therefore, under actual reaction conditions with disturbances, existing methods are prone to blurred boundaries in identifying similar fault modes, increasing the risk of misdiagnosis and missed diagnosis, and failing to meet the engineering requirements for refined fault mode identification in castor oil hydrogenation reactors. Summary of the Invention

[0005] To address the problems in existing technologies such as insufficient distinction between primary and secondary dimensions in high-dimensional monitoring data, easy disruption of manifold topology during dimensionality reduction, lack of weighted differentiation in fault distance metrics, and difficulty in adaptively adjusting judgment thresholds, according to one aspect of the present invention, a fault mode recognition and diagnosis method for castor oil hydrogenation reactors is provided, the method comprising:

[0006] A multidimensional time-series feature matrix was constructed by collecting status monitoring parameters of the castor oil hydrogenation reactor; the synergistic correlation coefficient matrix between dimensions was calculated using orthogonal vector projection, the dimensional contribution weight vector was extracted, and the multidimensional features were divided into primary and secondary dimensions. Extract the main dimension and auxiliary dimension data, calculate the cross covariance matrix and perform singular value decomposition to obtain the orthogonal basis eigenvectors corresponding to the singular value sequence and the main dimension; use the orthogonal basis eigenvectors corresponding to the main dimension to determine the mapping relationship, fold the main dimension data to a low-dimensional subspace, and calculate the folded projection distance of the data points; The singular value energy concentration is calculated based on the singular value sequence. A composite folding scaling factor is constructed by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance. The manifold topology is reconstructed on the low-dimensional subspace based on the composite folding scaling factor to obtain the two-dimensional folding feature vector. The ratio of the geometric distance before and after reconstruction is calculated as the manifold metric preservation coefficient. Based on the eigenvalue weight sequence of manifold topology reconstruction, the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library is calculated; the initial fault judgment threshold is corrected by the manifold metric preservation coefficient to obtain the dynamic threshold; when the weighted Chebyshev distance is less than the dynamic threshold, the corresponding fault mode diagnosis result is output.

[0007] According to another aspect of the present invention, a fault mode identification and diagnostic system for a castor oil hydrogenation reactor is provided, the system comprising the following modules: The module is used to collect the status monitoring parameters of the castor oil hydrogenation reactor to construct a multi-dimensional time-series feature matrix; the synergistic correlation coefficient matrix between dimensions is calculated using orthogonal vector projection, the dimensional contribution weight vector is extracted, and the multi-dimensional features are divided into main dimensions and auxiliary dimensions. The calculation module is used to extract the main dimension and auxiliary dimension data, calculate the cross covariance matrix and perform singular value decomposition to obtain the orthogonal basis eigenvectors corresponding to the singular value sequence and the main dimension; use the orthogonal basis eigenvectors corresponding to the main dimension to determine the mapping relationship, fold the main dimension data to a low-dimensional subspace, and calculate the folded projection distance of the data points. The reconstruction module is used to calculate the singular value energy concentration based on the singular value sequence, construct a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance, and perform manifold topology reconstruction on the low-dimensional subspace based on the composite folding scaling factor to obtain a two-dimensional folded feature vector, and calculate the ratio of the geometric distance before and after reconstruction as the manifold metric preservation coefficient. The diagnostic module is used to calculate the weight sequence of eigenvalues ​​based on manifold topology reconstruction, calculate the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library, correct the initial fault judgment threshold using the manifold metric preservation coefficient to obtain the dynamic threshold, and output the corresponding fault mode diagnostic result when the weighted Chebyshev distance is less than the dynamic threshold.

[0008] As can be seen from the above technical solutions, the castor oil hydrogenation reactor fault mode identification and diagnosis method provided in this specification constructs a multi-dimensional time-series feature matrix by collecting status monitoring parameters such as reactor temperature, pressure, hydrogen flow rate, and stirring speed, and divides the matrix into main and auxiliary dimensions based on co-correlation relationships. This approach considers the correlation between different monitoring parameters while reducing the impact of redundant data and environmental noise on fault identification. Furthermore, two-dimensional folded features are extracted through cross-covariance matrix, singular value decomposition, and manifold topology reconstruction, and fault mode matching is performed using weighted Chebyshev distance and dynamic thresholds, thereby improving the stability and reliability of reactor fault identification. Attached Figure Description

[0009] Figure 1 A flowchart of a fault mode identification and diagnosis method for a castor oil hydrogenation reactor; Figure 2 This is a flowchart of the overall process flow of the reactor; Figure 3 This is a schematic diagram of a high-dimensional reduction topological transformation. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] Example 1 In Embodiment 1 of the present invention, a fault mode identification and diagnosis method for a castor oil hydrogenation reactor is provided, such as... Figure 1 As shown, it includes: S1. Collect the status monitoring parameters of the castor oil hydrogenation reactor to construct a multi-dimensional time-series feature matrix; use orthogonal vector projection to calculate the synergistic correlation coefficient matrix between dimensions, extract the dimensional contribution weight vector, and divide the multi-dimensional features into main dimensions and auxiliary dimensions.

[0012] Before performing real-time fault mode identification, offline calibration is first performed based on a historical training sample set, which includes normal operation samples and fault samples corresponding to each known fault mode. During offline calibration, the standardized mean and standardized standard deviation of each state monitoring parameter are calculated, and the co-correlation coefficient matrix, dimensional contribution weight vector, main dimension set, auxiliary dimension set, spatial mapping matrix, fractal coordinate scaling coefficient, and fault mode benchmark library are calculated based on the historical training sample set. The standardized parameters, main dimension set, auxiliary dimension set, and spatial mapping matrix obtained from the offline calibration are used in the real-time identification stage.

[0013] Temperature, pressure, hydrogen flow rate, and stirring speed data within the castor oil hydrogenation reactor were collected using temperature sensors, pressure sensors, flow meters, and stirring speed sensors deployed on the reactor. Continuous sampling was performed at a fixed sampling frequency to acquire data from multiple sampling times. The values ​​of different physical quantities at the same sampling time were used as row vectors, and the row vectors corresponding to different sampling times were arranged in chronological order to construct a multidimensional time-series feature matrix of size N×m, where N is the number of sampling times and m is the number of dimensions of the status monitoring parameters.

[0014] Z-score standardization is applied to the multidimensional time-series feature matrix to reduce the influence of dimensions. This standardization is performed using the standardized mean and standardized standard deviation calculated from the historical training sample set during the offline calibration phase. Any two dimension vectors in the multidimensional time-series feature matrix are selected, and the inner product of the two dimension vectors and the product of their Euclidean norms are calculated. The ratio of the inner product to the norm product is used as the orthogonal projection correlation coefficient between the corresponding two dimensions. A symmetric co-correlation coefficient matrix is ​​constructed by traversing all dimension combinations in the multidimensional time-series feature matrix.

[0015] In the offline calibration phase, the absolute values ​​of the correlation coefficients within the corresponding rows of each feature dimension in the co-correlation coefficient matrix are calculated and accumulated to obtain the total absolute correlation value of each feature dimension. This total absolute correlation value is then normalized to obtain a dimension contribution weight vector. The state monitoring parameters are sorted in descending order of dimension contribution weight, and in this embodiment, the two top-ranked dimensions are selected as primary dimensions, with the remaining two dimensions designated as secondary dimensions. In the real-time identification phase, data is extracted from the primary and secondary dimension sets obtained from the offline calibration. The overall diagnostic process logic is as follows: Figure 2 As shown.

[0016] In some embodiments, the calculation of the synergistic correlation coefficient matrix between dimensions using orthogonal vector projection includes: Two columns of data with different column numbers in the multidimensional time series feature matrix are selected sequentially to form a paired column vector combination; The inner product of two vectors within a paired column vector combination is used as the correlation numerator. Calculate the Euclidean norm of each vector separately and multiply them together to obtain the result as the denominator of the correlation. The orthogonal projection correlation coefficient between the two corresponding dimensions is calculated by dividing the correlation numerator by the correlation denominator. Traverse the column vector combinations with different column numbers in the multidimensional time series feature matrix, calculate the corresponding orthogonal projection correlation coefficients, and arrange the orthogonal projection correlation coefficients into a symmetric matrix according to the feature dimension correspondence. At the same time, set the self-correlation coefficient corresponding to the data in the same column to 1 to obtain the co-correlation coefficient matrix.

[0017] During the status monitoring of the castor oil hydrogenation reactor, data at N time points across m monitoring dimensions (temperature, pressure, stirring speed, and hydrogen flow rate) are collected according to a set sampling frequency, resulting in a multidimensional time-series feature matrix X of size N×m. Let the column vectors in X... and This forms a pairing combination, where i and j are unequal column numbers. During orthogonal vector projection calculations, the temperature dimension vector is extracted, for example. and pressure dimension vector First, perform a vector inner product operation on these two dimensions to obtain the correlation numerator. Calculate the vectors respectively. and Euclidean norm and Multiply them to obtain the correlation denominator .pass The correlation coefficient of the two-dimensional orthogonal projection is calculated. When i=j, the corresponding autocorrelation coefficient Set to 1.

[0018] For example, if the extracted temperature feature dimension vector over a certain period of time... Pressure feature dimension vector The sum of the product of corresponding elements of the two vectors is calculated as the inner product, which is 316.96, and this is used as the numerator of the correlation coefficient. The Euclidean norms of the two vectors are calculated separately: the norm of the temperature vector is 140.75, and the norm of the pressure vector is 2.26. Their product is used as the denominator of the correlation coefficient, which is 318.09. The orthogonal projection correlation coefficient between these two dimensions is then calculated. ≈0.996. Repeat the above operation for all paired column vectors in the matrix, and define the correlation coefficient between themselves as 1. Arrange all m×m elements in order according to the original dimension row and column number correspondence to generate a symmetric matrix of dimension m×m, that is, the co-correlation coefficient matrix.

[0019] In some embodiments, the extraction of dimensional contribution weight vectors and the division of multidimensional features into primary and secondary dimensions includes: Calculate the absolute value of all values ​​in the corresponding row of each feature dimension in the co-correlation coefficient matrix, and sum the absolute values ​​of the same row to obtain the total absolute correlation value of the feature dimension; Divide the total absolute correlation value of a single feature dimension by the sum of the total absolute correlation values ​​of all feature dimensions to obtain the normalized contribution weight of each feature dimension, and arrange them in order to generate a dimension contribution weight vector. The feature dimensions are sorted in descending order of the normalized contribution weights. The two feature dimensions with the highest ranking are assigned to the primary dimension set, and the remaining two feature dimensions are assigned to the secondary dimension set.

[0020] After obtaining the co-correlation coefficient matrix of size m×m, traverse each row i of the matrix and extract the m elements in the row. For the i-th feature dimension, calculate the absolute value of all correlation coefficients within its row, and then sum them to obtain the total absolute correlation value. After obtaining the total absolute correlation values ​​for all m dimensions, these m values ​​are globally summed to obtain the final sum. Using the formula Calculate the normalized contribution weights for each dimension, and arrange them sequentially according to the column numbers of the original multidimensional time series feature matrix to generate a dimension contribution weight vector of length m. .

[0021] The components of the dimensional contribution weight vector W are extracted and sorted in descending order of weight value. In this embodiment, with four monitoring dimensions—temperature, pressure, stirring speed, and hydrogen flow rate—the two top-ranked dimensions are selected as the primary dimension set, and the remaining two dimensions are selected as the secondary dimension set. Assuming a system monitors four feature dimensions with absolute correlation values ​​of 45, 35, 15, and 5 respectively, the normalized contribution weight vector is calculated as W = [0.45, 0.35, 0.15, 0.05]. The sorting shows that the first and second dimensions have higher weights, therefore their corresponding dimensions are assigned to the primary dimension set; the third and fourth dimensions are assigned to the secondary dimension set, thus completing the primary and secondary dimension division in the offline calibration stage.

[0022] S2, extract the main dimension and auxiliary dimension data, calculate the cross covariance matrix and perform singular value decomposition to obtain the orthogonal basis eigenvectors corresponding to the singular value sequence and the main dimension; use the orthogonal basis eigenvectors corresponding to the main dimension to determine the mapping relationship, fold the main dimension data to the low-dimensional subspace, and calculate the folded projection distance of the data points.

[0023] During the offline calibration phase, data columns corresponding to the main dimension and data columns corresponding to the auxiliary dimension are extracted from the standardized historical training sample set to form the main dimension data matrix and the auxiliary dimension data matrix, respectively. The transpose of the main dimension data matrix is ​​multiplied by the auxiliary dimension data matrix, and the result of the multiplication is divided by the total number of sampling points minus 1 to obtain the cross covariance matrix.

[0024] Singular value decomposition is performed on the cross covariance matrix to obtain a left singular matrix, a singular value diagonal matrix, and a right singular matrix; the diagonal elements in the singular value diagonal matrix are extracted to form a singular value sequence, and the column vectors in the left singular matrix are used as orthogonal basis eigenvectors corresponding to the principal dimensions.

[0025] According to the order of singular values ​​from largest to smallest, the orthogonal basis eigenvectors corresponding to the first two valid singular values ​​are selected, and the orthogonal basis eigenvectors are concatenated column by column to form a fixed spatial mapping matrix. In the real-time recognition stage, the main dimension data matrix in the current sampling window is multiplied by the fixed spatial mapping matrix to map the main dimension data to a two-dimensional low-dimensional subspace, thus obtaining a low-dimensional subspace data matrix.

[0026] For each data point in the low-dimensional subspace data matrix, calculate the squared Euclidean distance between its coordinate vector and the origin, and use the squared Euclidean distance as the folded projection distance of the corresponding data point.

[0027] In some embodiments, determining the mapping relationship using the orthogonal basis eigenvectors corresponding to the principal dimension, folding the principal dimension data to a low-dimensional subspace, and calculating the folded projection distance of the data points includes: Extract the orthogonal basis eigenvectors corresponding to the principal dimensions associated with the first two valid singular values ​​in descending order of singular values ​​to construct a spatial mapping matrix, and save the spatial mapping matrix as an offline calibration parameter. Perform matrix multiplication between the main dimension data in the real-time sampling window and the spatial mapping matrix to obtain the set of mapped data point coordinates for each data point in the two-dimensional coordinate system; The set of coordinates of the mapped data points is determined as the low-dimensional subspace, and the square of the Euclidean distance from each mapped data point to the origin is calculated as the folded projection distance.

[0028] For the singular value sequence and orthogonal basis matrix obtained by performing singular value decomposition on the cross covariance matrix of the principal and secondary dimensions, select the first two valid singular values ​​from the singular value sequence. Extract the eigenvectors of the principal dimension orthogonal basis corresponding to these two valid singular values, and let them be denoted as . , Construct a spatial mapping matrix of size k×2 by piecing together columns. , where k is the number of principal dimensions. During the real-time recognition phase, the spatial mapping matrix P remains fixed.

[0029] Retrieve the main dimension data matrix within the current window Perform standard matrix multiplication on it and the space mapping matrix P. This method projects and folds the high-dimensional coordinates in k-dimensional space, outputting a set of mapped data point coordinates for each time step. The coordinate vector of each point in the two-dimensional coordinate system is denoted as... This two-dimensional set is output as a low-dimensional subspace.

[0030] After completing coordinate folding, extract each coordinate vector from the set of mapped data point coordinates. Treat it as a point at a distance from the origin in the new coordinate system, and calculate the square of its Euclidean distance to the origin. The square value This is defined as the folded projection distance of the i-th data point. For example, if the folded coordinates of a sample in a principal dimension at a certain sampling time are (0.8, -0.6) in two-dimensional space, the folded projection distance is obtained by summing the squares of each coordinate component. =1.00. This folded projection distance parameter comprehensively reflects the spatial energy distribution of the original high-dimensional sample points on the main feature component axes. It not only removes the numerical fluctuations of non-main parameters, but also indicates the degree of shift in the basic state of the hydrogenation reactor.

[0031] S3. Calculate the singular value energy concentration based on the singular value sequence, and construct a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance. Based on the composite folding scaling factor, perform manifold topology reconstruction on the low-dimensional subspace to obtain a two-dimensional folded feature vector, and calculate the ratio of the geometric distance before and after reconstruction as the manifold metric preservation coefficient.

[0032] During the offline calibration phase, a composite folding scaling factor is calculated based on the low-dimensional subspace data corresponding to the historical training samples. Specifically, the square of the first effective singular value is divided by the sum of the squares of the previous two effective singular values ​​to obtain the singular value energy concentration; the covariance matrix of the low-dimensional subspace data distribution is calculated, and the sum of the main diagonal elements of the covariance matrix is ​​used as the matrix trace of the low-dimensional subspace; the arithmetic mean of the folded projection distances of each data point in the low-dimensional subspace is calculated, and the global data variance sum of the dimensionless original multidimensional temporal feature matrix is ​​calculated. The arithmetic mean is divided by the global data variance sum to obtain the variance ratio. The composite folding scaling factor is obtained by multiplying the singular value energy concentration, the matrix trace of the low-dimensional subspace, and the variance ratio. This composite folding scaling factor serves as an offline calibration parameter and remains fixed during the real-time identification phase.

[0033] Based on the composite folding scaling factor and the variance of each coordinate dimension in the low-dimensional subspace, the fractal coordinate scaling coefficients corresponding to each coordinate dimension are calculated. Using these fractal coordinate scaling coefficients, the coordinate values ​​of different coordinate dimensions in the low-dimensional subspace data matrix are differentially scaled to obtain an expanded coordinate system cluster. The fractal coordinate scaling coefficients are determined during the offline calibration stage and remain fixed during the real-time recognition stage. A nearest-neighbor topology is constructed based on the expanded coordinate system clusters corresponding to historical training samples, and the two-dimensional embedding coordinates of the historical training samples are obtained using the Local Linear Embedding (LLE) algorithm, forming a fixed LLE reference coordinate system. The manifold dimensionality reduction transformation process is as follows: Figure 3 As shown.

[0034] In the real-time recognition stage, for the data point to be detected in the current sampling window, a fixed spatial mapping matrix and a fixed fractal coordinate scaling factor are used to obtain the expanded coordinates corresponding to the data point to be detected. Then, neighboring nodes are retrieved from the expanded coordinate system cluster of historical training samples obtained from offline calibration, and the local reconstruction weights are solved based on the expanded coordinate vectors of the data point to be detected and the expanded coordinate vectors of the neighboring nodes. Subsequently, the local reconstruction weights are applied to the historical two-dimensional embedded coordinates of the same neighboring nodes to obtain the two-dimensional folded feature vector of the data point to be detected in the fixed LLE reference coordinate system.

[0035] Calculate the sum of Euclidean distances between the data point to be detected and its neighboring nodes in the extended coordinate system cluster, and calculate the sum of geodesic distances of the same neighbor relationship in the nearest neighbor graph structure; divide the sum of Euclidean distances by the sum of geodesic distances to obtain the manifold metric preservation coefficient.

[0036] In some embodiments, the step of calculating the singular value energy concentration based on the singular value sequence and constructing a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance includes: Extract the square of the first effective singular value, divide it by the sum of the squares of the previous two effective singular values, and obtain the singular value energy concentration. Solve for the covariance matrix of the low-dimensional subspace data distribution, and calculate the matrix trace value by summing the elements on the main diagonal of the covariance matrix. Calculate the arithmetic mean of the folded projection distances of all data points in the low-dimensional subspace, and calculate the sum of global data variances after the original multidimensional temporal feature matrix has been dimensionless. Divide the former by the latter to obtain the variance ratio. The calculated singular value energy concentration, matrix trace value, and variance ratio are multiplied together, and the resulting product is determined as the composite folding scaling factor.

[0037] Extract all singular value sequences, calculate the quotient of the squared value of the first effective singular value and the sum of the squares of the first two effective singular values, and thus obtain the singular value energy concentration, which represents the energy proportion of the dominant singular direction. .

[0038] Based on the low-dimensional subspace dataset Y obtained by mapping, solve its 2×2 covariance matrix Cov(Y), and extract the two variance elements on the main diagonal to perform accumulation and calculate the matrix trace value Tr.

[0039] Sum the folded projected distances of all N data points and divide by N to obtain the arithmetic mean of the folded projected distances. After performing Z-score dimensionless standardization on the original m-dimensional acquisition matrix, the variance of each feature dimension is calculated, and the m variances are summed to obtain the total global data variance. Through the formula The variance ratio was calculated. .

[0040] After calculating the above three parameters, they are multiplied together, and the product is used as the composite folding scaling factor α. For example, assuming that in a certain monitoring batch, the sum of the singular value squared sequence calculated by SVD shows that the energy proportion of the first effective singular value in the first two effective singular values ​​is 85%, that is... =0.85. The sum of the main diagonal elements of the covariance matrix of the low-dimensional subspace data coordinates yields a matrix trace Tr = 1.20. Simultaneously, the arithmetic mean of the folded projected distance is 1.20, and the sum of the variances of the original multidimensional global data is 4.00. The variance ratio is then calculated. =0.30. Multiplying these three values ​​yields a composite folding scaling factor α = 0.306. Since the low-dimensional subspace is obtained by mapping the standardized data, this composite folding scaling factor is a dimensionless scaling parameter.

[0041] In some embodiments, the step of reconstructing the manifold topology of the low-dimensional subspace based on a composite folding scaling factor to obtain a two-dimensional folded feature vector, and calculating the ratio of the geometric distance before and after reconstruction as a manifold metric preservation coefficient, includes: The fractal coordinate scaling factor is calculated based on the composite folding scaling factor and the variance of each coordinate dimension in the low-dimensional subspace, and the scale of the low-dimensional subspace is adjusted using the fractal coordinate scaling factor to obtain the expanded coordinate system cluster. Based on the extended coordinate system cluster in the offline calibration stage, a local connectivity topology reconstruction weight matrix is ​​constructed, and the two-dimensional embedded coordinates are solved according to the local connectivity topology reconstruction weight matrix. In the real-time recognition stage, based on the proximity relationship between the data point to be detected and the historical training samples, the data point to be detected is mapped to the reference coordinate system corresponding to the two-dimensional embedded coordinates to obtain the two-dimensional folded feature vector corresponding to the data point to be detected. Calculate the sum of Euclidean distances between the data point to be detected and its neighboring nodes in the extended coordinate system cluster, and calculate the sum of geodesic distances for the same neighbor relationship in the nearest neighbor graph structure; Dividing the sum of the Euclidean distances by the sum of the geodesic distances yields the manifold preservation coefficient.

[0042] The coordinate matrix Y of all data points obtained from the low-dimensional subspace is scaled according to the scaling factor of the fractal coordinates corresponding to each coordinate dimension to obtain an expanded coordinate system cluster. Specifically, based on the variance of the composite folding scaling factor α and the a-th coordinate dimension... Constructing the scaling factor of the fractal coordinate , where ε is a preset minimum constant to prevent the denominator from being zero, and a is the coordinate dimension number of the low-dimensional subspace. Thus, different coordinate dimensions obtain different scaling ratios according to their degree of discreteness.

[0043] Within the extended coordinate system cluster, for each data point The K-nearest neighbor algorithm is invoked, and the initial number of neighbors K=3 is set to retrieve the three nearest neighbor nodes in space. Based on the nearest neighbor graph structure composed of each data point and its neighbor nodes, it is determined whether the nearest neighbor graph is connected; if the nearest neighbor graph is not connected, the number of neighbors K is incremented and the nearest neighbor graph is reconstructed to satisfy the connectivity condition.

[0044] The constrained least squares method is used to solve for the local reconstruction weights. Specifically, for the i-th data point, its neighboring node coordinate vectors are used as the local reconstruction basis. Under the constraint that the sum of the reconstruction weights of each neighboring node is 1, the weight combination that minimizes the error between the weighted combination of neighboring nodes and the coordinate vector of the i-th data point is solved, and this weight combination is used as the local reconstruction weight of the i-th data point. Unconnected point pairs are assigned weights of 0, thus forming a local connectivity topology reconstruction weight matrix of size N×N. .

[0045] In obtaining Next, calculate its corresponding cost matrix. , where I is an N×N identity matrix. The eigenvalues ​​and eigenvector matrices of this cost matrix M are solved, and the smallest zero eigenvalue is discarded. The eigenvector column corresponding to the two smallest non-zero eigenvalues ​​is extracted. and For the i-th historical training sample data, the two components in the i-th row of the feature vector matrix are extracted to form a new historical two-dimensional embedding coordinate, completing offline nonlinear dimensionality reduction. For the real-time data point to be detected, the historical two-dimensional embedding coordinates of the neighboring nodes are weighted and summed according to the local reconstruction weights of their neighboring nodes to obtain the two-dimensional folded feature vector of the real-time data point to be detected in the same LLE reference coordinate system.

[0046] Calculate the sum of the straight-line Euclidean distances between the data point to be detected and its neighboring nodes within the extended coordinate system cluster. And calculate the sum of geodesic distances in the nearest neighbor graph structure corresponding to the same neighbor relationship. Perform division operation Output the manifold metric preservation coefficient. For example, if the calculated sum of Euclidean distances is 450.0 and the sum of geodesic distances is 500.0, then the manifold metric preservation coefficient is... =0.90, this coefficient is used to characterize the degree of preservation between the linear distance relationship and the geodesic distance relationship of the nearest neighbor map within the extended coordinate system cluster.

[0047] S4 calculates the weighted sequence of eigenvalues ​​based on manifold topology reconstruction, calculates the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library, corrects the initial fault judgment threshold using the manifold metric preservation coefficient to obtain the dynamic threshold, and outputs the corresponding fault mode diagnosis result when the weighted Chebyshev distance is less than the dynamic threshold.

[0048] When constructing the fault mode benchmark library, the same standardized parameters, main dimension set, auxiliary dimension set, spatial mapping matrix, fractal coordinate scaling factor and LLE benchmark coordinate system as the real-time monitoring data are used to process historical fault samples to obtain two-dimensional benchmark vectors corresponding to each fault mode, so that the benchmark vectors in the fault mode benchmark library and the two-dimensional folded feature vectors corresponding to the data points to be detected are in the same mapping coordinate system.

[0049] Extract the two smallest non-zero eigenvalues ​​obtained in the process of solving the eigenvalue problem of manifold topology reconstruction, and calculate the reciprocals of the two smallest non-zero eigenvalues ​​respectively. The two reciprocal values ​​are then used to form a weight sequence.

[0050] Each baseline vector is extracted from a fault mode benchmark library containing various known reactor fault types, pre-constructed using historical fault samples. The two-dimensional folded feature vector corresponding to the data point to be detected is extracted from the two-dimensional folded feature vectors corresponding to each data point. For each baseline vector, the absolute value of the difference between the two-dimensional folded feature vector corresponding to the data point to be detected and the vector in the two dimensions of the two-dimensional space is calculated. The corresponding weight in the weight sequence is multiplied by the absolute value of the difference in these two dimensions, and the maximum value of these two products is taken as the weighted Chebyshev distance.

[0051] Based on the historical archived two-dimensional feature set of the corresponding fault mode, the weighted Chebyshev distance between the in-class samples of the fault mode and its two-dimensional reference vector is calculated, and the maximum value of the in-class weighted Chebyshev distance is set as the initial fault determination threshold. The manifold metric distortion is determined based on the manifold metric preservation coefficient, which is the difference between 1 and the manifold metric preservation coefficient. The manifold metric distortion is multiplied by a preset adjustment coefficient greater than zero, and 1 is added to the product to obtain the adaptive threshold correction multiplier. The initial fault determination threshold is multiplied by the adaptive threshold correction multiplier to obtain the dynamic threshold.

[0052] The calculated weighted Chebyshev distance of the current data is compared with the dynamic threshold. When a weighted Chebyshev distance is less than the dynamic threshold, the castor oil hydrogenation reactor is determined to have a fault represented by the reference vector corresponding to that weighted Chebyshev distance, and the corresponding fault mode diagnosis result is output.

[0053] In some embodiments, the calculation of the weighted Chebyshev distance between the two-dimensional folded feature vector and each reference vector in the fault mode reference library based on the eigenvalue calculation weight sequence of manifold topology reconstruction includes: Extract the two smallest non-zero eigenvalues ​​of the characteristic equation of the cost matrix, calculate the reciprocals of these two eigenvalues, and construct a weight sequence containing the two values. Extract the two-dimensional folded feature vector corresponding to the data point to be detected from the two-dimensional folded feature vector corresponding to each data point, obtain two feature values, and simultaneously extract the two benchmark values ​​corresponding to the benchmark vector of the compared fault mode. After performing subtraction calculation on each other, take the absolute value to obtain two feature difference values. Multiply the two feature difference values ​​by the corresponding numerical components in the weight sequence to obtain two weighted independent components; Compare the values ​​of the two weighted independent components and output the largest positive value as the determined weighted Chebyshev distance between them.

[0054] From the eigenvalue decomposition results of the cost matrix, extract the two least-two non-zero eigenvalues ​​corresponding to the dimension of the two-dimensional folded eigenvector, and let them be denoted as . and By calculating the reciprocals of both separately. and This forms a weight sequence containing two weight elements. For any two-dimensional folded feature vector obtained from real-time monitoring Retrieve fault reference vectors from the fault mode reference library For each corresponding data point in the same dimension, the absolute value of the difference is calculated to obtain the feature difference value. as well as .

[0055] The two eigenvalues ​​are independently multiplied by their corresponding values ​​in the weight sequence to obtain two weighted independent components. and Based on the principle of examining the maximum deviation of the coordinate axes using Chebyshev distance, the magnitudes of the two weighted independent components are compared, and the output is... The weighted Chebyshev distance between the current sample state and the baseline of this type of failure mode in the weighted manifold space.

[0056] For example, if the two extracted non-zero eigenvalues ​​are 0.25 and 0.40, their reciprocals generate a weight sequence of (4.0, 2.5). When the measured real-time two-dimensional folded feature vector is (0.6, 0.9) and the benchmark vector for a certain type of fault is (0.4, 1.2), the absolute values ​​of the feature difference terms are first calculated to be 0.2 and 0.3, respectively. Multiplying these by each term of the weight sequence yields the weighted independent components. =0.8, =0.75. After comparison, the largest positive value was determined to be 0.8, and its output was used as the weighted Chebyshev distance parameter between the two, realizing a quantitative assessment of the differences in sensitivity to information in different dimensions.

[0057] In some embodiments, the step of using the manifold metric preservation coefficient to correct the initial fault determination threshold to obtain a dynamic threshold includes: Retrieve the historical archived two-dimensional feature set of the corresponding fault mode, calculate the weighted Chebyshev distance between the in-class sample of the fault mode and its two-dimensional reference vector, and set the maximum value of the in-class weighted Chebyshev distance as the initial fault judgment threshold. The manifold distortion is determined based on the manifold preservation coefficient, wherein the manifold distortion is the difference between 1 and the manifold preservation coefficient. The manifold metric distortion is multiplied by a preset adjustment coefficient that is greater than zero, and 1 is added to the product to obtain the adaptive threshold correction multiplier. The initial fault determination threshold is multiplied by the adaptive threshold correction multiplier to obtain the dynamic threshold.

[0058] For each specific known fault mode, the database is connected to extract the historical archived two-dimensional feature set corresponding to the fault. A weighted Chebyshev distance is calculated between all historical two-dimensional feature samples of this fault mode and its two-dimensional baseline vector, and the maximum value is taken as the initial fault determination threshold for that specific fault mode. Obtain the manifold preservation coefficient calculated in the preceding step. And calculate the manifold metric distortion. Set a preset adjustment coefficient β greater than zero. Multiply the manifold metric distortion δ by the preset adjustment coefficient β and add 1 to obtain the adaptive threshold correction multiplier γ. Set the initial fault determination threshold... Multiply by the adaptive threshold correction multiplier γ to obtain the dynamic threshold. .

[0059] For example, the maximum weighted Chebyshev distance within this fault class, calculated from historical reactor blockage fault sets, is 1.8, which is the initial fault determination threshold. =1.8. If the manifold preservation coefficient calculated for the current monitoring period is... =0.90, with a preset adjustment coefficient β=0.5, then the manifold metric distortion δ=1 0.90 = 0.10, adaptive threshold correction multiplier γ = 1 + 0.5 × 0.10 = 1.05, thus the dynamic threshold is calculated. =1.89. When the weighted Chebyshev distance between the point to be detected and the fault baseline vector is less than 1.89, the reactor is determined to have the corresponding fault mode and the result is output. Optionally, if the point to be detected simultaneously meets the dynamic threshold conditions of multiple fault modes, the fault mode with the smallest ratio of weighted Chebyshev distance to the corresponding dynamic threshold is selected as the diagnostic result. This processing method can adaptively adjust the judgment threshold according to the distance maintenance during the manifold reconstruction process, which can improve the adaptability of the model diagnosis process to operating condition fluctuations.

[0060] Example 2 Embodiment 2 of the present invention proposes a fault mode identification and diagnosis system for a castor oil hydrogenation reactor, comprising the following modules: The module is used to collect the status monitoring parameters of the castor oil hydrogenation reactor to construct a multi-dimensional time-series feature matrix; the synergistic correlation coefficient matrix between dimensions is calculated using orthogonal vector projection, the dimensional contribution weight vector is extracted, and the multi-dimensional features are divided into main dimensions and auxiliary dimensions. The calculation module is used to extract the main dimension and auxiliary dimension data, calculate the cross covariance matrix and perform singular value decomposition to obtain the orthogonal basis eigenvectors corresponding to the singular value sequence and the main dimension; use the orthogonal basis eigenvectors corresponding to the main dimension to determine the mapping relationship, fold the main dimension data to a low-dimensional subspace, and calculate the folded projection distance of the data points. The reconstruction module is used to calculate the singular value energy concentration based on the singular value sequence, construct a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance, and perform manifold topology reconstruction on the low-dimensional subspace based on the composite folding scaling factor to obtain a two-dimensional folded feature vector, and calculate the ratio of the geometric distance before and after reconstruction as the manifold metric preservation coefficient. The diagnostic module is used to calculate the weight sequence of eigenvalues ​​based on manifold topology reconstruction, calculate the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library, correct the initial fault judgment threshold using the manifold metric preservation coefficient to obtain the dynamic threshold, and output the corresponding fault mode diagnostic result when the weighted Chebyshev distance is less than the dynamic threshold.

[0061] In some embodiments, the calculation of the synergistic correlation coefficient matrix between dimensions using orthogonal vector projection includes: Two columns of data with different column numbers in the multidimensional time series feature matrix are selected sequentially to form a paired column vector combination; The inner product of two vectors within a paired column vector combination is used as the correlation numerator. Calculate the Euclidean norm of each vector separately and multiply them together to obtain the result as the denominator of the correlation. The orthogonal projection correlation coefficient between the two corresponding dimensions is calculated by dividing the correlation numerator by the correlation denominator. Traverse the column vector combinations with different column numbers in the multidimensional time series feature matrix, calculate the corresponding orthogonal projection correlation coefficients, and arrange the orthogonal projection correlation coefficients into a symmetric matrix according to the feature dimension correspondence. At the same time, set the self-correlation coefficient corresponding to the data in the same column to 1 to obtain the co-correlation coefficient matrix.

[0062] In some embodiments, the extraction of dimensional contribution weight vectors and the division of multidimensional features into primary and secondary dimensions includes: Calculate the absolute value of all values ​​in the corresponding row of each feature dimension in the co-correlation coefficient matrix, and sum the absolute values ​​of the same row to obtain the total absolute correlation value of the feature dimension; Divide the total absolute correlation value of a single feature dimension by the sum of the total absolute correlation values ​​of all feature dimensions to obtain the normalized contribution weight of each feature dimension, and arrange them in order to generate a dimension contribution weight vector. The feature dimensions are sorted in descending order of the normalized contribution weights. The two feature dimensions with the highest ranking are assigned to the primary dimension set, and the remaining two feature dimensions are assigned to the secondary dimension set.

[0063] In some embodiments, determining the mapping relationship using the orthogonal basis eigenvectors corresponding to the principal dimension, folding the principal dimension data to a low-dimensional subspace, and calculating the folded projection distance of the data points includes: Extract the orthogonal basis eigenvectors corresponding to the principal dimensions associated with the first two valid singular values ​​in descending order of singular values ​​to construct a spatial mapping matrix, and save the spatial mapping matrix as an offline calibration parameter. Perform matrix multiplication between the main dimension data in the real-time sampling window and the spatial mapping matrix to obtain the set of mapped data point coordinates for each data point in the two-dimensional coordinate system; The set of coordinates of the mapped data points is determined as the low-dimensional subspace, and the square of the Euclidean distance from each mapped data point to the origin is calculated as the folded projection distance.

[0064] In some embodiments, the step of calculating the singular value energy concentration based on the singular value sequence and constructing a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance includes: Extract the square of the first effective singular value, divide it by the sum of the squares of the previous two effective singular values, and obtain the singular value energy concentration. Solve for the covariance matrix of the low-dimensional subspace data distribution, and calculate the matrix trace value by summing the elements on the main diagonal of the covariance matrix. Calculate the arithmetic mean of the folded projection distances of all data points in the low-dimensional subspace, and calculate the sum of global data variances after the original multidimensional temporal feature matrix has been dimensionless. Divide the former by the latter to obtain the variance ratio. The calculated singular value energy concentration, matrix trace value, and variance ratio are multiplied together, and the resulting product is determined as the composite folding scaling factor.

[0065] In some embodiments, the step of reconstructing the manifold topology of the low-dimensional subspace based on a composite folding scaling factor to obtain a two-dimensional folded feature vector, and calculating the ratio of the geometric distance before and after reconstruction as a manifold metric preservation coefficient, includes: The fractal coordinate scaling factor is calculated based on the composite folding scaling factor and the variance of each coordinate dimension in the low-dimensional subspace, and the scale of the low-dimensional subspace is adjusted using the fractal coordinate scaling factor to obtain the expanded coordinate system cluster. Based on the extended coordinate system cluster in the offline calibration stage, a local connectivity topology reconstruction weight matrix is ​​constructed, and the two-dimensional embedded coordinates are solved according to the local connectivity topology reconstruction weight matrix. In the real-time recognition stage, based on the proximity relationship between the data point to be detected and the historical training samples, the data point to be detected is mapped to the reference coordinate system corresponding to the two-dimensional embedded coordinates to obtain the two-dimensional folded feature vector corresponding to the data point to be detected. Calculate the sum of Euclidean distances between the data point to be detected and its neighboring nodes in the extended coordinate system cluster, and calculate the sum of geodesic distances for the same neighbor relationship in the nearest neighbor graph structure; Dividing the sum of the Euclidean distances by the sum of the geodesic distances yields the manifold preservation coefficient.

[0066] In some embodiments, the calculation of the weighted Chebyshev distance between the two-dimensional folded feature vector and each reference vector in the fault mode reference library based on the eigenvalue calculation weight sequence of manifold topology reconstruction includes: Extract the two smallest non-zero eigenvalues ​​of the characteristic equation of the cost matrix, calculate the reciprocals of these two eigenvalues, and construct a weight sequence containing the two values. Extract the two-dimensional folded feature vector corresponding to the data point to be detected from the two-dimensional folded feature vector corresponding to each data point, obtain two feature values, and simultaneously extract the two benchmark values ​​corresponding to the benchmark vector of the compared fault mode. After performing subtraction calculation on each other, take the absolute value to obtain two feature difference values. Multiply the two feature difference values ​​by the corresponding numerical components in the weight sequence to obtain two weighted independent components; Compare the values ​​of the two weighted independent components and output the largest positive value as the determined weighted Chebyshev distance between them.

[0067] In some embodiments, the step of using the manifold metric preservation coefficient to correct the initial fault determination threshold to obtain a dynamic threshold includes: Retrieve the historical archived two-dimensional feature set of the corresponding fault mode, calculate the weighted Chebyshev distance between the in-class sample of the fault mode and its two-dimensional reference vector, and set the maximum value of the in-class weighted Chebyshev distance as the initial fault judgment threshold. The manifold distortion is determined based on the manifold preservation coefficient, wherein the manifold distortion is the difference between 1 and the manifold preservation coefficient. The manifold metric distortion is multiplied by a preset adjustment coefficient that is greater than zero, and 1 is added to the product to obtain the adaptive threshold correction multiplier. The initial fault determination threshold is multiplied by the adaptive threshold correction multiplier to obtain the dynamic threshold.

[0068] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for fault mode identification and diagnosis of a castor oil hydrogenation reactor, characterized in that, include: A multidimensional time-series feature matrix was constructed by collecting status monitoring parameters of the castor oil hydrogenation reactor. The co-correlation coefficient matrix between dimensions is calculated using orthogonal vector projection, the dimensional contribution weight vector is extracted, and the multidimensional features are divided into primary and secondary dimensions. Extract the main dimension and auxiliary dimension data, calculate the cross covariance matrix and perform singular value decomposition to obtain the orthogonal basis eigenvectors corresponding to the singular value sequence and the main dimension; use the orthogonal basis eigenvectors corresponding to the main dimension to determine the mapping relationship, fold the main dimension data to a low-dimensional subspace, and calculate the folded projection distance of the data points; The singular value energy concentration is calculated based on the singular value sequence, and a composite folding scaling factor is constructed by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance. Based on the composite folding scaling factor, the manifold topology of the low-dimensional subspace is reconstructed to obtain the two-dimensional folded feature vector, and the ratio of the geometric distance before and after reconstruction is calculated as the manifold metric preservation coefficient. Based on the eigenvalue weight sequence of manifold topology reconstruction, the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library is calculated; the initial fault judgment threshold is corrected by the manifold metric preservation coefficient to obtain the dynamic threshold; when the weighted Chebyshev distance is less than the dynamic threshold, the corresponding fault mode diagnosis result is output.

2. The method according to claim 1, characterized in that, The calculation of the synergistic correlation coefficient matrix between dimensions using orthogonal vector projection includes: Two columns of data with different column numbers in the multidimensional time series feature matrix are selected sequentially to form a paired column vector combination; The inner product of two vectors within a paired column vector combination is used as the correlation numerator. Calculate the Euclidean norm of each vector separately and multiply them together to obtain the result as the denominator for the correlation. The orthogonal projection correlation coefficient between the two corresponding dimensions is calculated by dividing the correlation numerator by the correlation denominator. Traverse the column vector combinations with different column numbers in the multidimensional time series feature matrix, calculate the corresponding orthogonal projection correlation coefficients, and arrange the orthogonal projection correlation coefficients into a symmetric matrix according to the feature dimension correspondence. At the same time, set the self-correlation coefficient corresponding to the data in the same column to 1 to obtain the co-correlation coefficient matrix.

3. The method according to claim 2, characterized in that, The extracted dimensional contribution weight vector is used to divide the multidimensional features into primary and secondary dimensions, including: Calculate the absolute value of all values ​​in the corresponding row of each feature dimension in the co-correlation coefficient matrix, and sum the absolute values ​​of the same row to obtain the total absolute correlation value of the feature dimension; Divide the total absolute correlation value of a single feature dimension by the sum of the total absolute correlation values ​​of all feature dimensions to obtain the normalized contribution weight of each feature dimension, and arrange them in order to generate a dimension contribution weight vector. The feature dimensions are sorted in descending order of the normalized contribution weights. The two feature dimensions with the highest ranking are assigned to the primary dimension set, and the remaining two feature dimensions are assigned to the secondary dimension set.

4. The method according to claim 1, characterized in that, The step of determining the mapping relationship using the orthogonal basis eigenvectors corresponding to the principal dimension, folding the principal dimension data to a low-dimensional subspace, and calculating the folded projection distance of the data points includes: Extract the orthogonal basis eigenvectors corresponding to the principal dimensions associated with the first two valid singular values ​​in descending order of singular values ​​to construct a spatial mapping matrix, and save the spatial mapping matrix as an offline calibration parameter. Perform matrix multiplication between the main dimension data in the real-time sampling window and the spatial mapping matrix to obtain the set of mapped data point coordinates for each data point in the two-dimensional coordinate system; The set of coordinates of the mapped data points is determined as the low-dimensional subspace, and the square of the Euclidean distance from each mapped data point to the origin is calculated as the folded projection distance.

5. The method according to claim 1, characterized in that, The step of calculating the singular value energy concentration based on the singular value sequence and constructing a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance includes: Extract the square of the first effective singular value, divide it by the sum of the squares of the previous two effective singular values, and obtain the singular value energy concentration. Solve for the covariance matrix of the low-dimensional subspace data distribution, and calculate the matrix trace value by summing the elements on the main diagonal of the covariance matrix. Calculate the arithmetic mean of the folded projection distances of all data points in the low-dimensional subspace, and calculate the sum of global data variances after the original multidimensional temporal feature matrix has been dimensionless. Divide the former by the latter to obtain the variance ratio. The calculated singular value energy concentration, matrix trace value, and variance ratio are multiplied together, and the resulting product is determined as the composite folding scaling factor.

6. The method according to claim 1, characterized in that, The method of reconstructing the manifold topology of the low-dimensional subspace based on a composite folding scaling factor to obtain a two-dimensional folded feature vector, and calculating the ratio of the geometric distance before and after reconstruction as the manifold metric preservation coefficient, includes: The fractal coordinate scaling factor is calculated based on the composite folding scaling factor and the variance of each coordinate dimension in the low-dimensional subspace, and the scale of the low-dimensional subspace is adjusted using the fractal coordinate scaling factor to obtain the expanded coordinate system cluster. Based on the extended coordinate system cluster in the offline calibration stage, a local connectivity topology reconstruction weight matrix is ​​constructed, and the two-dimensional embedded coordinates are solved according to the local connectivity topology reconstruction weight matrix. In the real-time recognition stage, based on the proximity relationship between the data point to be detected and the historical training samples, the data point to be detected is mapped to the reference coordinate system corresponding to the two-dimensional embedded coordinates to obtain the two-dimensional folded feature vector corresponding to the data point to be detected. Calculate the sum of Euclidean distances between the data point to be detected and its neighboring nodes in the extended coordinate system cluster, and calculate the sum of geodesic distances for the same neighbor relationship in the nearest neighbor graph structure; Dividing the sum of the Euclidean distances by the sum of the geodesic distances yields the manifold preservation coefficient.

7. The method according to claim 6, characterized in that, The eigenvalue calculation weight sequence based on manifold topology reconstruction calculates the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library, including: Extract the two smallest non-zero eigenvalues ​​of the characteristic equation of the cost matrix, calculate the reciprocals of these two eigenvalues, and construct a weight sequence containing the two values. Extract the two-dimensional folded feature vector corresponding to the data point to be detected from the two-dimensional folded feature vector corresponding to each data point, obtain two feature values, and simultaneously extract the two benchmark values ​​corresponding to the benchmark vector of the compared fault mode. After performing subtraction calculation on each other, take the absolute value to obtain two feature difference values. Multiply the two feature difference values ​​by the corresponding numerical components in the weight sequence to obtain two weighted independent components; Compare the values ​​of the two weighted independent components and output the largest positive value as the determined weighted Chebyshev distance between them.

8. The method according to claim 1, characterized in that, The process of obtaining a dynamic threshold by correcting the initial fault determination threshold using the manifold metric preservation coefficient includes: Retrieve the historical archived two-dimensional feature set of the corresponding fault mode, calculate the weighted Chebyshev distance between the in-class sample of the fault mode and its two-dimensional reference vector, and set the maximum value of the in-class weighted Chebyshev distance as the initial fault judgment threshold. The manifold distortion is determined based on the manifold preservation coefficient, wherein the manifold distortion is the difference between 1 and the manifold preservation coefficient. The manifold metric distortion is multiplied by a preset adjustment coefficient that is greater than zero, and 1 is added to the product to obtain the adaptive threshold correction multiplier. The initial fault determination threshold is multiplied by the adaptive threshold correction multiplier to obtain the dynamic threshold.

9. A fault mode identification and diagnosis system for a castor oil hydrogenation reactor, characterized in that, include: A partitioning module is used to collect status monitoring parameters of the castor oil hydrogenation reactor and construct a multi-dimensional time-series feature matrix. The co-correlation coefficient matrix between dimensions is calculated using orthogonal vector projection, the dimensional contribution weight vector is extracted, and the multidimensional features are divided into primary and secondary dimensions. The calculation module is used to extract the main dimension and auxiliary dimension data, calculate the cross covariance matrix and perform singular value decomposition to obtain the orthogonal basis eigenvectors corresponding to the singular value sequence and the main dimension; use the orthogonal basis eigenvectors corresponding to the main dimension to determine the mapping relationship, fold the main dimension data to a low-dimensional subspace, and calculate the folded projection distance of the data points. The reconstruction module is used to calculate the singular value energy concentration based on the singular value sequence and construct a composite folding scaling factor by combining the matrix trace of the low-dimensional subspace with the variance ratio based on the folding projection distance. Based on the composite folding scaling factor, the manifold topology of the low-dimensional subspace is reconstructed to obtain the two-dimensional folded feature vector, and the ratio of the geometric distance before and after reconstruction is calculated as the manifold metric preservation coefficient. The diagnostic module is used to calculate the weight sequence of eigenvalues ​​based on manifold topology reconstruction, calculate the weighted Chebyshev distance between the two-dimensional folded eigenvector and each benchmark vector in the fault mode benchmark library, correct the initial fault judgment threshold using the manifold metric preservation coefficient to obtain the dynamic threshold, and output the corresponding fault mode diagnostic result when the weighted Chebyshev distance is less than the dynamic threshold.

10. The system according to claim 9, characterized in that, The calculation of the synergistic correlation coefficient matrix between dimensions using orthogonal vector projection includes: Two columns of data with different column numbers in the multidimensional time series feature matrix are selected sequentially to form a paired column vector combination; The inner product of two vectors within a paired column vector combination is used as the correlation numerator. Calculate the Euclidean norm of each vector separately and multiply them together to obtain the result as the denominator for the correlation. The orthogonal projection correlation coefficient between the two corresponding dimensions is calculated by dividing the correlation numerator by the correlation denominator. Traverse the column vector combinations with different column numbers in the multidimensional time series feature matrix, calculate the corresponding orthogonal projection correlation coefficients, and arrange the orthogonal projection correlation coefficients into a symmetric matrix according to the feature dimension correspondence. At the same time, set the self-correlation coefficient corresponding to the data in the same column to 1 to obtain the co-correlation coefficient matrix.

Images