Power plant intelligent early warning method and system based on big data

Abstract

The invention provides a power plant intelligent early warning method and system based on big data, and relates to the technical field of power systems, and the method comprises the steps: generating a high-dimensional feature map, generating a multi-modal feature tensor through adaptive chaotic mapping, constructing a feature relation network, generating a time sequence multi-dimensional feature vector, and reconstructing a multi-dimensional state vector based on topological entropy. In combination with a topological data analysis method, carrying out high-dimensional topological space mapping on the multi-dimensional state vector, identifying an operation mode of the power plant equipment, calculating a fuzzy region and a mode abrupt change point of the operation state of the equipment, and obtaining a dynamic evaluation result of the health state of the equipment; a multi-level recurrent neural network is combined, a prediction result is obtained through multi-scale time sequence prediction, uncertainty corresponding to the prediction result is quantitatively solved based on a quantum probability graph model, quantum probability distribution of an early warning signal is generated, a real-time operation strategy of equipment is adjusted, and an intelligent early warning signal of the power plant equipment is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a big data-based intelligent early warning method and system for power plants. Background Art

[0002] The power industry is an important basic industry. As the core component of the power system, the safe and stable operation of power plants is directly related to the reliability and economy of the power system. With the continuous development and complexity of the power system, the operating status of power plant equipment presents the characteristics of multi-source heterogeneity, high-dimensional dynamics, and nonlinear complexity. The traditional early warning method based on empirical models and a single data source is difficult to adapt to the real-time monitoring and fault early warning needs of power plant equipment.

[0003] In recent years, big data technology can realize the real-time collection, storage and processing of massive multi-source data, extract hidden patterns and laws from the data, and realize real-time evaluation and prediction of equipment operation status. However, the existing big data early warning methods are mainly based on shallow machine learning models, such as support vector machines, decision trees, etc. For the complex nonlinear characteristics and high-dimensional dynamic characteristics of power plant equipment, the accuracy and real-time performance of early warning still need to be improved; in addition, most of the existing early warning methods are based on a single data source, ignoring the multi-source heterogeneous characteristics of the operation status of power plant equipment, and it is difficult to fully characterize the health status of the equipment. At the same time, the existing methods are mainly based on the statistical characteristics of the data for modeling, lacking the mining of the intrinsic topological structure and dynamic evolution laws of the data, resulting in insufficient generalization and robustness of the early warning model.

[0004] Therefore, there is an urgent need for an intelligent early warning method for power plants based on big data, which can make full use of multi-source heterogeneous data, mine the intrinsic topological structure and dynamic evolution laws of the data, build an efficient and accurate early warning model, realize real-time health monitoring and fault early warning of power plant equipment, and ensure the safe and stable operation of the power plant. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a power plant intelligent early warning method and system based on big data, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a power plant intelligent early warning method based on big data, including:

[0008] Based on the operation data and environmental data of the power plant equipment, multi-scale features are extracted through a multi-dimensional data stream adaptive hierarchical sampling method to generate a high-dimensional feature map. The high-dimensional feature map is nonlinearly mapped and sparsely encoded through adaptive chaotic mapping to generate a multi-modal feature tensor. Based on a network community discovery algorithm and in combination with node importance determination, the multi-modal feature tensor is dynamically sampled at multiple levels to construct a feature relationship network and generate a time series multi-dimensional feature vector. Based on topological entropy, the structure of the feature relationship network is adjusted to reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector.

[0009] Based on the multidimensional state vector, combined with the topological data analysis method, the multidimensional state vector is mapped into a high-dimensional topological space through nonlinear uncertainty mapping, the operation mode of the power plant equipment is identified using a dynamic clustering algorithm, and the fuzzy area and mode mutation point of the equipment operation state are calculated through the combination of fuzzy logic and stochastic differential equations to obtain a dynamic evaluation result of the equipment health state;

[0010] Based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, the health status of the equipment is predicted through multi-scale time series prediction to obtain a prediction result. The uncertainty corresponding to the prediction result is quantified based on the quantum probability graph model to generate a quantum probability distribution of the warning signal. Based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to generate an intelligent warning signal for the power plant equipment.

[0011] In an optional embodiment,

[0012] Based on the operation data and environmental data of power plant equipment, multi-scale features are extracted through a multi-dimensional data stream adaptive layered sampling method to generate a high-dimensional feature map. The high-dimensional feature map is nonlinearly mapped and sparsely encoded through adaptive chaotic mapping to generate a multi-modal feature tensor including:

[0013] The operation data and the environment data constitute a multidimensional data stream, and the multidimensional data stream is adaptively layered sampled. Based on the basic statistical properties of the multidimensional data stream and the corresponding dynamic changes, the local variance of the multidimensional data stream is calculated using a sliding window. Based on the comparison result of the local variance with a preset variance threshold, the number of sampling layers and the sampling rate corresponding to each layer are adaptively determined to obtain the sampled multidimensional data stream;

[0014] The multi-dimensional data stream after sampling is decomposed into multiple scales by using multi-wavelet transform, a multi-wavelet basis is constructed by tensor product, a wavelet coefficient threshold is adaptively determined, and a soft threshold processing is performed on the wavelet coefficient to obtain the processed wavelet coefficient, statistical features of the processed wavelet coefficient are calculated at different scales, and statistical features of different scales are combined to obtain a high-dimensional feature map, wherein the statistical features include mean, variance, skewness and kurtosis;

[0015] The high-dimensional feature map is divided into networks of different sizes, and the fractal dimension is calculated by linear fitting in a double logarithmic coordinate system, and a chaotic mapping function and corresponding control parameters are adaptively determined according to the fractal dimension, and the high-dimensional feature map is nonlinearly transformed, and chaotic diffusion and sparse representation are performed in the feature space to determine the feature map after chaotic mapping;

[0016] The feature map after chaotic mapping is divided into multiple scales to obtain sub-maps of different scales. At each scale, chaotic mapping is recursively applied to the sub-map to generate multi-scale sparse coding. The number of recursive layers is adaptively determined based on the fractal dimension corresponding to each sub-map. The multi-level sparse coding is adaptively determined by adjusting the control parameters of the corresponding chaotic mapping. The multi-scale sparse coding and the multi-level sparse coding are cross-scale fused to obtain a multi-modal feature tensor.

[0017] Perform high-order singular value decomposition on the multimodal feature tensor, decompose the multimodal feature tensor into multiple matrices and a core tensor, each matrix corresponds to a mode, determine the relationship between the modes, obtain singular values ​​and singular vectors, arrange the singular values ​​in order of size, generate a singular value distribution, determine a truncation threshold of the singular values, retain the singular values ​​and corresponding singular vectors greater than the truncation threshold, and reconstruct the multimodal feature tensor.

[0018] In an optional embodiment,

[0019] Based on the network community discovery algorithm, combined with the determination of node importance, the multi-modal feature tensor is dynamically sampled at multiple levels to construct a feature relationship network, generate a time series multi-dimensional feature vector, and based on the topological entropy, adjust the structure of the feature relationship network, and reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector, including:

[0020] The features in the multimodal feature tensor are used as nodes, and the correlation strength between the features is used as the weight of the edge to construct a weighted undirected graph. Based on the network community discovery algorithm, different resolution parameters are set to perform community division on the weighted undirected graph. Based on different resolution parameters, feature community structures of different granularities are obtained. Feature nodes are sorted according to the importance of nodes in each feature community structure. According to a preset sampling ratio, nodes of a corresponding ratio are sampled from each feature community structure to form a multi-level feature subset.

[0021] Calculate the correlation coefficient matrix between any two features in the feature subset, determine the initial correlation strength between the features, generate an initial feature relationship network, binarize the correlation coefficient matrix according to a preset binarization threshold by a threshold method, obtain an initial adjacency matrix of the initial feature relationship network, perform random walks on the initial feature relationship network, correct the initial correlation strength, and obtain a corrected feature relationship network;

[0022] Based on the multimodal feature tensor, extract the feature tensor slice corresponding to each timestamp, map the feature tensor slice into a single multidimensional feature vector according to the topological structure corresponding to the modified feature relationship network, generate a multidimensional feature vector under a timestamp by traversing all nodes in the modified feature relationship network, and repeatedly iterate multiple consecutive timestamps to generate a time series multidimensional feature vector;

[0023] Using topological entropy as an evaluation index and adopting a heuristic search algorithm, the optimal network structure adjustment strategy is searched, and the network structure with the minimum topological entropy is determined through iterative search, and the structure of the modified characteristic relationship network is adjusted to obtain the final characteristic relationship network;

[0024] Through the graph convolutional neural network, a convolution operation is performed on each node in the final feature relationship network, the node's own feature information and the neighborhood node feature information are fused, the multidimensional feature vector of the time series is reconstructed, and a multidimensional state vector is generated.

[0025] In an optional embodiment,

[0026] Based on the multidimensional state vector, combined with the topological data analysis method, the high-dimensional topological space mapping of the multidimensional state vector is performed through nonlinear uncertainty mapping, including:

[0027] For each of the multidimensional state vectors, the Euclidean distance between the multidimensional state vector and the remaining multidimensional state vectors is calculated, and according to a preset distance threshold, the K multidimensional state vectors with the closest distance are selected to determine the neighbor state vectors, and the weight corresponding to the Euclidean distance is used as an edge to construct a neighbor graph;

[0028] Calculate the shortest path distance between any two nodes in the neighbor graph, determine the geodesic distance between the multidimensional state vectors, repeatedly traverse all nodes to generate a geodesic distance matrix, perform eigenvalue decomposition on the geodesic distance matrix, determine the eigenvector, select the eigenvector corresponding to the maximum eigenvalue of the corresponding number according to the preset eigenvalue threshold, construct a mapping matrix, and map the multidimensional state vector to the corresponding point in the popular space based on the mapping matrix to obtain a reduced-dimensional state vector;

[0029] Based on a preset distance threshold set, one distance threshold is selected each time, and in the popular space, it is determined that the distances between any two points are all less than the selected distance threshold, a neighborhood simplicial complex is constructed, and the distance threshold set is traversed in sequence, and the homology group corresponding to the neighborhood simplicial complex is calculated. Following the traversal process, the change of the homology group with the distance threshold is tracked to obtain a continuous homology barcode, and each bar in the continuous homology barcode corresponds to a coordinate in the high-dimensional topological space, and a continuous homology mapping is determined, and the reduced-dimensional state vector is mapped to the high-dimensional topological space to generate a high-dimensional topological representation.

[0030] In an optional embodiment,

[0031] The dynamic clustering algorithm is used to identify the operating mode of power plant equipment. By combining fuzzy logic with stochastic differential equations, the fuzzy area and mode mutation point of the equipment operating status are calculated. The dynamic evaluation results of the equipment health status include:

[0032] Based on high-dimensional topological space, a dynamic clustering algorithm is used to cluster the data points corresponding to the power plant equipment, a density threshold and a time decay factor are set as clustering parameters, the density of the data points is calculated, and the data points are assigned to different clusters in combination with time information, each of which corresponds to an operation mode;

[0033] Based on the operation mode, a fuzzy set is determined, and according to a pre-constructed membership function, a membership value of each data point under different operation modes is calculated, and based on the membership value, a fuzzy region of the operation state is determined;

[0034] Taking the operation mode as the steady-state solution of the stochastic differential equation, the fuzzy region of the operation state is collected, the random evolution trajectory of the operation state is determined, the parameters of the stochastic differential equation are set, the stochastic differential equation model is constructed, and the mode mutation point of the operation state is obtained by solving the stochastic differential equation;

[0035] The health status assessment result is determined based on the size corresponding to the fuzzy area of ​​the operating status and the frequency corresponding to the mode mutation point.

[0036] In an optional embodiment,

[0037] Based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, the health status of the equipment is predicted through multi-scale time series prediction to obtain a prediction result, and the uncertainty corresponding to the prediction result is quantified and solved based on the quantum probability graph model to generate a quantum probability distribution of the early warning signal. Based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to generate an intelligent early warning signal for the power plant equipment, including:

[0038] Based on the dynamic evaluation results of the equipment health status, a time series of equipment health status is generated, and the time series of equipment health status is decomposed according to different time scales to obtain time subsequences. At each time scale, the dynamic evolution trajectory corresponding to the time subsequence is determined, and through multiple hidden layers and recursive connections, long-term and short-term dependencies and multi-scale features are captured to construct a multi-level recursive neural network model;

[0039] Based on the multi-level recursive neural network model, the equipment health status is predicted in multi-scale time series, and the future health status prediction value is recursively generated at each time scale. The future health status prediction values ​​corresponding to multiple time scales are integrated to obtain the prediction result of the equipment health status.

[0040] Based on the quantum probability graph model, the prediction results are mapped to the quantum probability space, a quantum probability graph is constructed, the conditional dependency relationship and quantum entanglement results between multiple prediction results are determined, and based on the quantum probability reasoning algorithm, the quantum probability graph is iteratively evolved to calculate the quantum probability distribution of the prediction results;

[0041] Based on the quantum probability distribution of the prediction result, it is compared with the pre-set warning threshold, the quantum ground state corresponding to the warning signal is determined, the quantum ground state is superimposed, the square of the modulus of the superposition coefficient is calculated, the probability of the warning occurrence is determined, and the quantum probability distribution of the warning signal is generated;

[0042] Based on the quantum probability distribution of the warning signal, the real-time operation strategy of the equipment is adjusted, and the quantum probability distribution of the warning signal is visualized. Combined with the predetermined warning level and the corresponding operation content, the intelligent warning signal of the power plant equipment is output.

[0043] According to a second aspect of the embodiments of the present invention,

[0044] Provide a power plant intelligent early warning system based on big data, including:

[0045] The first unit is used to extract multi-scale features based on the operation data and environmental data of the power plant equipment through a multi-dimensional data stream adaptive hierarchical sampling method to generate a high-dimensional feature map, perform nonlinear mapping and sparse coding on the high-dimensional feature map through adaptive chaotic mapping to generate a multi-modal feature tensor, perform multi-level dynamic sampling on the multi-modal feature tensor based on a network community discovery algorithm and node importance determination, construct a feature relationship network, generate a time series multi-dimensional feature vector, adjust the structure of the feature relationship network based on topological entropy, and reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector;

[0046] The second unit is used to perform high-dimensional topological space mapping on the multidimensional state vector based on the multidimensional state vector and in combination with a topological data analysis method through nonlinear uncertainty mapping, identify the operation mode of the power plant equipment using a dynamic clustering algorithm, and calculate the fuzzy area and mode mutation point of the equipment operation state through a combination of fuzzy logic and stochastic differential equations to obtain a dynamic evaluation result of the equipment health state;

[0047] The third unit is used to predict the health status of the equipment based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, through multi-scale time series prediction, to obtain the prediction result, quantize the uncertainty corresponding to the prediction result based on the quantum probability graph model, generate the quantum probability distribution of the warning signal, adjust the real-time operation strategy of the equipment based on the quantum probability distribution, and generate the intelligent warning signal of the power plant equipment.

[0048] According to a third aspect of the embodiments of the present invention,

[0049] An electronic device is provided, comprising:

[0050] processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0053] According to a fourth aspect of the embodiments of the present invention,

[0054] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0055] In the embodiments of the present invention, through adaptive stratified sampling and multi-scale feature extraction methods, key features in power plant equipment and environmental data can be effectively captured, and the accuracy of data analysis can be improved; through adaptive chaotic mapping and sparse coding, high-dimensional feature maps can be converted into multimodal feature tensors, the expressive ability of features can be enhanced, and the model can better understand and utilize the complex relationships in the data; through network community discovery algorithms and node importance analysis, a multi-level feature relationship network is constructed to achieve accurate modeling of complex dependencies between features and enhance the prediction ability of the model; through topological data analysis and nonlinear uncertainty mapping, multi-dimensional state vectors can be effectively mapped to high-dimensional topological space to capture the complexity of equipment operation status. features, improving the accuracy of state representation; combining fuzzy logic and stochastic differential equations, it can calculate the fuzzy areas and mode mutation points of the equipment operating status, and provide more detailed and accurate dynamic evaluation results of the equipment health status; introducing the quantum probability graph model to quantify the uncertainty of the prediction results, making the early warning signal more scientific and explainable, and helping power plants better understand the credibility and risks of the prediction results; by generating the quantum probability distribution of the early warning signal, it can effectively identify potential risks, provide quantitative early warning information, and enhance the timeliness and reliability of the early warning; based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to achieve intelligent optimization of equipment operation, improve the safety and efficiency of the equipment, and reduce the possibility of failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of a power plant intelligent early warning method based on big data according to an embodiment of the present invention;

[0057] Figure 2 The figure is a schematic diagram of the structure of an intelligent early warning system for a power plant based on big data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of a power plant intelligent early warning method based on big data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] S101. Based on the operation data and environmental data of the power plant equipment, multi-scale features are extracted through a multi-dimensional data stream adaptive hierarchical sampling method to generate a high-dimensional feature map. The high-dimensional feature map is nonlinearly mapped and sparsely encoded through adaptive chaotic mapping to generate a multi-modal feature tensor. Based on the network community discovery algorithm, combined with the node importance determination, the multi-modal feature tensor is dynamically sampled at multiple levels to construct a feature relationship network and generate a time series multi-dimensional feature vector. Based on the topological entropy, the structure of the feature relationship network is adjusted to reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector;

[0062] In this embodiment, through adaptive stratified sampling and multi-scale feature extraction methods, key features in power plant equipment and environmental data can be effectively captured, and the accuracy of data analysis can be improved; through adaptive chaotic mapping and sparse coding, high-dimensional feature maps are converted into multimodal feature tensors, the expressive ability of features is enhanced, and the model can better understand and utilize complex relationships in the data; through network community discovery algorithms and node importance analysis, a multi-level feature relationship network is constructed to achieve accurate modeling of complex dependencies between features and enhance the prediction ability of the model; the feature relationship network structure is adjusted using topological entropy to ensure that the model can adapt to dynamic changes in time series data and improve the reliability and flexibility of prediction; by reconstructing the multidimensional feature vector of the time series into a multidimensional state vector, a systematic state monitoring and analysis method is provided, which helps to timely discover abnormalities or potential problems in equipment operation and improve the stability and safety of equipment operation.

[0063] In an optional embodiment, based on the operation data and environmental data of the power plant equipment, a multi-scale feature is extracted through a multi-dimensional data stream adaptive layered sampling method to generate a high-dimensional feature map, and the high-dimensional feature map is nonlinearly mapped and sparsely encoded through adaptive chaotic mapping to generate a multi-modal feature tensor, including:

[0064] The operation data and the environment data constitute a multidimensional data stream, and the multidimensional data stream is adaptively layered sampled. Based on the basic statistical properties of the multidimensional data stream and the corresponding dynamic changes, the local variance of the multidimensional data stream is calculated using a sliding window. Based on the comparison result of the local variance with a preset variance threshold, the number of sampling layers and the sampling rate corresponding to each layer are adaptively determined to obtain the sampled multidimensional data stream;

[0065] The multi-dimensional data stream after sampling is decomposed into multiple scales by using multi-wavelet transform, a multi-wavelet basis is constructed by tensor product, a wavelet coefficient threshold is adaptively determined, and a soft threshold processing is performed on the wavelet coefficient to obtain the processed wavelet coefficient, statistical features of the processed wavelet coefficient are calculated at different scales, and statistical features of different scales are combined to obtain a high-dimensional feature map, wherein the statistical features include mean, variance, skewness and kurtosis;

[0066] The high-dimensional feature map is divided into networks of different sizes, and the fractal dimension is calculated by linear fitting in a double logarithmic coordinate system, and a chaotic mapping function and corresponding control parameters are adaptively determined according to the fractal dimension, and the high-dimensional feature map is nonlinearly transformed, and chaotic diffusion and sparse representation are performed in the feature space to determine the feature map after chaotic mapping;

[0067] The feature map after chaotic mapping is divided into multiple scales to obtain sub-maps of different scales. At each scale, chaotic mapping is recursively applied to the sub-map to generate multi-scale sparse coding. The number of recursive layers is adaptively determined based on the fractal dimension corresponding to each sub-map. The multi-level sparse coding is adaptively determined by adjusting the control parameters of the corresponding chaotic mapping. The multi-scale sparse coding and the multi-level sparse coding are cross-scale fused to obtain a multi-modal feature tensor.

[0068] Perform high-order singular value decomposition on the multimodal feature tensor, decompose the multimodal feature tensor into multiple matrices and a core tensor, each matrix corresponds to a mode, determine the relationship between the modes, obtain singular values ​​and singular vectors, arrange the singular values ​​in order of size, generate a singular value distribution, determine a truncation threshold of the singular values, retain the singular values ​​and corresponding singular vectors greater than the truncation threshold, and reconstruct the multimodal feature tensor.

[0069] The power plant equipment operation data and environmental data are arranged in chronological order to form a multidimensional time series data stream, the size and step size of the sliding window are set, the window is slid on the multidimensional data stream, the local variance of the data in each window is calculated, a preset variance threshold is set, and the local variance is compared with the preset variance threshold. If the local variance is greater than the preset variance threshold, the number of sampling layers is increased and the sampling rate is increased within the time period corresponding to the window; if the local variance is less than the preset variance threshold, the number of sampling layers is reduced and the sampling rate is reduced. According to the adaptively determined number of sampling layers and sampling rate, the multidimensional data stream is layered sampled to obtain the sampled multidimensional data stream.

[0070] Select multi-wavelet basis functions, preferably such as Haar wavelet, Daubechies wavelet, etc., and construct multi-wavelet basis through tensor product; perform multi-scale decomposition on the sampled multidimensional data stream, and represent the data stream as wavelet coefficients at different scales; adaptively determine the wavelet coefficient threshold according to the statistical characteristics of the wavelet coefficients; perform soft threshold processing on the wavelet coefficients, set the coefficients with absolute values ​​less than the threshold to zero, and reduce the remaining coefficients proportionally; at each scale, calculate the statistical characteristics such as mean, variance, skewness and kurtosis of the processed wavelet coefficients; arrange the statistical characteristics of different scales in scale order, and combine them to form a high-dimensional feature map.

[0071] The high-dimensional feature map is divided into networks of different sizes, preferably 4×4 or 8×8; for each network, the cumulative probability distribution of the nodes in the network is calculated in a double logarithmic coordinate system; in the double logarithmic coordinate system, a linear fit is performed on the cumulative probability distribution, and the absolute value of the slope is the fractal dimension; according to the value range of the fractal dimension, a suitable chaos mapping function is selected from a predefined chaos mapping function library; according to the value of the fractal dimension, the control parameters of the chaos mapping function are adaptively adjusted to ensure the diffusion and sparse representation effects of the chaos mapping in the feature space; and the high-dimensional feature map is nonlinearly transformed using the determined chaos mapping function and control parameters to obtain a feature map after the chaos mapping.

[0072] The feature map after chaotic mapping is divided into multiple scales to obtain sub-maps of different scales; at each scale, chaotic mapping is recursively applied to the sub-map to generate multi-scale sparse codes; the fractal dimension of each sub-map is calculated, and the number of recursive layers is adaptively determined according to the fractal dimension, and the number of recursive layers is proportional to the fractal dimension; at each recursive layer, multi-level sparse codes are adaptively generated by adjusting the control parameters of the chaotic mapping function; the sparse codes of different scales and levels are combined in the order of scale and level to form a multimodal feature tensor.

[0073] A high-order singular value decomposition is performed on the multimodal feature tensor to decompose the tensor into multiple factor matrices and a core tensor; each factor matrix corresponds to a mode, and the column vectors of the matrix are the singular vectors of the mode; the elements of the core tensor are the interaction coefficients between the singular vectors of different modes, which characterize the mutual relationship between the modes; the elements of the core tensor are arranged in descending order to obtain the singular value distribution; according to the attenuation characteristics of the singular value distribution, the truncation threshold of the singular values ​​is adaptively determined; the singular values ​​and their corresponding singular vectors greater than the truncation threshold are retained, and the remaining singular values ​​and singular vectors are set to zero; the multimodal feature tensor is reconstructed using the truncated singular values ​​and singular vectors to obtain a compact representation after dimensionality reduction.

[0074] In this embodiment, through the adaptive stratified sampling technology, the sampling rate can be dynamically adjusted according to the local changes of the data, so as to realize the refined sampling of the power plant equipment operation data and the environmental data, and improve the representativeness and effectiveness of the data; the multidimensional data stream after sampling is decomposed into multiple scales by using multi-wavelet transform, the data features at different scales are extracted, and the high-dimensional feature map is constructed by statistical features to comprehensively capture the multi-scale information in the data; the high-dimensional feature map is nonlinearly transformed by chaotic mapping to enhance the sparse representation capability of the features and improve the recognition and separability of the features in complex systems; through multi-scale and multi-level sparse coding, the feature codes of different scales and levels are combined into a multi-modal feature tensor to realize the efficient fusion of multi-modal features and enhance the richness of feature expression; the multi-modal feature tensor is reduced in dimension by using high-order singular value decomposition technology to retain the most important feature information, reduce data redundancy, and improve the compactness and computational efficiency of feature representation.

[0075] In an optional embodiment, based on a network community discovery algorithm and in combination with node importance determination, the multimodal feature tensor is dynamically sampled at multiple levels to construct a feature relationship network, generate a time series multidimensional feature vector, and based on topological entropy, adjust the structure of the feature relationship network, and reconstruct the time series multidimensional feature vector into a multidimensional state vector, including:

[0076] The features in the multimodal feature tensor are used as nodes, and the correlation strength between the features is used as the weight of the edge to construct a weighted undirected graph. Based on the network community discovery algorithm, different resolution parameters are set to perform community division on the weighted undirected graph. Based on different resolution parameters, feature community structures of different granularities are obtained. Feature nodes are sorted according to the importance of nodes in each feature community structure. According to a preset sampling ratio, nodes of a corresponding ratio are sampled from each feature community structure to form a multi-level feature subset.

[0077] Calculate the correlation coefficient matrix between any two features in the feature subset, determine the initial correlation strength between the features, generate an initial feature relationship network, binarize the correlation coefficient matrix according to a preset binarization threshold by a threshold method, obtain an initial adjacency matrix of the initial feature relationship network, perform random walks on the initial feature relationship network, correct the initial correlation strength, and obtain a corrected feature relationship network;

[0078] Based on the multimodal feature tensor, extract the feature tensor slice corresponding to each timestamp, map the feature tensor slice into a single multidimensional feature vector according to the topological structure corresponding to the modified feature relationship network, generate a multidimensional feature vector under a timestamp by traversing all nodes in the modified feature relationship network, and repeatedly iterate multiple consecutive timestamps to generate a time series multidimensional feature vector;

[0079] Using topological entropy as an evaluation index and adopting a heuristic search algorithm, the optimal network structure adjustment strategy is searched, and the network structure with the minimum topological entropy is determined through iterative search, and the structure of the modified characteristic relationship network is adjusted to obtain the final characteristic relationship network;

[0080] Through the graph convolutional neural network, a convolution operation is performed on each node in the final feature relationship network, the node's own feature information and the neighborhood node feature information are fused, the multidimensional feature vector of the time series is reconstructed, and a multidimensional state vector is generated.

[0081] Each feature in the multimodal feature tensor is used as a node, and the strength of association between features is used as the weight of the edge to construct a weighted undirected graph. Select an appropriate network community discovery algorithm, preferably use the Louvain algorithm, set different resolution parameters, optionally 0.5, 1.0, 1.5, etc., and perform community division on the weighted undirected graph. Under each resolution parameter, a feature community structure of corresponding granularity is obtained. In each feature community structure, the feature nodes are sorted by importance based on the centrality measure of the node, preferably such as degree centrality. Set a preset sampling ratio, such as 20%, and sample the corresponding proportion of nodes from each feature community structure according to the order of node importance. Combine the node sets sampled under different resolution parameters to form a multi-level feature subset.

[0082] Calculate the Pearson correlation coefficient between any two features in the feature subset to form a correlation coefficient matrix. Use the correlation coefficient matrix as the initial correlation strength between the features to generate an initial feature relationship network. Set a preset binarization threshold, preferably 0.5, and binarize the correlation coefficient matrix. Set the correlation coefficient greater than or equal to the threshold to 1, and the correlation coefficient less than the threshold to 0 to obtain the initial adjacency matrix of the initial feature relationship network. Perform a random walk on the initial feature relationship network, and update the correlation strength between the nodes by the transition probability between the nodes. Repeat the random walk multiple times to obtain the revised feature correlation strength. According to the revised feature correlation strength, update the edge weights of the initial feature relationship network to obtain the revised feature relationship network.

[0083] When generating a time series multidimensional feature vector, based on the multimodal feature tensor, extract the feature tensor slice corresponding to each timestamp. According to the topological structure of the modified feature relationship network, map the feature tensor slice into a single multidimensional feature vector. Specifically, traverse all nodes in the modified feature relationship network, extract the corresponding eigenvalues ​​in order of the nodes, and form a multidimensional feature vector under a timestamp. Repeat the iteration of multiple consecutive timestamps to generate a time series multidimensional feature vector.

[0084] In order to optimize the structure of the feature relationship network, the topological entropy is used as an evaluation index to measure the complexity and heterogeneity of the feature relationship network. Select an appropriate heuristic search algorithm, such as a genetic algorithm, a particle swarm optimization algorithm, etc., and use the heuristic search algorithm to search for the optimal network structure adjustment strategy, such as adding or deleting edges, adjusting edge weights, etc. In each round of iteration, the feature relationship network is structurally adjusted and the adjusted topological entropy is calculated. Repeat multiple rounds of iterations to search for the network structure with the minimum topological entropy. According to the optimal adjustment strategy obtained by the search, the structure of the modified feature relationship network is adjusted to obtain the final feature relationship network.

[0085] A graph convolutional neural network is applied to generate a multidimensional state vector. The final feature relationship network is used as the input of the graph convolutional neural network. Each node in the final feature relationship network is convolved to fuse the node's own feature information and the feature information of the neighboring nodes. Feature representations of different scales are extracted through multi-layer stacking of graph convolutional layers. At the top layer of the graph convolutional neural network, a feature embedding vector for each node is generated. According to the feature order in the time series multidimensional feature vector, the feature embedding vectors of the corresponding nodes are sequentially spliced ​​to obtain the reconstructed multidimensional state vector. Multidimensional state vectors at multiple timestamps are repeatedly generated to form the final device health status representation.

[0086] In this embodiment, through community discovery algorithm and multi-level sampling, multimodal features can be refinedly extracted and screened, important features can be retained, redundancy can be reduced, and the effectiveness of feature representation can be improved; by constructing and correcting the feature relationship network, the correlation relationship between features can be accurately modeled and optimized, the representation ability of feature relationships is enhanced, and it is ensured that the network structure can reflect the real correlation between features; the generated time series multidimensional feature vector can dynamically reflect the changes in the health status of the equipment over time, and provide a reliable data basis for subsequent analysis and prediction; with topological entropy as an indicator, the heuristic search algorithm is used to optimize the structure of the feature relationship network, reduce the complexity of the network, and improve the structural rationality and computational efficiency of the network; through the graph convolutional neural network, the feature information is further extracted and fused to generate a multidimensional state vector, which improves the expression ability of the health status of the equipment and enhances the accuracy and robustness of the prediction model.

[0087] S102. Based on the multidimensional state vector, combined with the topological data analysis method, the multidimensional state vector is mapped into a high-dimensional topological space through nonlinear uncertainty mapping, the operation mode of the power plant equipment is identified using a dynamic clustering algorithm, and the fuzzy region and mode mutation point of the equipment operation state are calculated through the combination of fuzzy logic and stochastic differential equations to obtain a dynamic evaluation result of the equipment health state;

[0088] In this embodiment, through topological data analysis and nonlinear uncertainty mapping, the multidimensional state vector can be effectively mapped to a high-dimensional topological space, the complex characteristics of the equipment operating status can be captured, and the accuracy of the state representation can be improved; using the dynamic clustering algorithm, the different operating modes of the power plant equipment can be accurately identified, and the changes in the equipment operating status can be discovered in time, providing a basis for early warning and maintenance; combining fuzzy logic and stochastic differential equations, the fuzzy areas and mode mutation points of the equipment operating status can be calculated, providing more detailed and accurate dynamic evaluation results of the equipment health status; through nonlinear uncertainty mapping, the system's ability to handle equipment state uncertainty is enhanced, and the robustness and reliability of the evaluation results are improved.

[0089] In an optional embodiment, based on the multidimensional state vector, combined with a topological data analysis method, by nonlinear uncertainty mapping, high-dimensional topological space mapping of the multidimensional state vector includes:

[0090] For each of the multidimensional state vectors, the Euclidean distance between the multidimensional state vector and the remaining multidimensional state vectors is calculated, and according to a preset distance threshold, the K multidimensional state vectors with the closest distance are selected to determine the neighbor state vectors, and the weight corresponding to the Euclidean distance is used as an edge to construct a neighbor graph;

[0091] Calculate the shortest path distance between any two nodes in the neighbor graph, determine the geodesic distance between the multidimensional state vectors, repeatedly traverse all nodes to generate a geodesic distance matrix, perform eigenvalue decomposition on the geodesic distance matrix, determine the eigenvector, select the eigenvector corresponding to the maximum eigenvalue of the corresponding number according to the preset eigenvalue threshold, construct a mapping matrix, and map the multidimensional state vector to the corresponding point in the popular space based on the mapping matrix to obtain a reduced-dimensional state vector;

[0092] Based on a preset distance threshold set, one distance threshold is selected each time, and in the popular space, it is determined that the distances between any two points are all less than the selected distance threshold, a neighborhood simplicial complex is constructed, and the distance threshold set is traversed in sequence, and the homology group corresponding to the neighborhood simplicial complex is calculated. Following the traversal process, the change of the homology group with the distance threshold is tracked to obtain a continuous homology barcode, and each bar in the continuous homology barcode corresponds to a coordinate in the high-dimensional topological space, and a continuous homology mapping is determined, and the reduced-dimensional state vector is mapped to the high-dimensional topological space to generate a high-dimensional topological representation.

[0093] The neighborhood simplicial complex specifically refers to a distance-based topological structure used to characterize the geometric and topological properties of a set of points in a manifold space. Specifically, given a distance threshold ∈ and a point set P: any k+1 points in the point set form a k-dimensional simplex if and only if the distances between any two points in the point set are less than or equal to the threshold ∈. As the distance threshold changes, the structure of the simplicial complex will also change, and the structural change is used to analyze the topological properties of the data.

[0094] The homology group specifically refers to an algebraic structure used to characterize holes of different dimensions in a topological space, where the 0-dimensional homology group describes independent connected components; the 1-dimensional homology group describes rings in the space, i.e., one-dimensional "holes"; the 2-dimensional homology group describes two-dimensional "cavities" in the space; and the general k-dimensional homology group describes the topological features of k dimensions. The homology group reflects the number and type of topological features of data in different dimensions.

[0095] The persistent homology barcode specifically refers to a topological data analysis tool based on homology groups, which is used to analyze and represent the persistence of topological features of data at different scales. Each bar represents the "life cycle" of a topological feature - the entire process from appearance to disappearance; the length of the bar reflects the persistence and importance of the corresponding topological feature in the data. Longer bars correspond to more important and stable topological features, and short bars may correspond to noise. Persistent homology barcodes are used to analyze the shape and structural properties of data and generate topological representations by corresponding them to coordinates in a high-dimensional topological space.

[0096] For each multidimensional state vector, calculate the Euclidean distance between the vector and the remaining multidimensional state vectors. Euclidean distance is a measure of similarity between vectors, which is obtained by calculating the sum of squared differences in the corresponding dimensions of the vectors and then taking the square root. Set a preset distance threshold, select the K multidimensional state vectors closest to the current multidimensional state vector, and determine the neighboring state vector. The value of K can be determined based on the characteristics of the data and application requirements, and 5, 10, 20, etc. can usually be selected. Take the current multidimensional state vector as the central node in the neighbor graph, use the weight corresponding to the Euclidean distance as the edge, and take the neighbor state vector as the neighbor node to construct a neighbor graph. The neighbor graph reflects the local similarity relationship between multidimensional state vectors.

[0097] Compute the shortest path distance between any two nodes on the neighbor graph as the geodesic distance between the multidimensional state vectors. The geodesic distance is the shortest distance between two points on a manifold space, taking into account the intrinsic geometry of the manifold. Compute the shortest path distance between all pairs of nodes in the neighbor graph, optionally using the Floyd-Warshall algorithm. Repeat the traversal for all pairs of nodes to generate a geodesic distance matrix, which is symmetric and has zeros on the diagonal.

[0098] Perform eigenvalue decomposition on the geodesic distance matrix. Eigenvalue decomposition is to decompose a matrix into a combination of eigenvalues ​​and eigenvectors. The eigenvalues ​​represent the scaling factors of the matrix, and the eigenvectors represent the scaling directions. By calculating the eigenvalues ​​and eigenvectors of the geodesic distance matrix, the essential structural information between the multidimensional state vectors can be extracted. Set a preset eigenvalue threshold, select the eigenvectors corresponding to the corresponding number of maximum eigenvalues, and construct a mapping matrix. The choice of eigenvalue threshold can be determined based on the size distribution of the eigenvalues ​​and the need to retain the amount of information. Based on the mapping matrix, the multidimensional state vector is mapped to the corresponding point in the popular space to obtain a reduced-dimensional state vector. The popular space is a low-dimensional Euclidean space that retains the essential structural relationship between the multidimensional state vectors.

[0099] Based on the preset distance threshold set, a neighborhood simplicial complex is constructed in the manifold space. A simplicial complex is a topological structure used to describe the connectivity and hole structure between point sets. Each time a distance threshold is selected, it is determined that the distances between any two points in the manifold space are all less than the selected distance threshold, and the corresponding neighborhood simplicial complex is constructed. The distance threshold set can be an equally spaced numerical sequence, such as 0.1, 0.2, 0.3, etc., or it can be adaptively determined based on the distribution characteristics of the data. The distance threshold set is traversed in sequence to calculate the homology group corresponding to each neighborhood simplicial complex. The homology group is an algebraic tool for describing the hole structure in the simplicial complex, and homology groups of different dimensions correspond to holes of different dimensions. Following the traversal process of the distance threshold, the changes in the homology group are tracked to obtain a persistent homology barcode. The persistent homology barcode is a visualization tool that uses the length of the bar to represent the persistence of the homology group, revealing the stability of the topological structure with scale changes.

[0100] Each bar in the continuous homology barcode is mapped to a coordinate in the high-dimensional topological space to determine the continuous homology mapping. The continuous homology mapping maps points in the low-dimensional manifold space to the high-dimensional topological space, preserving the topological structure information of the data. Through the continuous homology mapping, the reduced-dimensional state vector is mapped to the high-dimensional topological space to generate a high-dimensional topological representation. The high-dimensional topological representation encodes the topological characteristics of the health status of the device in the form of a high-dimensional vector, providing a richer information representation for subsequent status assessment and anomaly detection.

[0101] In this embodiment, by constructing a neighbor graph, the local similarity relationship between multidimensional state vectors can be effectively captured, providing a basis for subsequent analysis; through geodesic distance and eigenvalue decomposition, the intrinsic geometric structure of the data is identified and retained, ensuring that the representation after dimensionality reduction still retains the essential characteristics of the data; the calculation of neighborhood simplicial complexes and homology groups reveals the topological structure characteristics of the data, especially the void information in dimension and scale, which is crucial to understanding the complexity and diversity of the data; the continuous homology barcode demonstrates the stability of the topological characteristics of the data through multi-scale analysis, helping to identify important topological structures and potential anomalies.

[0102] In an optional embodiment, a dynamic clustering algorithm is used to identify the operation mode of power plant equipment, and the fuzzy region and mode mutation point of the equipment operation state are calculated by combining fuzzy logic with stochastic differential equations to obtain a dynamic evaluation result of the equipment health state, including:

[0103] Based on high-dimensional topological space, a dynamic clustering algorithm is used to cluster the data points corresponding to the power plant equipment, a density threshold and a time decay factor are set as clustering parameters, the density of the data points is calculated, and the data points are assigned to different clusters in combination with time information, each of which corresponds to an operation mode;

[0104] Based on the operation mode, a fuzzy set is determined, and according to a pre-constructed membership function, a membership value of each data point under different operation modes is calculated, and based on the membership value, a fuzzy region of the operation state is determined;

[0105] Taking the operation mode as the steady-state solution of the stochastic differential equation, the fuzzy region of the operation state is collected, the random evolution trajectory of the operation state is determined, the parameters of the stochastic differential equation are set, the stochastic differential equation model is constructed, and the mode mutation point of the operation state is obtained by solving the stochastic differential equation;

[0106] The health status assessment result is determined based on the size corresponding to the fuzzy area of ​​the operating status and the frequency corresponding to the mode mutation point.

[0107] In the high-dimensional topological space, the dynamic clustering algorithm is used to cluster the data points corresponding to the power plant equipment. The dynamic clustering algorithm is a clustering method that can adapt to the dynamic changes of data streams. By setting the density threshold and time decay factor as clustering parameters, the clustering results can be updated in real time. The density threshold is used to determine whether a data point belongs to a certain cluster, and the time decay factor is used to control the degree of influence of historical data on the current cluster. By calculating the density of each data point and combining it with the time information, the data points are assigned to different clusters. Each cluster corresponds to a mode of operation of the power plant equipment, reflecting the operating characteristics of the equipment under different working conditions.

[0108] Based on the obtained operating mode, the corresponding fuzzy set is determined. Fuzzy set is a mathematical tool to describe the membership of elements, which can indicate the degree of belonging of elements in different sets. According to the pre-built membership function, the membership value of each data point in different operating modes is calculated. The membership function can be designed based on expert knowledge or data distribution characteristics. Usually, the membership function includes triangular function, trapezoidal function, Gaussian function, etc. Based on the membership value, the fuzzy region of the operating state is determined. The fuzzy region indicates the transition state of the data point between different operating modes, reflecting the uncertainty of the equipment operating state.

[0109] The operation mode is taken as the steady-state solution of the stochastic differential equation, and the stochastic evolution trajectory of the operation state is determined in combination with the fuzzy region of the operation state. The stochastic differential equation is a mathematical model that describes a random dynamic system, and the uncertainty of the system is characterized by introducing a random disturbance term. The parameters of the stochastic differential equation are set, such as drift terms and diffusion terms, to construct a stochastic differential equation model. By solving the stochastic differential equation, the mode mutation point of the operation state can be obtained. The mode mutation point indicates the moment when the operation state of the equipment changes significantly, which may indicate an abnormality or failure of the equipment.

[0110] The health status assessment result is determined based on the size of the fuzzy area of ​​the operating status and the frequency of the mode mutation points. The size of the fuzzy area of ​​the operating status reflects the degree of uncertainty of the equipment's operating status. The larger the fuzzy area, the more unstable the equipment's operating status. The frequency of mode mutation points indicates how frequently the equipment's operating status changes significantly. The higher the frequency, the more unstable the equipment's operating status. By comprehensively considering the size of the fuzzy area of ​​the operating status and the frequency of mode mutation points, the assessment result of the equipment's health status can be obtained. The assessment result can be a classification label, such as healthy, sub-healthy, abnormal, etc., or it can be a continuous health index indicating the degree of equipment health status.

[0111] In this embodiment, the dynamic clustering algorithm can update the clustering results in real time, adapt to the dynamic changes in the equipment operating status, and improve the processing ability of real-time data; the fuzzy area of ​​the equipment operating status is described by fuzzy sets, which can effectively deal with the uncertainty during the state transition period and enhance the accurate assessment of the equipment health status; different operating modes are identified as clustering results, which can reflect the characteristics of the equipment under different working conditions, thereby more comprehensively evaluating the equipment status; capturing the mutation points of the operating status through stochastic differential equations helps to detect equipment abnormalities or failures early and enhance prediction and early warning capabilities; combined with the analysis of fuzzy areas and mutation points, a comprehensive health status assessment result is provided, which can describe the equipment health status in detail and support equipment maintenance decisions.

[0112] S103. Based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, the health status of the equipment is predicted through multi-scale time series prediction to obtain a prediction result, and the uncertainty corresponding to the prediction result is quantified based on the quantum probability graph model to generate a quantum probability distribution of the warning signal. Based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to generate an intelligent warning signal for the power plant equipment.

[0113] In this embodiment, multi-scale time series prediction is performed through a multi-level recursive neural network, which can more accurately predict the health status of the equipment and capture the complex dynamic changes of the equipment operation status; the quantum probability graph model is introduced to quantify the uncertainty of the prediction results, making the early warning signal more scientific and explainable, helping the power plant to better understand the credibility and risk of the prediction results; by generating a quantum probability distribution of the early warning signal, potential risks can be effectively identified, quantitative early warning information can be provided, and the timeliness and reliability of the early warning can be enhanced; based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to achieve intelligent optimization of equipment operation, improve the safety and efficiency of the equipment, and reduce the possibility of failure; the intelligent early warning signal provides comprehensive decision-making support for the power plant, helping operators to take preventive measures when the risk is high to ensure the stable operation and safety of the equipment.

[0114] In an optional embodiment, based on the dynamic evaluation result of the health status of the equipment, combined with a multi-level recursive neural network, the health status of the equipment is predicted through multi-scale time series prediction to obtain a prediction result, and the uncertainty corresponding to the prediction result is quantified and solved based on a quantum probability graph model to generate a quantum probability distribution of a warning signal. Based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to generate an intelligent warning signal for power plant equipment, including:

[0115] Based on the dynamic evaluation results of the equipment health status, a time series of equipment health status is generated, and the time series of equipment health status is decomposed according to different time scales to obtain time subsequences. At each time scale, the dynamic evolution trajectory corresponding to the time subsequence is determined, and through multiple hidden layers and recursive connections, long-term and short-term dependencies and multi-scale features are captured to construct a multi-level recursive neural network model;

[0116] Based on the multi-level recursive neural network model, the equipment health status is predicted in multi-scale time series, and the future health status prediction value is recursively generated at each time scale. The future health status prediction values ​​corresponding to multiple time scales are integrated to obtain the prediction result of the equipment health status.

[0117] Based on the quantum probability graph model, the prediction results are mapped to the quantum probability space, a quantum probability graph is constructed, the conditional dependency relationship and quantum entanglement results between multiple prediction results are determined, and based on the quantum probability reasoning algorithm, the quantum probability graph is iteratively evolved to calculate the quantum probability distribution of the prediction results;

[0118] Based on the quantum probability distribution of the prediction result, it is compared with the pre-set warning threshold, the quantum ground state corresponding to the warning signal is determined, the quantum ground state is superimposed, the square of the modulus of the superposition coefficient is calculated, the probability of the warning occurrence is determined, and the quantum probability distribution of the warning signal is generated;

[0119] Based on the quantum probability distribution of the warning signal, the real-time operation strategy of the equipment is adjusted, and the quantum probability distribution of the warning signal is visualized. Combined with the predetermined warning level and the corresponding operation content, the intelligent warning signal of the power plant equipment is output.

[0120] Construct a multi-level recursive neural network model. Considering the time dependency and multi-scale characteristics of the equipment health status, a multi-level recursive neural network is used to model the equipment health status. The equipment health status time series is decomposed according to different time scales to obtain subsequences representing different frequency components. At each time scale, a recursive neural network is used to characterize the dynamic evolution of the subsequence. By setting multiple hidden layers and recursive connections, the long-term and short-term dependencies and multi-scale characteristics of the health status time series are captured.

[0121] Perform multi-scale time series prediction. Use the trained multi-level recursive neural network model to predict the future evolution trend of the equipment health status. At each time scale, the recursive neural network recursively generates the health status prediction value for a period of time in the future based on the historical health status data. By fusing the prediction results of multiple time scales, the multi-scale time series prediction results of the equipment health status are obtained. The prediction results include not only the mean of the future health status, but also the confidence interval of the health status to reflect the uncertainty of the prediction.

[0122] Quantify prediction uncertainty based on quantum probability graph model. In view of the uncertainty of prediction results of multi-level recurrent neural network, the quantum probability graph model is introduced for quantitative analysis. The prediction results are regarded as quantum states, and the randomness and uncertainty in the prediction process are mapped to the quantum probability space. By constructing a quantum probability graph, the conditional dependency relationship and quantum entanglement effect between the prediction results are characterized. Using quantum probability inference algorithms, such as quantum belief propagation algorithm or quantum Markov chain Monte Carlo method, the quantum probability graph is inferred to calculate the quantum probability distribution of the prediction results. This distribution can quantify the uncertainty of the prediction results and provide a probabilistic basis for subsequent early warning decisions.

[0123] Generate the quantum probability distribution of intelligent early warning signals. According to the quantum probability distribution of the equipment health status prediction results, judge the probability of equipment failure or abnormality in the future. Set the early warning threshold, compare the quantum probability distribution of the prediction results with the threshold, and generate the quantum probability distribution of the intelligent early warning signal. The quantum state of the early warning signal can be expressed as the superposition of the two base states of early warning and non-earning, and the square of the modulus of the superposition coefficient represents the probability of the early warning. By analyzing the quantum probability distribution of the early warning signal, the credibility and urgency of the early warning can be quantified.

[0124] Adjust the real-time operation strategy of the equipment. Based on the quantum probability distribution of the intelligent early warning signal, dynamically adjust the real-time operation strategy of the equipment to achieve intelligent optimization and decision support for equipment operation. According to the level of early warning probability, adopt different operation strategies: when the early warning probability is low, maintain normal operation of the equipment; when the early warning probability is high, adjust the equipment operation parameters in time, reduce the load or perform preventive maintenance; when the early warning probability exceeds the safety threshold, consider shutting down for maintenance to avoid accidents. By adjusting the operation strategy under the guidance of quantum probability distribution, the safety, reliability and efficiency of equipment operation can be improved.

[0125] Output intelligent early warning signals for power plant equipment. Visualize the generated intelligent early warning signal quantum probability distribution to display early warning information in an intuitive and easy-to-understand way. Combined with the quantum state of the early warning signal, give early warning levels and recommended measures. At the same time, correlate the early warning information with the equipment's historical operating data, health status assessment results, etc., generate a comprehensive early warning report, and provide decision-making references for power plant operators. The output of the early warning signal can be in the form of visual charts, early warning lights, SMS notifications, etc., to achieve timely transmission and sharing of early warning information.

[0126] In this embodiment, multi-scale modeling and prediction of the health status of the equipment are performed through a multi-level recursive neural network, which can more accurately capture the long-term and short-term dependencies and multi-scale characteristics of the health status of the equipment, thereby improving the accuracy and reliability of the prediction; a quantum probability graph model is introduced to quantify the uncertainty of the prediction results, so that the prediction includes not only the mean but also the confidence interval, providing more comprehensive prediction information, which is helpful to better evaluate the future evolution trend of the equipment status; intelligent early warning signals are generated through quantum probability distribution, which can quantify the credibility and urgency of the early warning, improve the accuracy and effectiveness of the early warning signals, and help power plants to identify potential risks in a timely manner; according to the quantum probability distribution of the early warning signal, the equipment operation strategy is dynamically adjusted to realize intelligent equipment operation management, thereby improving the safety, reliability and operation efficiency of the equipment; by combining the intelligent early warning signal with the equipment historical data and health assessment results, comprehensive decision support is provided to help power plant operators make better preventive maintenance and emergency decisions; the quantum probability distribution of the early warning signal is visualized to make the early warning information more intuitive and easy to understand, ensuring that the early warning information can be transmitted and shared in a timely manner, thereby improving the response speed and overall operation efficiency of the power plant.

[0127] Figure 2 FIG. 1 is a schematic diagram of the structure of a power plant intelligent early warning system based on big data according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0128] The first unit is used to extract multi-scale features based on the operation data and environmental data of the power plant equipment through a multi-dimensional data stream adaptive hierarchical sampling method to generate a high-dimensional feature map, perform nonlinear mapping and sparse coding on the high-dimensional feature map through adaptive chaotic mapping to generate a multi-modal feature tensor, perform multi-level dynamic sampling on the multi-modal feature tensor based on a network community discovery algorithm and node importance determination, construct a feature relationship network, generate a time series multi-dimensional feature vector, adjust the structure of the feature relationship network based on topological entropy, and reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector;

[0129] The second unit is used to perform high-dimensional topological space mapping on the multidimensional state vector based on the multidimensional state vector and in combination with a topological data analysis method through nonlinear uncertainty mapping, identify the operation mode of the power plant equipment using a dynamic clustering algorithm, and calculate the fuzzy area and mode mutation point of the equipment operation state through a combination of fuzzy logic and stochastic differential equations to obtain a dynamic evaluation result of the equipment health state;

[0130] The third unit is used to predict the health status of the equipment based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, through multi-scale time series prediction, to obtain the prediction result, quantize the uncertainty corresponding to the prediction result based on the quantum probability graph model, generate the quantum probability distribution of the warning signal, adjust the real-time operation strategy of the equipment based on the quantum probability distribution, and generate the intelligent warning signal of the power plant equipment.

[0131] According to a third aspect of the embodiments of the present invention,

[0132] An electronic device is provided, comprising:

[0133] processor;

[0134] a memory for storing processor-executable instructions;

[0135] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0136] According to a fourth aspect of the embodiments of the present invention,

[0137] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0138] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The intelligent early warning method for power plants based on big data is characterized by: include: Based on the operation data and environmental data of the power plant equipment, multi-scale features are extracted through a multi-dimensional data stream adaptive hierarchical sampling method to generate a high-dimensional feature map. The high-dimensional feature map is nonlinearly mapped and sparsely encoded through adaptive chaotic mapping to generate a multi-modal feature tensor. Based on a network community discovery algorithm and in combination with node importance determination, the multi-modal feature tensor is dynamically sampled at multiple levels to construct a feature relationship network and generate a time series multi-dimensional feature vector. Based on topological entropy, the structure of the feature relationship network is adjusted to reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector. Based on the multidimensional state vector, combined with the topological data analysis method, the multidimensional state vector is mapped into a high-dimensional topological space through nonlinear uncertainty mapping, the operation mode of the power plant equipment is identified using a dynamic clustering algorithm, and the fuzzy area and mode mutation point of the equipment operation state are calculated through the combination of fuzzy logic and stochastic differential equations to obtain a dynamic evaluation result of the equipment health state; Based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, the health status of the equipment is predicted through multi-scale time series prediction to obtain a prediction result. The uncertainty corresponding to the prediction result is quantified based on the quantum probability graph model to generate a quantum probability distribution of the warning signal. Based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to generate an intelligent warning signal for the power plant equipment.

2. The method according to claim 1, characterized in that: Based on the operation data and environmental data of power plant equipment, multi-scale features are extracted through a multi-dimensional data stream adaptive layered sampling method to generate a high-dimensional feature map. The high-dimensional feature map is nonlinearly mapped and sparsely encoded through adaptive chaotic mapping to generate a multi-modal feature tensor including: The operation data and the environment data constitute a multidimensional data stream, and the multidimensional data stream is adaptively layered sampled. Based on the basic statistical properties of the multidimensional data stream and the corresponding dynamic changes, the local variance of the multidimensional data stream is calculated using a sliding window. Based on the comparison result of the local variance with a preset variance threshold, the number of sampling layers and the sampling rate corresponding to each layer are adaptively determined to obtain the sampled multidimensional data stream; The multi-dimensional data stream after sampling is decomposed into multiple scales by using multi-wavelet transform, a multi-wavelet basis is constructed by tensor product, a wavelet coefficient threshold is adaptively determined, and a soft threshold processing is performed on the wavelet coefficient to obtain the processed wavelet coefficient, statistical features of the processed wavelet coefficient are calculated at different scales, and statistical features of different scales are combined to obtain a high-dimensional feature map, wherein the statistical features include mean, variance, skewness and kurtosis; The high-dimensional feature map is divided into networks of different sizes, and the fractal dimension is calculated by linear fitting in a double logarithmic coordinate system, and a chaotic mapping function and corresponding control parameters are adaptively determined according to the fractal dimension, and the high-dimensional feature map is nonlinearly transformed, and chaotic diffusion and sparse representation are performed in the feature space to determine the feature map after chaotic mapping; The feature map after chaotic mapping is divided into multiple scales to obtain sub-maps of different scales. At each scale, chaotic mapping is recursively applied to the sub-map to generate multi-scale sparse coding. The number of recursive layers is adaptively determined based on the fractal dimension corresponding to each sub-map. The multi-level sparse coding is adaptively determined by adjusting the control parameters of the corresponding chaotic mapping. The multi-scale sparse coding and the multi-level sparse coding are cross-scale fused to obtain a multi-modal feature tensor. Perform high-order singular value decomposition on the multimodal feature tensor, decompose the multimodal feature tensor into multiple matrices and a core tensor, each matrix corresponds to a mode, determine the relationship between the modes, obtain singular values ​​and singular vectors, arrange the singular values ​​in order of size, generate a singular value distribution, determine a truncation threshold of the singular values, retain the singular values ​​and corresponding singular vectors greater than the truncation threshold, and reconstruct the multimodal feature tensor.

3. The method according to claim 1, characterized in that Based on the network community discovery algorithm, combined with the determination of node importance, the multi-modal feature tensor is dynamically sampled at multiple levels to construct a feature relationship network, generate a time series multi-dimensional feature vector, and based on the topological entropy, adjust the structure of the feature relationship network, and reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector, including: The features in the multimodal feature tensor are used as nodes, and the correlation strength between the features is used as the weight of the edge to construct a weighted undirected graph. Based on the network community discovery algorithm, different resolution parameters are set to perform community division on the weighted undirected graph. Based on different resolution parameters, feature community structures of different granularities are obtained. Feature nodes are sorted according to the importance of nodes in each feature community structure. According to a preset sampling ratio, nodes of a corresponding ratio are sampled from each feature community structure to form a multi-level feature subset. Calculate the correlation coefficient matrix between any two features in the feature subset, determine the initial correlation strength between the features, generate an initial feature relationship network, binarize the correlation coefficient matrix according to a preset binarization threshold by a threshold method, obtain an initial adjacency matrix of the initial feature relationship network, perform random walks on the initial feature relationship network, correct the initial correlation strength, and obtain a corrected feature relationship network; Based on the multimodal feature tensor, extract the feature tensor slice corresponding to each timestamp, map the feature tensor slice into a single multidimensional feature vector according to the topological structure corresponding to the modified feature relationship network, generate a multidimensional feature vector under a timestamp by traversing all nodes in the modified feature relationship network, and repeatedly iterate multiple consecutive timestamps to generate a time series multidimensional feature vector; Using topological entropy as an evaluation index and adopting a heuristic search algorithm, the optimal network structure adjustment strategy is searched, and the network structure with the minimum topological entropy is determined through iterative search, and the structure of the modified characteristic relationship network is adjusted to obtain the final characteristic relationship network; Through the graph convolutional neural network, a convolution operation is performed on each node in the final feature relationship network, the node's own feature information and the neighborhood node feature information are fused, the multidimensional feature vector of the time series is reconstructed, and a multidimensional state vector is generated.

4. The method according to claim 1, characterized in that Based on the multidimensional state vector, combined with the topological data analysis method, the high-dimensional topological space mapping of the multidimensional state vector is performed through nonlinear uncertainty mapping, including: For each of the multidimensional state vectors, the Euclidean distance between the multidimensional state vector and the remaining multidimensional state vectors is calculated, and according to a preset distance threshold, the K multidimensional state vectors with the closest distance are selected to determine the neighbor state vectors, and the weight corresponding to the Euclidean distance is used as an edge to construct a neighbor graph; Calculate the shortest path distance between any two nodes in the neighbor graph, determine the geodesic distance between the multidimensional state vectors, repeatedly traverse all nodes to generate a geodesic distance matrix, perform eigenvalue decomposition on the geodesic distance matrix, determine the eigenvector, select the eigenvector corresponding to the maximum eigenvalue of the corresponding number according to the preset eigenvalue threshold, construct a mapping matrix, and map the multidimensional state vector to the corresponding point in the popular space based on the mapping matrix to obtain a reduced-dimensional state vector; Based on a preset distance threshold set, one distance threshold is selected each time, and in the popular space, it is determined that the distances between any two points are all less than the selected distance threshold, a neighborhood simplicial complex is constructed, and the distance threshold set is traversed in sequence, and the homology group corresponding to the neighborhood simplicial complex is calculated. Following the traversal process, the change of the homology group with the distance threshold is tracked to obtain a continuous homology barcode, and each bar in the continuous homology barcode corresponds to a coordinate in the high-dimensional topological space, and a continuous homology mapping is determined, and the reduced-dimensional state vector is mapped to the high-dimensional topological space to generate a high-dimensional topological representation.

5. The method according to claim 1, characterized in that The dynamic clustering algorithm is used to identify the operating mode of power plant equipment. By combining fuzzy logic with stochastic differential equations, the fuzzy area and mode mutation point of the equipment operating status are calculated. The dynamic evaluation results of the equipment health status include: Based on high-dimensional topological space, a dynamic clustering algorithm is used to cluster the data points corresponding to the power plant equipment, a density threshold and a time decay factor are set as clustering parameters, the density of the data points is calculated, and the data points are assigned to different clusters in combination with time information, each of which corresponds to an operation mode; Based on the operation mode, a fuzzy set is determined, and according to a pre-constructed membership function, a membership value of each data point under different operation modes is calculated, and based on the membership value, a fuzzy region of the operation state is determined; Taking the operation mode as the steady-state solution of the stochastic differential equation, the fuzzy region of the operation state is collected, the random evolution trajectory of the operation state is determined, the parameters of the stochastic differential equation are set, the stochastic differential equation model is constructed, and the mode mutation point of the operation state is obtained by solving the stochastic differential equation; The health status assessment result is determined based on the size corresponding to the fuzzy area of ​​the operating status and the frequency corresponding to the mode mutation point.

6. The method according to claim 1, characterized in that Based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, the health status of the equipment is predicted through multi-scale time series prediction to obtain a prediction result, and the uncertainty corresponding to the prediction result is quantified and solved based on the quantum probability graph model to generate a quantum probability distribution of the early warning signal. Based on the quantum probability distribution, the real-time operation strategy of the equipment is adjusted to generate an intelligent early warning signal for the power plant equipment, including: Based on the dynamic evaluation results of the equipment health status, a time series of equipment health status is generated, and the time series of equipment health status is decomposed according to different time scales to obtain time subsequences. At each time scale, the dynamic evolution trajectory corresponding to the time subsequence is determined, and through multiple hidden layers and recursive connections, long-term and short-term dependencies and multi-scale features are captured to construct a multi-level recursive neural network model; Based on the multi-level recursive neural network model, the equipment health status is predicted in multi-scale time series, and the future health status prediction value is recursively generated at each time scale. The future health status prediction values ​​corresponding to multiple time scales are integrated to obtain the prediction result of the equipment health status. Based on the quantum probability graph model, the prediction results are mapped to the quantum probability space, a quantum probability graph is constructed, the conditional dependency relationship and quantum entanglement results between multiple prediction results are determined, and based on the quantum probability reasoning algorithm, the quantum probability graph is iteratively evolved to calculate the quantum probability distribution of the prediction results; Based on the quantum probability distribution of the prediction result, it is compared with the pre-set warning threshold, the quantum ground state corresponding to the warning signal is determined, the quantum ground state is superimposed, the square of the modulus of the superposition coefficient is calculated, the probability of the warning occurrence is determined, and the quantum probability distribution of the warning signal is generated; Based on the quantum probability distribution of the warning signal, the real-time operation strategy of the equipment is adjusted, and the quantum probability distribution of the warning signal is visualized. Combined with the predetermined warning level and the corresponding operation content, the intelligent warning signal of the power plant equipment is output.

7. A power plant intelligent early warning system based on big data, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to extract multi-scale features based on the operation data and environmental data of the power plant equipment through a multi-dimensional data stream adaptive hierarchical sampling method to generate a high-dimensional feature map, perform nonlinear mapping and sparse coding on the high-dimensional feature map through adaptive chaotic mapping to generate a multi-modal feature tensor, perform multi-level dynamic sampling on the multi-modal feature tensor based on a network community discovery algorithm and node importance determination, construct a feature relationship network, generate a time series multi-dimensional feature vector, adjust the structure of the feature relationship network based on topological entropy, and reconstruct the time series multi-dimensional feature vector into a multi-dimensional state vector; The second unit is used to perform high-dimensional topological space mapping on the multidimensional state vector based on the multidimensional state vector and in combination with a topological data analysis method through nonlinear uncertainty mapping, identify the operation mode of the power plant equipment using a dynamic clustering algorithm, and calculate the fuzzy area and mode mutation point of the equipment operation state through a combination of fuzzy logic and stochastic differential equations to obtain a dynamic evaluation result of the equipment health state; The third unit is used to predict the health status of the equipment based on the dynamic evaluation results of the health status of the equipment, combined with a multi-level recursive neural network, through multi-scale time series prediction, to obtain the prediction result, quantize the uncertainty corresponding to the prediction result based on the quantum probability graph model, generate the quantum probability distribution of the warning signal, adjust the real-time operation strategy of the equipment based on the quantum probability distribution, and generate the intelligent warning signal of the power plant equipment.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images