A power plant automation operation and maintenance management system based on big data analysis

Abstract

The application discloses a power plant automatic operation and maintenance management system based on big data analysis, comprising the following steps: the system collects equipment operation, environment and operation and maintenance record data, constructs a standardized time sequence, adopts a subspace tracking algorithm to update an operation state characteristic structure in real time, identifies operation characteristic deviation, combines with Granger causality analysis, constructs a causality weight matrix between time sequence variables and embeds a modeling process, realizes dynamic subspace structure adjustment, the system further identifies state abnormal sections based on subspace angle change, combines with a causality path to construct a fault link, realizes accurate identification of power plant operation and maintenance states and intelligent presentation of potential problem chains. The application is suitable for operation and maintenance management scenes of complex systems of power plants.

Description

Technical Field

[0001] This invention relates to the field of power system operation and maintenance technology, and in particular to an automated operation and maintenance management system for power plants based on big data analysis. Background Technology

[0002] With the continuous development of digital transformation and intelligent operation and maintenance systems in power plants, large-scale equipment status awareness and fault diagnosis based on multi-source data fusion have become core requirements for ensuring the reliability of power systems. Power plant operation and maintenance management generally relies on data collected by a large number of monitoring devices and sensors, which are then analyzed and alerted in real time through edge computing or back-end systems. In practical applications, due to the high dimensionality, rapid changes, and complex causal relationships of equipment operation data, existing data-driven analysis methods still have significant limitations in scenarios such as sudden changes in operating conditions, the initial stage of faults, and multivariate co-evolution.

[0003] On the one hand, traditional fixed-structure models are difficult to adapt to the dynamic changes in equipment operating status and cannot respond in a timely manner to the drift of high-dimensional feature space, resulting in delayed anomaly identification and high misjudgment rate. On the other hand, most existing causal analysis methods are decoupled from the feature modeling process, lack the ability to model the dynamic causal influence between variables, and cannot build traceable fault propagation paths, which limits the system's ability to locate the root cause of the fault and the operation and maintenance support at the link level.

[0004] Therefore, how to provide a power plant automation operation and maintenance management system based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a power plant automated operation and maintenance management system based on big data analysis. This invention adopts a method combining subspace tracking algorithm and Granger causality analysis to construct a dynamic operating feature subspace and identify the causal structure between variables. It systematically realizes data acquisition, state modeling, anomaly detection and fault link construction, and has the advantages of lightweight model, strong dynamic adaptability and interpretable fault diagnosis.

[0006] An automated operation and maintenance management system for power plants based on big data analysis, according to an embodiment of the present invention, includes the following steps:

[0007] The data acquisition module is used to collect operating parameters of power plant equipment, environmental monitoring data and operation and maintenance records, and generate multi-source raw data sequences with a unified time identifier;

[0008] The data preprocessing module is used to receive multi-source raw data sequences, perform preprocessing, and output standardized time series data tensors.

[0009] The subspace construction module receives standardized time series data tensors, constructs the initial subspace structure of the running state through the subspace tracking algorithm, and iteratively updates it within a continuous time window to generate a subspace structure sequence.

[0010] The causal relationship identification module is used to receive standardized time series data tensors, perform Granger causal analysis, construct a causal graph structure between time series variables, and output a causal weight matrix.

[0011] The subspace fusion module is used to receive the subspace structure sequence and the causal weight matrix, embed the causal weights into the subspace update process, and generate a dynamic subspace structure sequence under causal constraints.

[0012] The state recognition module is used to receive the dynamic subspace structure sequence, calculate the subspace angle change between consecutive time points, and determine the state offset time period based on the subspace angle change.

[0013] The link construction module receives the state offset time period and the causal weight matrix, combines the variable indices that have shifted in the subspace structure sequence, establishes a causal variable path graph related to the state offset, and outputs the fault link structure.

[0014] Optionally, the data acquisition module includes:

[0015] The operating parameter acquisition unit is used to acquire operating parameters from the power plant equipment's operation monitoring terminal. The operating parameters include equipment speed, voltage, current, power, frequency, and equipment operating status indicators.

[0016] The environmental monitoring data acquisition unit is used to acquire environmental monitoring data from environmental monitoring sensors. The environmental monitoring data includes ambient temperature, humidity, emission monitoring signals and external air pressure measurements.

[0017] The operation and maintenance record collection unit is used to obtain operation and maintenance records related to the equipment operation cycle from the power plant operation and maintenance information system. The operation and maintenance records include maintenance time markers, fault record entries, maintenance task information and spare parts registration information.

[0018] All types of data output from the operation parameter acquisition unit, environmental monitoring acquisition unit, and operation and maintenance record acquisition unit are accompanied by a unified time identifier and are collected in a structured form as a multi-source raw data sequence.

[0019] Optional preprocessing includes: data cleaning, missing data imputation, time alignment, and format normalization.

[0020] Optionally, subspace building blocks include:

[0021] The feature vector generation unit is used to extract multiple fields related to the operating status of power plant equipment from the standardized time series data tensor, construct a continuous feature vector set in time order, and output a feature matrix for subspace modeling.

[0022] An initial subspace modeling unit is used to receive the feature matrix, extract the first k principal component vectors through principal component analysis, and perform orthogonalization on the principal component vectors to construct the initial subspace structure, wherein the value of k is determined according to the cumulative variance contribution rate. The initial subspace structure is used as the initialization input for the subspace tracking algorithm.

[0023] The subspace trend prediction unit is used to receive the initial subspace structure and the subspace structure generated at each time step iteration, construct the subspace structure change sequence, and perform vector direction fitting based on the subspace structure change sequence to generate the predicted subspace structure.

[0024] The projection residual calculation unit is used to receive the feature vector set of the current time step and the subspace structure of the previous time step, project the feature vector set onto the previous subspace structure and calculate the residual vector, which is used to measure the difference between the current input and the subspace representation.

[0025] The noise stripping unit receives the residual vector, performs energy spectrum analysis and threshold decomposition, divides the residual vector into signal residual and noise residual, and outputs only the signal residual to the subspace update unit as the basis for effective update.

[0026] The subspace update unit is used to receive the previous subspace structure, the current signal residual and the predicted subspace structure, perform QR decomposition driven by the signal residual to obtain the candidate updated subspace structure, and then weight and fuse the candidate updated subspace structure and the predicted subspace structure according to the set weight to output the subspace structure of the current time step.

[0027] The sequence generation unit and the subspace update unit perform update operations multiple times within a continuous time window to form a subspace structure sequence that changes over time.

[0028] Optionally, the causal relationship identification module includes:

[0029] The lag order determination unit is used to receive the standardized time series data tensor, construct a candidate set of lag orders for any two time series variables in the standardized time series data tensor, calculate the evaluation value corresponding to each lag order based on the preset Bayesian information criterion, and select the lag order with the smallest evaluation value as the optimal lag order of the time series variable pair composed of any two time series variables.

[0030] The restricted regression modeling unit is used to receive the optimal lag order, construct a restricted regression model for the currently analyzed time series variable based on the optimal lag order, which only contains the lagged terms of the corresponding time series variable itself, and calculate the residual sum of squares of the restricted regression model.

[0031] The unrestricted regression modeling unit is used to construct an unrestricted regression model for the currently analyzed time series variable based on the optimal lag order, which includes the lag term of the corresponding time series variable itself and the lag term of another time series variable, and to calculate the sum of squared residuals of the unrestricted regression model.

[0032] The significance determination unit is used to receive the residual sum of squares of the restricted regression model and the residual sum of squares of the unrestricted regression model, perform differential statistical analysis based on the preset sample size and the number of model parameters, and determine whether the difference between the two regression models is significant according to the set significance level. When the significance determination result rejects the no-causal hypothesis, it records that there is a Granger causal relationship between the current time series variables.

[0033] The causal weight matrix generation unit is used to receive all time series variable pairs with Granger causal relationships, construct a two-dimensional matrix based on the variable index, mark the time series variable pairs with causal relationships as non-zero elements, and assign weight values ​​according to the corresponding statistical test indicators to form a causal weight matrix.

[0034] The causal graph structure construction unit is used to receive the causal weight matrix, map each time series variable in the causal weight matrix to a node in the causal graph structure, map the time series variable pairs corresponding to non-zero elements to directed edges, the direction of the directed edges is from the dependent variable to the affected variable, the edge weight of the directed edges is generated using the corresponding values ​​in the causal weight matrix, and output the causal graph structure between time series variables.

[0035] Optionally, the subspace fusion module includes:

[0036] The feature dimension correspondence unit is used to map each column of feature vectors in the subspace structure of the current time step to the time series variables in the standardized time series data tensor, and establish the correspondence between feature vectors and variable indices.

[0037] The causal weight extraction unit is used to extract the causal influence factors corresponding to each feature vector from the causal weight matrix based on the correspondence between feature vectors and variable indices, forming a set of causal weight factors consistent with the number of column vectors in the subspace structure.

[0038] The causal weighted adjustment unit is used to receive the candidate updated subspace structure and the predicted subspace structure during the candidate updated subspace structure generation process, and to perform weighted transformation on each column feature vector of the candidate updated subspace structure according to its corresponding causal influence factor to form an intermediate subspace structure with embedded causal adjustment.

[0039] The fusion update unit is used to linearly weight the intermediate subspace structure with embedded causal adjustment and the prediction subspace structure according to a preset fusion ratio to generate the fusion update subspace structure of the current time step.

[0040] The structure sequence generation unit is used to take the fusion update subspace structure of the current time step as the final subspace structure of the corresponding time step, and iterate and update it according to the continuous time steps. The final subspace structures under all time steps are stored in sequence to generate a dynamic subspace structure sequence.

[0041] Optionally, the status recognition module includes:

[0042] The subspace structure pairing unit is used to receive the dynamic subspace structure sequence and extract the subspace structure of any two adjacent time steps to construct the subspace structure pair sequence.

[0043] Angle change extraction unit is used to perform vector space similarity analysis on each pair of subspace structures, calculate the angle change between subspaces based on the directional differences between corresponding orthogonal feature vectors, and bind the angle change to the time step index;

[0044] Angle change sequence construction unit is used to arrange angle changes in chronological order to construct a subspace angle change sequence, which reflects the temporal evolution characteristics of the dynamic subspace structure.

[0045] An anomaly marker generation unit is used to receive the angle change sequence and mark the time steps that are greater than a preset angle threshold as an anomaly time step.

[0046] The offset segment identification unit is used to perform continuity analysis on abnormal time steps, identify continuous subsequences with an interval of less than or equal to one time step between abnormal time steps, and delineate the corresponding time segments as candidate state offset segments.

[0047] The offset time period output unit is used to perform minimum duration verification on candidate state offset segments and output time segments that meet the duration requirements as state offset time periods.

[0048] Optionally, the link building module includes:

[0049] The offset variable identification unit is used to receive the dynamic subspace structure sequence and the state offset time period, extract the subspace structure corresponding to each time step in the state offset time period, perform directional change analysis on each column feature vector, and when the directional difference of the same column feature vector in adjacent time steps exceeds the preset offset judgment threshold, record the variable index corresponding to the corresponding column and summarize it to form an offset variable index set.

[0050] The causal path extraction unit is used to receive the offset variable index set and the causal weight matrix. For each offset variable index, it extracts all non-zero weight variable indices pointing to the corresponding variable from the causal weight matrix to construct the initial causal path set.

[0051] The path backtracking analysis unit is used to perform reverse traversal on the initial causal path set, identify the predecessor variable index on the causal chain, filter out path segments with propagation strength higher than a set threshold by combining causal weight values, and mark variable indexes that appear frequently in the path as key causal nodes.

[0052] The link structure generation unit is used to connect the offset variable index with its corresponding key causal node, construct a set of directed paths with key causal nodes as the source and offset variables as the target based on the causal path direction, summarize them to form a complete causal variable path graph, and finally output the fault link structure.

[0053] The beneficial effects of this invention are:

[0054] (1) This invention introduces a subspace tracking algorithm to construct a dynamically evolving operating state feature structure, which can realize real-time updates of the feature subspace in the scenario of multiple operating conditions and high-frequency data of power plant equipment, significantly improve the system's sensitivity to state drift and modeling accuracy, and break through the limitation of traditional static models that cannot adapt to changes in operating conditions.

[0055] (2) Based on Granger causality analysis, this invention identifies the temporal causal structure between variables and embeds causal weights into the subspace modeling and updating process to realize the linkage modeling of dynamic dependencies between variables, forming an interpretable causal-driven updating mechanism, and enhancing the accuracy and diagnostic capability of subspace representation.

[0056] (3) By constructing a sequence of subspace angle changes and combining it with time continuity analysis, this invention can accurately identify the state deviation period in the operation of the power plant, effectively support early anomaly detection and operation status trend judgment, and improve the system's response speed in the process of anomaly evolution.

[0057] (4) This invention combines the causal weight matrix and the state offset variable index to construct the causal variable path graph in reverse, output the fault link structure, realize the structured identification of the fault source and propagation path, break through the black box limitation of traditional anomaly detection that cannot trace the root cause, and improve the system's decision support capability and intelligent diagnosis depth. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a framework diagram of a power plant automated operation and maintenance management system based on big data analysis proposed in this invention;

[0060] Figure 2 This is a schematic diagram of the modeling process of the subspace construction module based on the subspace tracking algorithm of the present invention;

[0061] Figure 3 This is a schematic diagram of the structure of the subspace fusion module of the present invention, which embeds causal weights into the subspace update process. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figures 1-3 An automated operation and maintenance management system for power plants based on big data analytics includes the following steps:

[0064] The data acquisition module is used to collect operating parameters of power plant equipment, environmental monitoring data and operation and maintenance records, and generate multi-source raw data sequences with a unified time identifier;

[0065] The data preprocessing module is used to receive multi-source raw data sequences, perform preprocessing, and output standardized time series data tensors.

[0066] The subspace construction module receives standardized time series data tensors, constructs the initial subspace structure of the running state through the subspace tracking algorithm, and iteratively updates it within a continuous time window to generate a subspace structure sequence.

[0067] The causal relationship identification module is used to receive standardized time series data tensors, perform Granger causal analysis, construct a causal graph structure between time series variables, and output a causal weight matrix.

[0068] The subspace fusion module is used to receive the subspace structure sequence and the causal weight matrix, embed the causal weights into the subspace update process, and generate a dynamic subspace structure sequence under causal constraints.

[0069] The state recognition module is used to receive the dynamic subspace structure sequence, calculate the subspace angle change between consecutive time points, and determine the state offset time period based on the subspace angle change.

[0070] The link construction module receives the state offset time period and the causal weight matrix, combines the variable indices that have shifted in the subspace structure sequence, establishes a causal variable path graph related to the state offset, and outputs the fault link structure.

[0071] In this embodiment, the data acquisition module includes:

[0072] The operating parameter acquisition unit is used to acquire operating parameters from the power plant equipment's operation monitoring terminal. The operating parameters include equipment speed, voltage, current, power, frequency, and equipment operating status indicators.

[0073] The environmental monitoring data acquisition unit is used to acquire environmental monitoring data from environmental monitoring sensors. The environmental monitoring data includes ambient temperature, humidity, emission monitoring signals and external air pressure measurements.

[0074] The operation and maintenance record collection unit is used to obtain operation and maintenance records related to the equipment operation cycle from the power plant operation and maintenance information system. The operation and maintenance records include maintenance time markers, fault record entries, maintenance task information and spare parts registration information.

[0075] All types of data output from the operation parameter acquisition unit, environmental monitoring acquisition unit, and operation and maintenance record acquisition unit are accompanied by a unified time identifier and are collected in a structured form as a multi-source raw data sequence.

[0076] In this embodiment, preprocessing includes: data cleaning, missing data imputation, time alignment, and format standardization;

[0077] The data preprocessing module receives multi-source raw data sequences and sequentially performs data cleaning, missing data imputation, time alignment, and format standardization. Data cleaning removes abnormal data that does not conform to the numerical range by setting operating parameters and upper and lower limit threshold rules for environmental monitoring fields. Missing data imputation uses linear interpolation or moving average methods based on field type to fill in missing data points within the time window. Time alignment constructs a target time axis based on a unified time step and resamples various types of data to this time axis using nearest neighbor or linear interpolation. Format standardization converts all fields to a unified data type and a unified unit of measurement and rearranges the field order, finally outputting a standardized time series data tensor that meets the time series structure constraints.

[0078] In this embodiment, the subspace construction module includes:

[0079] The feature vector generation unit is used to extract multiple fields related to the operating status of power plant equipment from the standardized time series data tensor, construct a continuous feature vector set in time order, and output a feature matrix for subspace modeling.

[0080] An initial subspace modeling unit is used to receive the feature matrix, extract the first k principal component vectors through principal component analysis, and perform orthogonalization on the principal component vectors to construct the initial subspace structure, wherein the value of k is determined according to the cumulative variance contribution rate. The initial subspace structure is used as the initialization input for the subspace tracking algorithm.

[0081] The orthogonalization of principal component vectors refers to the process of extracting the first k principal component vectors and performing an orthogonalization operation on them to ensure that the subspace formed by these vectors is linearly independent. The specific implementation method is as follows: the principal component vectors are sorted according to the size of their eigenvalues ​​and introduced sequentially. The Gram-Schmidt orthogonalization method is used to gradually remove the projection components of each newly introduced vector on the existing orthogonal basis, retaining only the parts orthogonal to the previous vectors, thereby constructing a set of pairwise orthogonal and normalized principal component orthogonal bases.

[0082] The subspace trend prediction unit is used to receive the initial subspace structure and the subspace structure generated at each time step iteration, construct the subspace structure change sequence, and perform vector direction fitting based on the subspace structure change sequence to generate the predicted subspace structure.

[0083] The subspace structure generated by each time step iteration refers to the subspace basis matrix generated within a continuous time window by receiving the feature vector input of each time step and performing an update operation by combining the subspace structure of the previous time step, the signal residual, and the prediction structure. This matrix represents the main feature direction of the current running state.

[0084] Vector direction fitting refers to using the least squares method to perform linear regression on the trajectory of the subspace basis vectors of the same dimension based on the subspace basis vectors under multiple consecutive time steps, calculating the trend of their direction change, and generating the predicted vector direction for the next time step.

[0085] The projection residual calculation unit is used to receive the feature vector set of the current time step and the subspace structure of the previous time step, project the feature vector set onto the previous subspace structure and calculate the residual vector, which is used to measure the difference between the current input and the subspace representation.

[0086] The specific implementation of projecting the feature vector set onto the previous subspace structure and calculating the residual vector is as follows: At each time step, the current feature vector is used as input, and the orthogonal matrix formed by the subspace basis generated in the previous time step is used to perform a projection operation on the feature vector, that is, the projection of the vector in each direction of the subspace basis is summed to obtain its representation vector in the subspace. The difference between the original feature vector and its projection vector in the subspace is the residual vector. The residual reflects the part that cannot be explained between the current input and the existing subspace structure.

[0087] The noise stripping unit receives the residual vector, performs energy spectrum analysis and threshold decomposition, divides the residual vector into signal residual and noise residual, and outputs only the signal residual to the subspace update unit as the basis for effective update.

[0088] The specific implementation of energy spectrum analysis and threshold decomposition is as follows: A fast Fourier transform is performed on the residual vector to obtain the energy spectrum distribution in the frequency domain. The energy intensity of each frequency component is calculated, and the overall energy distribution characteristics are statistically analyzed. An adaptive energy threshold based on historical residual statistical characteristics is set to identify high-frequency energy concentration areas. Frequency components with energy above the threshold are identified as noise components, and their corresponding time-domain components are reconstructed into noise residuals through inverse transform. The remaining frequency components are retained as signal residuals. Finally, the residual vector is decomposed into signal and noise components.

[0089] The subspace update unit is used to receive the previous subspace structure, the current signal residual and the predicted subspace structure, perform QR decomposition driven by the signal residual to obtain the candidate updated subspace structure, and then weight and fuse the candidate updated subspace structure and the predicted subspace structure according to the set weight to output the subspace structure of the current time step.

[0090] The specific implementation method for obtaining the candidate update subspace structure by performing QR decomposition is as follows: at each time step, the subspace basis matrix of the previous time step is concatenated with the signal residual vector of the current time step to form an extended matrix; QR decomposition operation is performed on the extended matrix to obtain an orthogonal matrix Q and an upper triangular matrix R; wherein, the first k columns of matrix Q are used as new orthogonal basis to form the candidate update subspace structure, which is used to characterize the main feature direction of the device operating state at the current time step;

[0091] The specific implementation method of weighted fusion according to the set weights is as follows: the candidate updated subspace structure and the predicted subspace structure obtained by QR decomposition are linearly weighted according to the corresponding positions of the column vectors. The fusion formula is: the subspace structure is equal to the candidate updated subspace structure multiplied by the weight coefficient α, plus the predicted subspace structure multiplied by the weight coefficient 1−α, where α is a preset or adaptively adjusted fusion parameter.

[0092] The sequence generation unit and the subspace update unit perform update operations multiple times within a continuous time window to form a subspace structure sequence that changes over time.

[0093] The specific implementation method for performing multiple update operations is as follows: Within a set continuous time window, the system receives new feature vector inputs sequentially at each time step and calls the subspace structure generated at the previous time step as the update basis; it sequentially performs projection residual calculation, noise stripping, subspace update and prediction fusion processing, and uses the subspace structure after each update as the input basis for the next time step, iterating cyclically until the sliding window ends; each iteration outputs the subspace structure corresponding to the current time step and records it in the subspace structure sequence to reflect the changing trend of the device's operating status on the time axis.

[0094] In this embodiment, the causal relationship identification module includes:

[0095] The lag order determination unit is used to receive the standardized time series data tensor, construct a candidate set of lag orders for any two time series variables in the standardized time series data tensor, calculate the evaluation value corresponding to each lag order based on the preset Bayesian information criterion, and select the lag order with the smallest evaluation value as the optimal lag order of the time series variable pair composed of any two time series variables.

[0096] The specific implementation of the pre-defined Bayesian information criterion in this system is as follows: For any two time series variables in the standardized time series data tensor, regression models with different lag orders are constructed. The Bayesian information criterion (BIC) value corresponding to each order is calculated, and this value is used as a trade-off index between model complexity and goodness of fit. Among all candidate orders, the lag order with the smallest BIC value is selected as the input parameter for subsequent causal relationship modeling. This is used to construct restricted and unrestricted regression models, thereby achieving pre-optimization of the causal determination process between time series variables.

[0097] The restricted regression modeling unit is used to receive the optimal lag order, construct a restricted regression model for the currently analyzed time series variable based on the optimal lag order, which only contains the lagged terms of the corresponding time series variable itself, and calculate the residual sum of squares of the restricted regression model.

[0098] A restricted regression model is a univariate autoregressive model that uses only the lagged values ​​of the time series variable itself as independent variables. It extracts the values ​​of the target variable in several lagged periods before the current time point from the standardized time series data tensor and uses them as regression inputs to build a linear regression model to predict the variable value at the current time. This model does not contain information from any other variables.

[0099] The unrestricted regression modeling unit is used to construct an unrestricted regression model for the currently analyzed time series variable based on the optimal lag order, which includes the lag term of the corresponding time series variable itself and the lag term of another time series variable, and to calculate the sum of squared residuals of the unrestricted regression model.

[0100] An unrestricted regression model refers to a model that, when constructing a predictive model for the current value of a target time series variable, uses not only the lagged value of the time series variable itself as an independent variable, but also introduces the lagged value of another time series variable under test within the same lag period as an additional independent variable. The model extracts the historical values ​​of the target variable and the external variable within consecutive lag periods from the standardized time series data tensor, combines them to form a multidimensional independent variable vector, and constructs a multiple linear regression model based on this vector. By comparing the sum of squared residuals of this model with the sum of squared residuals of a restricted regression model containing only its own lagged terms, the model can determine whether the external variable has a Granger causal relationship with the target variable.

[0101] The significance determination unit is used to receive the residual sum of squares of the restricted regression model and the residual sum of squares of the unrestricted regression model, perform differential statistical analysis based on the preset sample size and the number of model parameters, and determine whether the difference between the two regression models is significant according to the set significance level. When the significance determination result rejects the no-causal hypothesis, it records that there is a Granger causal relationship between the current time series variables.

[0102] The specific implementation method for statistical analysis of differences based on preset sample size and number of model parameters is as follows: After constructing a restricted regression model and an unrestricted regression model, obtain the sum of squared residuals of the two models and calculate the residual difference between the two models; combine the total sample size of the standardized time series data tensor and the number of parameters included in the two models, construct a test statistic based on these statistics to measure whether the model differences are significant; by comparing the statistic with the preset significance level, determine whether to reject the no-causal hypothesis, in order to support the determination that there is a causal relationship between the time series variables;

[0103] The established significance level is used to determine whether the difference in residuals between two regression models is sufficient to reject the hypothesis of no causation. Specifically, based on the sum of squared residuals, sample size, and number of parameters of the restricted and unrestricted regression models, an F-statistic is constructed for hypothesis testing. The preset significance level α is typically set to 0.05 or 0.01. When the p-value corresponding to the calculated F-statistic is less than this significance level, the difference between the two models is considered statistically significant, thus determining that the external variable has a Granger causal effect on the target variable.

[0104] The causal weight matrix generation unit is used to receive all time series variable pairs with Granger causal relationships, construct a two-dimensional matrix based on the variable index, mark the time series variable pairs with causal relationships as non-zero elements, and assign weight values ​​according to the corresponding statistical test indicators to form a causal weight matrix.

[0105] The specific implementation method for constructing a two-dimensional matrix using variable indexes is as follows: all time series variables involved in causal analysis in the standardized time series data tensor are numbered in a fixed order and used as the row index and column index of the matrix; each element in the matrix corresponds to a pair of ordered variable combinations, indicating whether the row variable has a causal influence on the column variable, and the element value is generated by the causal determination result or causal weight to form a complete causal relationship structure;

[0106] The specific implementation method for assigning weight values ​​to the corresponding statistical test indicators is as follows: For variable pairs with Granger causal relationship, extract the absolute value of the F statistic or regression coefficient calculated in the significance test as the causal effect strength index, and directly assign this value to the corresponding element in the causal weight matrix as the directed edge weight for subsequent causal graph structure construction.

[0107] The causal graph structure construction unit is used to receive the causal weight matrix, map each time series variable in the causal weight matrix to a node in the causal graph structure, map the time series variable pairs corresponding to non-zero elements to directed edges, the direction of the directed edges is from the dependent variable to the affected variable, the edge weight of the directed edges is generated using the corresponding values ​​in the causal weight matrix, and output the causal graph structure between time series variables.

[0108] In this embodiment, the subspace fusion module includes:

[0109] The feature dimension correspondence unit is used to map each column of feature vectors in the subspace structure of the current time step to the time series variables in the standardized time series data tensor, and establish the correspondence between feature vectors and variable indices.

[0110] The causal weight extraction unit is used to extract the causal influence factors corresponding to each feature vector from the causal weight matrix based on the correspondence between feature vectors and variable indices, forming a set of causal weight factors consistent with the number of column vectors in the subspace structure.

[0111] The causal weighted adjustment unit is used to receive the candidate updated subspace structure and the predicted subspace structure during the candidate updated subspace structure generation process, and to perform weighted transformation on each column feature vector of the candidate updated subspace structure according to its corresponding causal influence factor to form an intermediate subspace structure with embedded causal adjustment.

[0112] A weighted transformation is performed on each column of feature vectors in the candidate updated subspace structure. The vectors are scaled according to their corresponding causal influence factors as weight coefficients. That is, the original vectors are multiplied by the causal weight values ​​to adjust their proportion in the fusion process, thereby realizing the quantitative guidance of causal relationships in the subspace structure update.

[0113] The fusion update unit is used to linearly weight the intermediate subspace structure with embedded causal adjustment and the prediction subspace structure according to a preset fusion ratio to generate the fusion update subspace structure of the current time step.

[0114] After obtaining the intermediate subspace structure and the prediction subspace structure with embedded causal adjustment by linear weighting with a preset fusion ratio, the two are multiplied by the preset fusion ratio coefficient, and then the corresponding column vectors are summed element by element to form the fusion update subspace structure of the current time step, so that the two input structures jointly determine the final subspace representation according to the set ratio.

[0115] The structure sequence generation unit is used to take the fusion update subspace structure of the current time step as the final subspace structure of the corresponding time step, and iterate and update it according to the continuous time steps. The final subspace structures under all time steps are stored in sequence to generate a dynamic subspace structure sequence.

[0116] During system operation, for each time step, the final subspace structure generated by the fusion update of the current time step is obtained and stored in the time series data structure as the state representation vector of that time step. The data structure is indexed in time order and supports the sequential appending of records by sliding window over time until the window ends. Finally, an ordered sequence containing the subspace structures of all time steps is formed, which serves as the dynamic subspace structure sequence.

[0117] In this embodiment, the status recognition module includes:

[0118] The subspace structure pairing unit is used to receive the dynamic subspace structure sequence and extract the subspace structure of any two adjacent time steps to construct the subspace structure pair sequence.

[0119] Angle change extraction unit is used to perform vector space similarity analysis on each pair of subspace structures, calculate the angle change between subspaces based on the directional differences between corresponding orthogonal feature vectors, and bind the angle change to the time step index;

[0120] The specific implementation of vector space similarity analysis is as follows: For the subspace structure of two adjacent time steps, extract their orthogonal feature vector sets respectively, and calculate the cosine value of the angle between the corresponding vector sets to measure their directional consistency; the cosine value reflects the degree of overlap between the two subspaces in the feature dimension, and the smaller the value, the greater the difference.

[0121] The directional difference between orthogonal eigenvectors refers to the angle between the vector directions corresponding to the same feature dimension in the subspace representation at adjacent time steps in a high-dimensional vector space. By comparing the directional changes of corresponding column vectors in adjacent subspaces, the evolution magnitude and trend of the subspace structure in the time dimension can be reflected.

[0122] Angle change sequence construction unit is used to arrange angle changes in chronological order to construct a subspace angle change sequence, which reflects the temporal evolution characteristics of the dynamic subspace structure.

[0123] An anomaly marker generation unit is used to receive the angle change sequence and mark the time steps that are greater than a preset angle threshold as an anomaly time step.

[0124] The preset angle threshold is determined by statistically analyzing the historical subspace angle change data, selecting the high quantile (such as the 90th percentile) in the change distribution as the identification boundary for abnormal changes, and fine-tuning the angle threshold according to the system operating characteristics to finally determine a fixed reference standard for identifying state offsets.

[0125] The offset segment identification unit is used to perform continuity analysis on abnormal time steps, identify continuous subsequences with an interval of less than or equal to one time step between abnormal time steps, and delineate the corresponding time segments as candidate state offset segments.

[0126] The specific implementation method for continuity analysis is as follows: sort all time step indices marked as anomalous, compare the differences between adjacent index values ​​one by one, and when the index difference between two adjacent time steps is not greater than 1, they are considered to belong to the same continuous anomalous segment. In this way, the anomalous time steps are divided into several continuous segments.

[0127] The offset time period output unit is used to perform minimum duration verification on the candidate state offset segment and output the time segment that meets the duration requirement as the state offset time period.

[0128] The specific implementation method for minimum duration verification is as follows: For each identified candidate state offset segment, count its time step count, and compare the duration with the preset minimum duration threshold. Only segments with a duration not less than the threshold are retained as valid state offset periods, and the remaining segments are discarded.

[0129] In this embodiment, the link construction module includes:

[0130] The offset variable identification unit is used to receive the dynamic subspace structure sequence and the state offset time period, extract the subspace structure corresponding to each time step in the state offset time period, perform directional change analysis on each column feature vector, and when the directional difference of the same column feature vector in adjacent time steps exceeds the preset offset judgment threshold, record the variable index corresponding to the corresponding column and summarize it to form an offset variable index set.

[0131] Directional change analysis is performed on the subspace structural feature vectors of the same column in continuous time steps. The angle difference between their normalized vectors is calculated, and the structural drift magnitude of the variable is judged based on the degree of directional change between the vectors. When the angle is greater than a set threshold, it is determined that the variable has changed direction in the current time step.

[0132] The preset offset judgment threshold is calculated based on the statistical distribution of the subspace angle change of the dynamic subspace structure sequence at each time step during the historical normal operation period. The high percentile value, such as the 95th percentile value, is then set as the offset judgment threshold for subspace direction change.

[0133] The causal path extraction unit is used to receive the offset variable index set and the causal weight matrix. For each offset variable index, it extracts all non-zero weight variable indices pointing to the corresponding variable from the causal weight matrix to construct the initial causal path set.

[0134] The path backtracking analysis unit is used to perform reverse traversal on the initial causal path set, identify the predecessor variable index on the causal chain, filter out path segments with propagation strength higher than a set threshold by combining causal weight values, and mark variable indexes that appear frequently in the path as key causal nodes.

[0135] The reverse traversal starts from the offset variable index and searches for all predecessor variables pointing to the variable in the causal graph structure constructed by the causal weight matrix. It then recursively traces its predecessor node upwards until there are no upper-level nodes or the preset traversal depth limit is reached. During the process, all variable indices and path structures are recorded.

[0136] The causal weight value filtering process extracts the causal weight value corresponding to the edge weight of each path from all causal paths obtained by reverse traversal, compares it with the preset propagation strength threshold, retains the path segments with causal weight greater than the threshold, removes low-weight paths, and outputs only the causal influence paths with significant propagation strength.

[0137] The link structure generation unit is used to connect the offset variable index with its corresponding key causal node, construct a set of directed paths with key causal nodes as the source and offset variables as the target based on the causal path direction, summarize them to form a complete causal variable path graph, and finally output the fault link structure.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to a large-scale thermal power plant. With its massive equipment and complex operating environment, continuous monitoring and timely early warning of equipment health status have become crucial for ensuring stable power generation and safe operation. Traditional monitoring systems often employ univariate methods based on threshold settings, lacking the ability to characterize the evolution trend of equipment operating status and analyze the causal linkage mechanism between multi-source data, thus failing to meet the real-time and accuracy requirements under complex fault modes.

[0140] This embodiment demonstrates the formal deployment and operation of a power plant automated operation and maintenance management system based on big data analytics at the power plant. This system is used to collect, model, analyze, and identify potential anomalies and fault propagation links during equipment operation. The entire system consists of a data acquisition module, a data preprocessing module, a subspace construction module, a causal relationship identification module, a subspace fusion module, a state identification module, and a link construction module. Through end-to-end processing and intelligent modeling, it effectively replaces traditional static rule-based methods.

[0141] During the data acquisition phase, the system deploys multiple acquisition ports to collect equipment operating parameters (such as current, voltage, and power) from components including the main steam pump, desulfurization fan, and cooling tower water pump; environmental data (such as humidity, flue gas flow rate, and outdoor temperature); and maintenance record data (such as maintenance cycle, maintenance tasks, and equipment anomaly logs). The acquisition frequency is once every 5 seconds, and a timestamp is uniformly added to form a structured data stream.

[0142] In the data preprocessing stage, the collected raw data is cleaned and missing values ​​are filled in. Linear interpolation is used to repair univariate null values. At the same time, all time series are aligned to a unified time window and tensor data structures are constructed through Z-Score standardization.

[0143] The subspace construction module employs an improved subspace tracking algorithm. It utilizes principal component analysis to extract the first five principal components within each time period to construct the initial subspace structure, and dynamically updates the subspace representation in consecutive time steps using eigenvector orthogonalization, direction fitting, and noise stripping mechanisms. Each update generates a new subspace for the current time step through QR decomposition and fusion mechanisms, and subsequently links with the causal module.

[0144] The causal relationship identification module performs Granger causal analysis on all time series variable pairs, determines the optimal lag order, and constructs restricted and unrestricted regression models respectively. The model quality is evaluated based on the Bayesian information criterion, and the significance is tested by the difference of residual sum of squares. Finally, a causal weight matrix containing the causal strength of variables is formed, and a causal graph structure of time series variables is constructed.

[0145] During the subspace fusion stage, the system embeds the aforementioned causal weights into the subspace structure update process, extracts corresponding causal factors based on the index mapping of features and variables, adjusts the principal component direction by weighting, and then linearly weights it with the prediction direction by a fixed fusion ratio to form a dynamically evolving subspace structure sequence.

[0146] The state recognition module performs continuous time step pairing analysis on the sequence, forms an angle sequence by calculating the angle change between orthogonal subspaces, marks abnormal time steps with a preset threshold, and then determines the final state offset period by the minimum duration determination method.

[0147] The link construction module analyzes the triggering variables from the state offset period, extracts the corresponding variable index in the subspace structure, judges the offset variable by the direction change, and then combines the cause-effect graph structure to backtrack the high causal weight path and output one or more complete fault propagation links.

[0148] In the case of a sudden vibration failure of the desulfurization fan No. A at the power plant, this system detected a state deviation 15 minutes before the failure occurred, identified a significant causal relationship between the change in the fan input current and the fluctuation in the cooling water pump frequency, and constructed the following fault path diagram:

[0149] Table 1 Fault Link Paths and Key Indicators

[0150] Time step initial variables Target variable Subspace angle change Causal weight t-4 Water pump operating frequency Main fan current 12.4° 0.78 t-3 Main fan current Main fan bearing temperature 14.2° .82 t-2 Main fan bearing temperature Fan vibration signal 19.7° 0.88

[0151] By identifying the aforementioned links, the system provided early warning of potential faults in wind turbine A and prompted relevant maintenance personnel to perform shutdown inspections. On-site maintenance records confirmed that reduced bearing lubricating oil flow led to abnormal temperature rise, which in turn induced abnormal vibrations in the mechanical structure, perfectly matching the system's automatic identification results.

[0152] Table 2 Comparison of Fault Event Recognition Performance

[0153] method Advance identification time (minutes) False alarm rate (%) Diagnostic accuracy (%) Method of the present invention 15 3.2 94.6 Traditional threshold detection method 0 12.7 67.4 Static PCA model 5 7.5 82.1

[0154] Analysis of the table data reveals that the power plant automation operation and maintenance management system of this invention has significant advantages in fault event identification. Table 1 clearly shows the progressive propagation process from pump frequency fluctuations to abnormal fan vibration. The subspace angle change at each time step exceeds 12°, reaching 19.7° at t-2, indicating a drastic shift in the system state at that point. Simultaneously, the corresponding causal weight values ​​are all greater than 0.75, reflecting a strong dynamic causal relationship between variables, verifying the system's high sensitivity and high-confidence diagnostic capability for abnormal behavior.

[0155] The comparative data in Table 2 further highlights the system's performance advantages. The method of this invention can identify faults 15 minutes in advance, demonstrating a significant time advantage compared to the traditional threshold method (0 minutes) and the static PCA method (5 minutes). The false alarm rate is only 3.2%, far lower than the traditional threshold method (12.7%) and the static PCA method (7.5%), indicating superior noise tolerance and stability. The diagnostic accuracy also reaches 94.6%, more than 10% higher than traditional methods, effectively reducing the false positive and false negative rates.

[0156] In summary, from the interpretability of fault links and the ability to diagnose in advance to the accuracy of prediction, this system demonstrates significant and superior practical effects in the complex operating conditions of actual power plants.

[0157] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A power plant automatic operation and maintenance management system based on big data analysis, characterized in that, The method comprises the following steps: A data acquisition module is used to collect the operating parameters, environmental monitoring data and operation and maintenance records of the power plant equipment, and generate a multi-source original data sequence with a unified time identifier; A data preprocessing module is used to receive the multi-source original data sequence, perform preprocessing, and output a standardized time series data tensor; A subspace construction module is used to receive the standardized time series data tensor, construct an initial subspace structure of the operating state by using a subspace tracking algorithm, and iteratively update the initial subspace structure within a continuous time window to generate a subspace structure sequence; A causal relationship identification module is used to receive the standardized time series data tensor, perform Granger causality analysis, construct a causal graph structure between time series variables, and output a causal weight matrix; A subspace fusion module is used to receive the subspace structure sequence and the causal weight matrix, embed the causal weight into the subspace update process, and generate a dynamic subspace structure sequence; A state identification module is used to receive the dynamic subspace structure sequence, calculate the subspace angle change amount between continuous time points, and determine the state offset period based on the subspace angle change amount; A link construction module is used to receive the state offset period and the causal weight matrix, combine the variable index that has occurred offset in the subspace structure sequence, establish a causal variable path graph related to the state offset, and output a fault link structure; The subspace construction module comprises: A feature vector generation unit is used to extract a plurality of fields related to the operating state of the power plant equipment from the standardized time series data tensor, construct a continuous feature vector set in time sequence, and output a feature matrix; An initial subspace modeling unit is used to receive the feature matrix, extract the first k principal component vectors by principal component analysis, and construct an initial subspace structure by orthogonalizing the principal component vectors; A subspace trend prediction unit is used to receive the initial subspace structure and the subspace structures generated at each time step, construct a subspace structure change sequence, and generate a predicted subspace structure based on the subspace structure change sequence; A projection residual calculation unit is used to receive the feature vector set at the current time step and the subspace structure at the previous time step, project the feature vector set to the previous subspace structure and calculate a residual vector; A noise stripping unit is used to receive the residual vector, perform energy spectrum analysis and threshold decomposition, divide the residual vector into signal residual and noise residual, and output only the signal residual to the subspace update unit as an effective update basis; A subspace update unit is used to receive the previous subspace structure, the current signal residual and the predicted subspace structure, obtain a candidate updated subspace structure by QR decomposition driven by the signal residual, and output the subspace structure at the current time step by weighting and fusing the candidate updated subspace structure and the predicted subspace structure according to a set weight; A sequence generation unit is used to execute the update operation multiple times within a continuous time window to form a subspace structure sequence that changes over time.

2. The power plant automation operation and maintenance management system based on big data analysis according to claim 1, characterized in that, The data acquisition module comprises: An operating parameter acquisition unit is used to collect operating parameters from the operating monitoring end of the power plant equipment, and the operating parameters include equipment speed, voltage, current, power, frequency and equipment operating state indicators; An environmental monitoring collection unit is configured to collect environmental monitoring data from an environmental monitoring sensor device, the environmental monitoring data including ambient temperature, humidity, emission monitoring signals, and external air pressure measurement values; An operation and maintenance record collection unit is configured to obtain operation and maintenance records related to the operation cycle of the equipment from a power plant operation and maintenance information system, the operation and maintenance records including maintenance time markers, fault record entries, maintenance task information, and spare part registration information; The various types of data output by the operation parameter collection unit, the environmental monitoring collection unit, and the operation and maintenance record collection unit are all attached to a unified time identifier, and are collected in a structured form as a multi-source original data sequence.

3. The power plant automation operation and maintenance management system based on big data analysis according to claim 1, characterized in that, The preprocessing includes: data cleaning, missing data filling, time alignment, and format standardization.

4. The power plant automation operation and maintenance management system based on big data analysis according to claim 1, characterized in that, The causal relationship identification module includes: A lag order determination unit is configured to receive the standardized time series data tensor, construct a lag order candidate set for any two time series variables in the standardized time series data tensor, calculate an evaluation value corresponding to each lag order based on a preset Bayesian information criterion, and select the lag order with the smallest evaluation value as the optimal lag order for the pair of time series variables. A restricted regression modeling unit is configured to receive the optimal lag order, construct a restricted regression model containing only the lag items of the corresponding time series variable itself for the currently analyzed time series variable based on the optimal lag order, and calculate the residual sum of squares of the restricted regression model. An unrestricted regression modeling unit is configured to construct an unrestricted regression model containing lag items of the corresponding time series variable itself and lag items of another time series variable for the currently analyzed time series variable based on the optimal lag order, and calculate the residual sum of squares of the unrestricted regression model. A significance determination unit is configured to receive the residual sum of squares of the restricted regression model and the residual sum of squares of the unrestricted regression model, perform difference statistical analysis based on a preset sample size and the number of model parameters, and determine whether the difference between the two regression models is significant according to a set significance level. When the significance determination result rejects the no-causality hypothesis, it is recorded that the current pair of time series variables has a Granger causal relationship. A causal weight matrix generation unit is configured to receive all pairs of time series variables that have a Granger causal relationship, construct a two-dimensional matrix according to the variable index, mark the pairs of time series variables that have a causal relationship as non-zero elements, assign weight values according to the corresponding statistical test indicators, and form a causal weight matrix. A causal graph structure construction unit is configured to receive the causal weight matrix, map each time series variable in the causal weight matrix to a node in the causal graph structure, map the pairs of time series variables corresponding to the non-zero elements to directed edges, and output the causal graph structure between the time series variables.

5. The power plant automation operation and maintenance management system based on big data analysis according to claim 1, characterized in that, The subspace fusion module includes: A feature dimension correspondence unit is configured to one-to-one correspond each column feature vector in the subspace structure at the current time step to a time series variable in the standardized time series data tensor, and establish a correspondence between the feature vectors and the variable indices. A causal weight extraction unit is configured to extract causal influence factors corresponding to each feature vector from the causal weight matrix based on the correspondence between the feature vectors and the variable indices, and form a set of causal weight factors consistent with the number of column vectors in the subspace structure. a causal weighting adjustment unit, configured to perform, in a candidate update subspace structure generation process, weighting transformation on each column of feature vectors in the candidate update subspace structure according to a corresponding causal influence factor of the column, to form an intermediate subspace structure embedded with causal adjustment; a fusion update unit, configured to perform linear weighting on the intermediate subspace structure embedded with causal adjustment and the prediction subspace structure according to a preset fusion ratio, to generate a fusion update subspace structure at a current time step; a structure sequence generation unit, configured to take the fusion update subspace structure at the current time step as a final subspace structure at a corresponding time step, to perform iterative update according to consecutive time steps, to sequentially store the final subspace structures at all time steps, and to generate a dynamic subspace structure sequence.

6. The power plant automation operation and maintenance management system based on big data analysis according to claim 1, characterized in that, The state identification module includes: a subspace structure pairing unit, configured to receive the dynamic subspace structure sequence, and extract subspace structures at any two adjacent time steps from the dynamic subspace structure sequence to construct a subspace structure pair sequence; an angle change extraction unit, configured to perform vector space similarity analysis on each pair of subspace structures, to calculate an angle change amount between the subspace structures according to a direction difference between corresponding orthogonal feature vectors, and to bind the angle change amount with a time step index; an angle change sequence construction unit, configured to arrange the angle change amounts in time sequence to construct a subspace angle change amount sequence; an abnormality marker generation unit, configured to receive the angle change amount sequence, and to mark a time step in the angle change amount sequence as an abnormal time step if an angle change amount of the time step is greater than a preset angle threshold; an offset section identification unit, configured to perform continuity analysis on the abnormal time steps, to identify a continuous subsequence between the abnormal time steps, the continuous subsequence having a time step interval less than or equal to one, and to demarcate a corresponding time section as a candidate state offset segment; an offset period output unit, configured to perform minimum duration verification on the candidate state offset segment, and to output a time segment that meets a duration requirement as a state offset period.

7. The power plant automation operation and maintenance management system based on big data analysis of claim 1, wherein, The link construction module includes: an offset variable identification unit, configured to receive the dynamic subspace structure sequence and the state offset period, to extract subspace structures corresponding to time steps in the state offset period, to perform direction change analysis on feature vectors in each column, and to record a variable index corresponding to a column when a direction difference of the column between adjacent time steps exceeds a preset offset determination threshold, to aggregate the variable indexes to form an offset variable index set; a causal path extraction unit, configured to receive the offset variable index set and a causal weight matrix, to extract, for each offset variable index, all non-zero weight variable indexes pointing to the corresponding variable from the causal weight matrix, and to construct an initial causal path set; a path backtracking analysis unit, configured to perform reverse traversal on the initial causal path set, to identify predecessor variable indexes on a causal chain, to filter out path segments with a propagation strength higher than a set threshold according to causal weight values, and to mark variable indexes appearing frequently in the path as key causal nodes. The link structure generating unit is configured to connect the offset variable index with the corresponding key causal node, construct a set of directed paths with the key causal node as a source and the offset variable as a target based on the causal path direction, aggregate to form a complete causal variable path graph, and finally output a fault link structure.

Images