A sensor anomaly diagnosis method and device for a power equipment monitoring system

Abstract

This invention belongs to the field of abnormal data detection technology and discloses a method and device for diagnosing sensor anomalies in a power equipment monitoring system. The method includes: collecting continuous monitoring variable data of the sensor and preprocessing it, selecting the first N data as a reference data matrix; selecting network nodes of the target sensor based on the magnitude of the regularization kernel typical nonlinear correlation between the target sensor and other sensors; calculating the regularization kernel typical nonlinear correlation and causal directionality of the time lag windows between the target sensor and the associated nodes of the target sensor, thereby constructing a directed causal difference network; evaluating the variation of each node in the directed causal difference network in the current time window based on the information transmission flow entropy, and calculating the comprehensive state index (SAMS) of the target sensor for each time window; when the SAMS of the target sensor in a certain time window is higher than a threshold, the target sensor is considered to have an anomaly.

Description

Technical Field

[0001] This invention relates to the field of abnormal data detection technology, and in particular to a method and apparatus for diagnosing sensor anomalies in a power equipment monitoring system. Background Technology

[0002] In recent years, big data and artificial intelligence technologies such as deep learning have been regarded as the core driving force for building new power systems in the power industry. From load forecasting to intelligent equipment status diagnosis and fault early warning, and then to autonomous decision-making in dispatching and operation, artificial intelligence is gradually realizing the transformation from theory to practice. These algorithms do not rely on traditional physical models or human experience, but make decisions based on massive historical and real-time operational data. However, the performance of the algorithms is highly dependent on the quality of the input data, and the accuracy, completeness, and consistency of the data directly determine the reliability and security of intelligent applications. As the source of industrial data acquisition, the reliability of sensor data is particularly critical. Therefore, in modern power systems that are highly dependent on data-driven approaches, ensuring the accuracy of sensor data has become a core link in ensuring the reliable operation of the system. Sensor data-based fault diagnosis methods for sensors themselves have also received extensive and in-depth research from scholars at home and abroad. Currently, most research on sensor faults at home and abroad can be roughly divided into two major directions: one is the early sensor fault diagnosis method, which is based on statistics and signal processing for sensor fault diagnosis; the other is the data-driven sensor fault diagnosis method derived from large artificial intelligence models. Sensor fault diagnosis methods based on signal processing usually rely on the modeling and analysis of signal characteristics, but there are still obvious shortcomings in practical applications. This approach, often based on predefined mathematical models or statistical assumptions, has limited adaptability to nonlinearity, time-varying characteristics, and complex operating conditions. It primarily targets typical fault types such as offset and jamming, showing low sensitivity to complex, gradual, or unknown faults. On the other hand, while data-driven methods can learn complex fault modes from historical data and reduce reliance on precise physical models, their diagnostic performance is highly dependent on the scale, quality, and operating condition coverage of the training data. In practical applications, they are prone to declining generalization ability due to data imbalance or insufficient representativeness. Furthermore, the decision-making process of these methods often lacks interpretability, making it difficult to provide intuitive fault mechanism analysis. Real-time deployment also requires significant computational resources, limiting their application potential in edge devices or low-power scenarios. Overall, both types of methods face challenges in reliability, applicability, and interpretability when dealing with complex industrial scenarios. Summary of the Invention

[0003] The purpose of this invention is to solve the problems in the prior art.

[0004] The technical solution adopted by this invention to solve its technical problem is: to provide a method for diagnosing sensor anomalies in a power equipment monitoring system, comprising the following steps:

[0005] Data of continuously monitored variables collected by sensors in the power equipment monitoring system were collected. Data from the continuous sampling times of each sensor under the same operating condition of K power equipment units were selected as experimental samples and preprocessed.

[0006] Set the sliding window length, select the first N data points of the preprocessed data as the reference data matrix, and construct the time window data matrix at the current time step by setting the sliding step size according to the same sliding window length.

[0007] The typical nonlinear correlation of the regularization kernel between each sensor is calculated based on the reference data matrix. The target sensor associated node sensor is selected based on the magnitude of the typical nonlinear correlation of the regularization kernel between the target sensor and other sensors. The target sensor and the target sensor associated node sensor constitute the target sensor network node.

[0008] Calculate the typical nonlinear correlation of the regularization kernel between the target sensor and its associated nodes in the time lag window, and determine the direction of the edges between the target sensor network nodes as causal directionality; calculate the typical nonlinear correlation of the regularization kernel between the target sensor and its associated nodes in the current time window and compare it with the typical nonlinear correlation of the regularization kernel in the reference time window to determine the strength of the edges between the target sensor network nodes; construct a directed causal difference network for all time windows based on the direction and strength of the edges between the target sensor network nodes.

[0009] The variation of each node in the directed causal differential network in the current time window is evaluated based on the information transmission flow entropy, and the comprehensive state index (SAMS) of the target sensor for each time window is calculated based on the information transmission entropy flow of the variation.

[0010] A threshold is set, and if the SAMS of the target sensor exceeds the threshold within a certain time window, the target sensor is considered to be abnormal.

[0011] Preferably, the continuous monitoring data from sensors in the power equipment unit is collected, and all data is processed, assuming a total of There are [number] sensors, and the number of time points for each sensor is [number]. , No. The sequence of sensors is denoted as:

[0012] , ;

[0013] in, Indicates sensor In the Observations at a given time point;

[0014] Normalization is represented as:

[0015] ;

[0016] in, For sensors exist The observed value at time, For sensors exist Normalized data of observations at time points; For sensors The average of the observations at all time points; For sensors The standard deviation of the observations at all time points.

[0017] Preferably, the first N data points selected from the preprocessed data are used as a reference data matrix, represented as follows:

[0018] ;

[0019] Among them, the reference data matrix elements in This represents preprocessed data representing observations from different sensors at different time points.

[0020] Setting a time-lag sliding window is represented as follows:

[0021] ;

[0022] Among them, time-lag sliding window elements in This represents the values ​​of different variables from different sensors within the current time window; the last row of data represents the current... Data at any given time, number of columns Represents the number of sensors;

[0023] Set the sliding window size to L and the sliding step size to 1. Combine the time-lag sliding window with the reference sample matrix to form the current time-time window matrix, represented as:

[0024] ;

[0025] in, for The window matrix at time points;

[0026] Repeat the above steps Next, obtain the window matrix that iterates through all time points.

[0027] Preferably, the calculation of the typical nonlinear correlation of the regularization kernel includes the following steps:

[0028] Window sample pair construction, in any time window Inside, take the sensor With sensors The normalized sequences are as follows:

[0029] ;

[0030] ;

[0031] Where L represents the number of sampling points within the window;

[0032] Using Gaussian kernel function to calculate time window sensor Gram matrix With sensors Gram matrix , is represented as:

[0033] ;

[0034] ;

[0035] in, For sensors The value at time a, For sensors The value at time b; For sensors The value at time a, For sensors The value at time b;

[0036] Building sensors kernel Gram matrix With sensors kernel Gram matrix :

[0037] ; ;

[0038] The kernel Gram matrix is ​​centered to obtain the sensor. Centralized kernel Gram matrix With sensors Centralized kernel Gram matrix , respectively represented as:

[0039] ;

[0040] ;

[0041] Where H is the centering matrix, E is The identity matrix, where Q is all 1s and dimension is 1. Matrix;

[0042] To obtain the principal components of the centered kernel Gram matrix, we solve for the eigenvalues ​​of the two centered kernel Gram matrices, which are expressed as follows:

[0043] ;

[0044] ;

[0045] in, and They are respectively eigenvalues ​​and eigenvectors, and They are respectively eigenvalues ​​and eigenvectors;

[0046] Before selecting eigenvalues Using 10 feature vectors as principal components to reduce dimensionality and avoid The numerical instability caused by singularities is represented as:

[0047] ;

[0048] ;

[0049] Where m≤L; and Sensors With sensors Principal components;

[0050] Computational Sensors With sensors The projection scores of time a along the directions of the c principal components are expressed as:

[0051] ;

[0052] ;

[0053] in, Indicates sensor principal components Column c The p-th component, Indicates sensor The value in the a-th row and p-th column of the Gram matrix; Indicates sensor principal components Column c The One component; Indicates sensor The value in the a-th row and p-th column of the Gram matrix;

[0054] Sensors respectively With sensors The m kernel principal component scores at time a are concatenated to obtain an m-dimensional score vector. and , is represented as:

[0055] ;

[0056] ;

[0057] Stack the m-kernel principal component scores at each time step to construct an in-window sensor. Kernel principal component score matrix With sensors Kernel principal component score matrix , is represented as:

[0058] ;

[0059] ;

[0060] Perform CCA in the kernel principal component space to construct the autocovariance matrix and crosscovariance matrix:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] in, and Sensors Autocovariance matrix and sensor The autocovariance matrix, Indicates from The amount to The cross-covariance matrix form of the components, Indicates from The amount to The cross-covariance matrix form of the components;

[0066] To avoid the autocovariance matrix becoming singular or ill-conditioned due to the limited number of samples in the window, a regularization term is introduced, expressed as:

[0067] ;

[0068] ;

[0069] in, Indicates sensor The introduction of the autocovariance matrix of the regularization term... Indicates sensor The introduction of the regularization term introduces the autocovariance matrix; and All are identity matrices. and Both represent regularization coefficients;

[0070] Define the regularized kernel canonical nonlinear correlation function as follows:

[0071] ;

[0072] in, This represents the regularized kernel canonical correlation coefficient. Indicates sensor The corresponding weighted coefficient vector of the kernel principal component scores; Indicates sensor The corresponding weighted coefficient vector of the kernel principal component scores;

[0073] Imposing constraints on the canonical nonlinear correlation function of the regularized kernel, expressed as:

[0074] ;

[0075] ;

[0076] The eigenvalues ​​are solved using the following formula:

[0077] ;

[0078] ;

[0079] Solving yields a set of eigenvalues its square root , , The coefficients are canonical correlation coefficients, where the largest is... Represented as the first kernel canonical regularized correlation coefficient;

[0080] The first kernel canonical regularized correlation coefficient is used as a measure of the sensor. With sensors Indices of typical nonlinear correlation of the regularization kernel within the window. , is represented as:

[0081] ;

[0082] in, ; The larger the value, the stronger the nonlinear correlation between the two sensors in the kernel principal component space;

[0083] Inter-sensor Calculate the correlation coefficient matrix in sequence. , is represented as:

[0084] .

[0085] Preferably, the step of calculating the regularized kernel typical nonlinear correlation of the time lag windows between the target sensor and the sensors associated with the target sensor, and determining the direction of the edges between the target sensor network nodes as causal directionality, includes the following steps:

[0086] Set time lag The largest constant is With sensors As the target sensor, with sensor As the sensor associated with the target sensor, for each hysteresis... Computational Sensors and sensors The regularized kernel's typical nonlinear correlation is expressed as:

[0087] ;

[0088] ;

[0089] in, This represents an index used to calculate the typical nonlinear correlation of the regularization kernel. Indicates sensor For sensors The hysteresis regularization kernel is an index of typical nonlinear correlation. Indicates sensor For sensors The hysteresis regularization kernel is an indicator of typical nonlinear correlation. This represents the kernel principal component score matrix of sensor i at time t. This represents the kernel principal component score matrix of sensor j at time t. This represents the kernel principal component score matrix of sensor i at time (th). Let represent the kernel principal component score matrix of sensor j at time (th);

[0090] The decision quantile is calculated by selecting the index of the canonical nonlinear correlation of the largest hysteresis regularization kernel. , is represented as:

[0091] ;

[0092] ;

[0093] ;

[0094] in, Indicates the largest sensor The hysteresis regularization kernel is an index of typical nonlinear correlation. express The corresponding lag time constant; Indicates the largest sensor The hysteresis regularization kernel is an index of typical nonlinear correlation. express The corresponding lag time constant;

[0095] Causal directionality is represented by the direction of the edges, if Then the direction of the edge is taken ;like Then the direction of the edge is taken ;like The direction is bidirectional; among them, It is an adjustable parameter, which is the threshold used to determine the pointer.

[0096] Preferably, the strength of the edge is determined in the following way:

[0097] When the direction of the edge is When the correlation strength is... ;

[0098] When the direction is When the correlation strength is... ;

[0099] When the direction is bidirectional, then and The correlation strength is:

[0100] .

[0101] Preferably, the step of evaluating the variation of each node in the directed causal differential network of the current time window based on the information transmission flow entropy, and calculating the comprehensive state index (SAMS) of the target sensor for each time window based on the information transmission entropy of the variation, includes the following steps:

[0102] For the target sensor, normal data from the previous N time steps are selected to form a reference time window, and sensor data from the current time step is added to form the current time window. The norm of the canonical nonlinear correlation matrix of the regularization kernel between the reference time window and the current time window is used as the difference between the directed regularization kernel canonical nonlinear correlation coefficient matrix.

[0103] ;

[0104] in, This is the regularized kernel of the window at the current time, which is a typical nonlinear correlation matrix. The regularized kernel of the reference time window is a typical nonlinear correlation matrix. The regularized kernel typical nonlinear correlation coefficient between sensor i and sensor j in the current time window; The regularized kernel typical nonlinear correlation coefficients of sensor i and sensor j within the reference time window;

[0105] The sum of the differences in the directed regularized kernel canonical nonlinear correlation coefficient matrices at all time points is taken as the total anomaly difference. , is represented as:

[0106] ;

[0107] Probability of calculating information flow entropy for:

[0108] ;

[0109] Where K represents the total number of time windows;

[0110] Normalizing the probability of the information flow entropy yields the normalized probability of the information flow entropy. ;

[0111] For sensors Calculate the variance on its reference sample set. , is represented as:

[0112] ;

[0113] in, It is a sensor The reading at time t in the reference sample set; It is a sensor The mean of the reference sample set; N is the number of time points in the reference sample.

[0114] For sensors Calculate its variance on the sample set at the current time. , is represented as:

[0115] ;

[0116] in, It is a sensor The reading at time t in the sample set at the current moment; It is a sensor The mean of the sample set at the current time; N is the number of reference sample time points, and Q is the number of samples added at the current time.

[0117] Computational Sensors Its own fluctuations are represented as follows:

[0118] ;

[0119] in, Reflection sensor Abnormal changes in reading fluctuations; larger values ​​indicate greater sensor instability. (This is relevant to the sensor.) The self-fluctuation was normalized to obtain a normalized sensor. Reading fluctuation ;

[0120] Calculate the abnormal differences of the sensor at the current moment. , is represented as:

[0121] ;

[0122] Computational Sensors The Comprehensive Status Index (SAMS) is expressed as:

[0123] ;

[0124] Among them, k and B are adjustable parameters that can be adjusted according to different operating conditions and different power equipment.

[0125] Preferably, it also includes: a target sensor The SAMS (Comprehensive Status Index) scores for all time windows are mapped to a line chart, including the following steps:

[0126] A two-dimensional coordinate system is constructed, with the X-axis representing the time window and the Y-axis representing the Comprehensive State Index (SAMS) score. The SAMS scores calculated by the target sensor within each time window are placed into the two-dimensional coordinate system to visualize the abnormal degradation of the target sensor, resulting in a two-dimensional visualization of the entire process.

[0127] Preferably, the method further includes: after removing the target sensor from the network, calculating the remaining nodes, and sequentially performing fault analysis on the associated nodes of the target sensor to obtain the removed sensors. The SAMS score, the comprehensive status index of the associated nodes, is calculated after removing sensors. If the SAMS (Status Index) score of the associated nodes is normal, then the target sensor is verified to be faulty.

[0128] The present invention also provides a sensor anomaly diagnosis device for a power equipment monitoring system, comprising:

[0129] The sample acquisition module collects data of continuously monitored variables from sensors in the power equipment monitoring system. It selects data from continuous sampling times of each sensor under the same operating condition of K power equipment units as experimental samples and preprocesses them.

[0130] The time window construction module sets the sliding window length, selects the first N data points after preprocessing as a reference data matrix, and constructs the current time window data matrix by setting the sliding step size based on the same sliding window length.

[0131] The associated node selection module calculates the typical nonlinear correlation of the regularization kernel between each sensor based on the reference data matrix, and selects the associated node sensors of the target sensor based on the magnitude of the typical nonlinear correlation of the regularization kernel between the target sensor and other sensors; the target sensor and the associated node sensors of the target sensor constitute the target sensor network nodes;

[0132] The directed causal difference network construction module calculates the typical nonlinear correlation of the regularization kernel between the target sensor and its associated nodes within the time lag window, determining the direction of the edges between the target sensor network nodes as causal directionality; it calculates and compares the typical nonlinear correlation of the regularization kernel between the target sensor and its associated nodes within the current time window with the typical nonlinear correlation of the regularization kernel within the reference time window to determine the strength of the edges between the target sensor network nodes; and based on the direction and strength of the edges between the target sensor network nodes, it constructs a directed causal difference network for all time windows.

[0133] The node score calculation module evaluates the variation of each node in the directed causal differential network in the current time window based on the information transmission flow entropy, and calculates the comprehensive state index (SAMS) of the target sensor for each time window based on the information transmission entropy flow of the variation.

[0134] The fault detection module sets a threshold. If the SAMS of the target sensor within a certain time window is higher than the threshold, the target sensor is considered to have malfunctioned.

[0135] The present invention has the following beneficial effects:

[0136] Compared with existing technologies, this invention can simultaneously consider the nonlinear correlation, time lag relationship and directed information transmission characteristics between sensors, construct a directed causal difference network, and form a comprehensive state index (SAMS) by combining the sensor's own fluctuation. Therefore, it can not only improve the accuracy, stability and interpretability of sensor anomaly detection under complex working conditions, but also achieve more sensitive identification of early anomalies such as gradual degradation and has better anomaly localization capabilities. At the same time, this method can achieve stable anomaly detection under small sample conditions with a limited number of samples.

[0137] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0138] Figure 1 This is a diagram illustrating the method steps of an embodiment of the present invention;

[0139] Figure 2 A flowchart of a fault monitoring method for power equipment sensor groups based on dynamic network markers of regularized kernel typical nonlinear correlation analysis is provided for embodiments of the present invention.

[0140] Figure 3 A visualization of the neighboring sensors of the target sensor for implementing the present invention;

[0141] Figure 4 This figure shows the typical nonlinear correlation changes of the regularized kernel of the temperature sensor of phase a of the high-voltage transformer in a wind turbine generator under healthy conditions.

[0142] Figure 5 A visualization of a typical nonlinear correlation dynamic network of a regularized kernel for a phase a temperature sensor of a high-voltage transformer in a wind turbine generator under healthy conditions.

[0143] Figure 6 SAMS chart showing the comprehensive status indicators of the high-voltage transformer A-phase temperature sensor in a wind turbine generator set from 00:00:00 on August 1, 2017 to 23:50:00 on August 11, 2017 under healthy conditions.

[0144] Figure 7 Figure showing the typical nonlinear correlation changes of the regularized kernel under drift fault for the temperature sensor of phase a high-voltage transformer in a wind turbine generator set under the condition of injecting drift fault.

[0145] Figure 8 Visualization of a typical nonlinear correlation dynamic network with a regularized kernel under drift fault for the temperature sensor of phase a high-voltage transformer in a wind turbine generator set;

[0146] Figure 9SAMS (Status Analysis and Management System) chart of the comprehensive status indicators from 00:00:00 on August 1, 2017 to 23:50:00 on August 11, 2017 under the condition of drift fault injection into the temperature sensor of phase A of high voltage transformer in wind turbine generator set;

[0147] Figure 10 A three-dimensional visualization of the SAMS (Comprehensive Status Index) values ​​of neighboring sensors after removing the temperature sensor of phase a of the high-voltage transformer in the wind turbine generator set;

[0148] Figure 11 This is a structural diagram of the device according to an embodiment of the present invention. Detailed Implementation

[0149] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0150] See Figure 1 The diagram shown is a method step diagram and a method flowchart of an embodiment of the present invention, including the following steps:

[0151] S101: Collect data of continuous monitoring variables collected by sensors in the power equipment monitoring system, select data from the continuous sampling times of each sensor under the same operating condition of K power equipment units as experimental samples and preprocess them.

[0152] S102, Set the sliding window length, select the first N data points of the preprocessed data as the reference data matrix, and construct the current time window data matrix by setting the sliding step size according to the same sliding window length;

[0153] S103, calculate the typical nonlinear correlation of the regularization kernel between each sensor based on the reference data matrix, and select the target sensor associated node sensor based on the magnitude of the typical nonlinear correlation of the regularization kernel between the target sensor and other sensors; the target sensor and the target sensor associated node sensor constitute the target sensor network node;

[0154] S104, calculate the typical nonlinear correlation of the regularization kernel between the target sensor and the sensors associated with the target sensor in the time lag window, and determine the direction of the edge between the target sensor network nodes as causal directionality; calculate and compare the typical nonlinear correlation of the regularization kernel between the target sensor and the sensors associated with the target sensor in the current time window with the typical nonlinear correlation of the regularization kernel in the reference time window to determine the strength of the edge between the target sensor network nodes; construct a directed causal difference network for all time windows based on the direction and strength of the edge between the target sensor network nodes.

[0155] S105, assess the variation of each node in the directed causal differential network in the current time window based on the information transmission flow entropy, and calculate the comprehensive state index (SAMS) of the target sensor for each time window based on the flow of information transmission entropy of the variation.

[0156] S106, Set a threshold. If the SAMS of the target sensor within a certain time window is higher than the threshold, the target sensor is considered to be abnormal.

[0157] Specifically, in step S101, continuous monitoring data from sensors in the power equipment unit is collected, and all data is processed. Let there be a total of... There are [number] sensors, and the number of time points for each sensor is [number]. , No. The sequence of sensors is denoted as:

[0158] ;

[0159] ;

[0160] in Indicates sensor The observation at time point j. The normalized representation is:

[0161] ;

[0162] in, for Sensors collected The observed value at time, For sensors Collected Normalized data of observations at time points; For sensors The average of the observations at all time points; For sensors The standard deviation of the observations at all time points.

[0163] Specifically, the construction of the reference sample data matrix and the current time window data matrix in S102 includes:

[0164] The first N data points are selected as reference sample data, and a reference sample data matrix is ​​constructed from them. :

[0165] ;

[0166] Set the sliding window size to L and the sliding step size to 1, iterate through each time point, and construct the time lag sliding window matrix as follows:

[0167] ;

[0168] The time-lag sliding window and the reference sample matrix are combined to form the current-time window matrix. The current-time window matrix is ​​then constructed as follows:

[0169] ;

[0170] Specifically, in S103, the regularized kernel typical nonlinear correlation coefficient matrix between the target sensor and each sensor is calculated, and other nodes with strong correlation to the target sensor are selected. The specific steps include:

[0171] Window sample pair construction, in any time window (Window length is) Within each sampling point, the sensor... With sensors The normalized sequences are as follows:

[0172] ;

[0173] ;

[0174] The Gram matrix of sensor i and sensor j within the time window is calculated using the Gaussian kernel function, and is represented using the Gaussian kernel function:

[0175] ;

[0176] ;

[0177] in, For sensors The value at time a, For sensors The value at time b; For sensors The value at time a, For sensors The value at time b.

[0178] Building sensors kernel Gram matrix and sensor The kernel Gram matrix (the Gram matrix characterizes the similarity of samples within a window in a nonlinear feature space):

[0179] ; ;

[0180] Center the Gram matrix:

[0181] ;

[0182] Where H is the centering matrix, and E is... An identity matrix, where Q is all 1s, with dimension 1. The matrix. Sensor With sensors The centering result of the kernel Gram matrix is:

[0183] ;

[0184] ;

[0185] To obtain the principal components of the centered kernel Gram matrix, solve for the eigenvalues ​​of the centered kernel Gram matrix:

[0186] ;

[0187] ;

[0188] in, For eigenvalues, This is a feature vector, and the subscript indicates the sensor.

[0189] Since the eigenvectors are not unique, before selecting eigenvalues... Using 10 eigenvectors (m≤L) as principal components to reduce dimensionality and avoid Numerical instability caused by singularities:

[0190] ;

[0191] ;

[0192] Calculate the kernel principal component score for the sensor. With sensors The projection scores of time a in the directions of the c principal components are calculated as follows:

[0193] ;

[0194] ;

[0195] in, Indicates sensor principal components Column c The p-th component, Indicates the Gaussian kernel function as a sensor The Gram matrix at row a and column p. Indicates sensor principal components Column c The Each component. This represents the Gaussian kernel function, which is the sensor. The Gram matrix at row a and column p.

[0196] sensor With sensors The m kernel principal component scores at time a are concatenated to obtain an m-dimensional score vector:

[0197] ;

[0198] ;

[0199] Stack the m-kernel principal component scores at each time step to construct an in-window sensor. With sensors Kernel principal component score matrix:

[0200] ;

[0201] ;

[0202] Perform CCA in the kernel principal component space to construct the autocovariance matrix and crosscovariance matrix:

[0203] ;

[0204] ;

[0205] ;

[0206] ;

[0207] in and It is the autocovariance matrix. This is the cross-covariance matrix. To avoid issues arising from the limited number of samples in the window... or Singular or pathological, introduce regularization terms:

[0208] ;

[0209] ;

[0210] Where I is the identity matrix. , This is the regularization coefficient.

[0211] Since industrial processes are often coupled with the combined effects of multiple variables, which cannot be represented by a single column, it is necessary to construct a CCA optimization. To find two weight vectors, the two sets of multidimensional data are weighted and synthesized into two new one-dimensional sequences, maximizing the correlation between the two sequences. This requires defining a regularization kernel, a typical nonlinear correlation function.

[0212] ;

[0213] Furthermore, since there are no boundary constraints, The covariance can be increased infinitely, which would render the objective function meaningless. Therefore, it is necessary to impose constraints on the objective function.

[0214] ;

[0215] ;

[0216] The above optimization problem can be transformed into a generalized eigenvalue problem:

[0217] ;

[0218] ;

[0219] Solving for this feature value yields a set of feature values. its square root , , The coefficients are canonical correlation coefficients, where the largest is the highest. It is represented as the first kernel canonical regularized correlation coefficient.

[0220] The first kernel canonical regularized correlation coefficient is used as a sensor. With sensors Non-linear correlation metrics within the window:

[0221] ;

[0222] in . The larger the value, the stronger the nonlinear correlation between the two sensors in the kernel principal component space.

[0223] The correlation coefficient matrix is ​​constructed by sequentially calculating the correlation between each sensor. :

[0224] .

[0225] Based on the regularized kernel's typical nonlinear correlation coefficient matrix, select the correlation coefficients with the target sensor. The most strongly correlated One sensor is the target sensor. The neighboring nodes.

[0226] Specifically, in S104, as the sampling time increases sequentially, the data samples contained in the window are also updated accordingly. The RKCCA correlation strength between the target sensor and each neighboring sensor with time lag is calculated:

[0227] ;

[0228] ;

[0229] Select The decision value is the one with the highest correlation to RKCCA during the lag time.

[0230] ;

[0231] ;

[0232] ;

[0233] in, This is the lag time constant with the highest correlation to RKCCA. If , direction ;like , direction ;like The direction is bidirectional; Parameters are adjusted according to different field conditions; in this embodiment, a value of 0.4 is used. A directed regularized kernel typical nonlinear correlation coefficient matrix is ​​constructed based on the directionality. When the direction is... When the correlation strength is... When the direction is When the correlation strength is... When the direction is bidirectional, then and Correlation strength is Finally, a directed kernel canonical nonlinear correlation matrix is ​​constructed. As shown in the table below:

[0234] Thus, a directed dynamic network at each time point is constructed based on the connection direction and correlation strength between the target sensor and the node sensors.

[0235] The difference between the regularized kernel canonical nonlinear correlation coefficient matrix in the reference sample matrix and the regularized kernel canonical nonlinear correlation coefficient matrix in the current time sample matrix is ​​characterized as the degree of change between the current network situation and the normal situation (reference sample). The difference is defined using the norm between the two regularized kernel canonical nonlinear correlation coefficient matrices:

[0236] ;

[0237] This represents the perturbation change in the canonical nonlinear correlation coefficients of the regularization kernels between the sensors as the window data changes over time. The directional determinations of the target sensor node and its neighboring sensor nodes remain unchanged, and the edges represent the norm difference between the two canonical nonlinear correlation coefficient matrices of the regularization kernels, thus constructing a directed causal difference network.

[0238] Specifically, in S105, to quantify the anomaly level of the target sensor, a Comprehensive State Index (SAMS) is assigned to the target sensor. One of the characteristics for quantifying the anomaly level of the target sensor is the flow of anomaly information entropy. The sum of the differences in the directed regularized kernel typical nonlinear correlation coefficient matrices at all time points is taken as the total anomaly difference.

[0239] ;

[0240] The probability of calculating the entropy of the information flow is:

[0241] ;

[0242] Where K represents the number of time windows. Since the information flow entropy values ​​obtained from different networks vary, we normalized the probability of the information flow entropy to ensure consistency in subsequent score calculations across different networks. Specifically, we used a min-max normalization method to linearly map each feature value to the [0, 1] interval. This process avoids model instability caused by significant differences in the scale of subsequent feature calculations. The normalized probability of the information flow entropy is set as... .

[0243] The second characteristic for quantifying the degree of anomaly is the fluctuation of the sensor itself. The fluctuation of the target sensor is calculated. Calculate the variance on its reference sample set:

[0244] ;

[0245] in It is a sensor The reading at time t in the reference sample set; It is a sensor The mean value on the reference sample set; N is the number of time points in the reference sample. For the target sensor Calculate its variance on the sample set at the current time:

[0246] ;

[0247] in It is a sensor The reading at time t in the sample set at the current moment; It is a sensor The mean of the sample set at the current time; N is the number of reference sample time points, and Q is the number of samples added at the current time. Calculate the target sensor. Self-fluctuation status:

[0248] ;

[0249] in, Reflection sensor Abnormal fluctuations in readings indicate sensor instability; larger values ​​suggest greater sensor instability. Similar to the previous section, this is to ensure the probability of the information flow entropy matches the target sensor. The consistency of its own fluctuations, for sensors The inherent fluctuations were normalized. Specifically, a minimum-maximum normalization method was used to linearly map each eigenvalue to the [0, 1] interval. The sensor... After normalization of reading fluctuations, it becomes .

[0250] The current anomalous difference of the target sensor is calculated as AS:

[0251] ;

[0252] Calculate target sensor The Comprehensive Status Index (SAMS) is as follows:

[0253] ;

[0254] Where k and B are adjustable parameters that can be adjusted according to different operating conditions and different power equipment.

[0255] Specifically, to further verify whether the fault lies with the target sensor, after removing the target sensor, fault analysis is performed on the associated node sensors within the target sensor to determine which sensors were removed. For the comprehensive status index SAMS score of the post-associated node, if the comprehensive status index SAMS score is normal after removing the target sensor, it can be confirmed that the target sensor has failed.

[0256] Continuously repeat the fault monitoring method and device for the power equipment sensor group of the dynamic network biomarker by regularized kernel canonical nonlinear correlation analysis. For the monitoring of different target sensors, put the comprehensive status index SAMS scores of each neighbor sensor after removing the target sensor into a three-dimensional coordinate system, and obtain a three-dimensional coordinate diagram with the time window on the X-axis, the associated node sensor of the target sensor on the Y-axis, and the comprehensive status index SAMS score of the associated node sensor on the Z-axis, and finally find all the faulty sensors.

[0257] Specifically, when it is confirmed that the comprehensive status index SAMS score of the associated node sensor after removing the target sensor is normal, the target sensor Map the comprehensive status index SAMS scores within all time windows into a line chart, construct a two-dimensional coordinate system, use the X-axis to represent the time window, and the Y-axis to represent the comprehensive status index SAMS score. Put the comprehensive status index SAMS scores calculated for the target sensor within each time window into the two-dimensional coordinate system to visualize the abnormal deterioration of the target sensor. Obtain a full-process two-dimensional visualization chart.

[0258] Conduct a verification experiment on the embodiment of the present invention. Take the data collected by the sensors of the T01 wind turbine in a certain wind farm as the object. This unit has a horizontal-axis three-blade structure, a rated power of 2 MW, a rated wind speed of 12 m / s, and cut-in and cut-out wind speeds of 4 m / s and 25 m / s respectively. The rotor diameter of the wind turbine is 90 m, and its maximum rotational speed can reach 14.9 r / min. The gearbox adopts a three-stage planetary gear structure to drive an asynchronous generator, with a maximum rotational speed of 2016 r / min, a rated voltage of 690 V, and a grid-connected power frequency of 50 Hz. The tower of the unit is made of steel pipe structure, and the hub height is 80 m. Data acquisition is completed by the SCADA system supporting the unit, and the sampling period is 10 min. In this experiment, data at 50 consecutive sampling moments (i.e., N = 50) are selected to construct a reference time window. The current time window sample adopts a sliding window mechanism, the length of the time sliding window is 20, and the window sliding step is set to 1 sampling moment, so that all time points can be traversed. Combine the sliding time window with the reference time window as the current time window. Set the target sensor as the temperature sensor of phase a of the high-voltage transformer, and set the top 20 (i.e., Q = 20) sensors with the strongest correlation with it as neighbor nodes. If the comprehensive status index SAMS of the temperature sensor of phase a of the high-voltage transformer is abnormally high, and the time window continuously mutates within a short period of time, it is determined that the time period meets the early warning trigger condition, and the time period from when the comprehensive status index SAMS starts to be abnormally high to when it returns to normal is the sensor fault time period.

[0259] Under normal operating conditions, the operating data of the target wind turbine's health sensor from 00:00:00 on August 1, 2017 to 23:50:00 on August 11, 2017, was imported, totaling 1576 data collection points (i.e., K=1576). A reference sample window and a sample window for the current time were constructed based on the set parameters, and experiments were conducted. First, sensors with strong correlations were selected as neighbor sensors. Figure 3 The process of selecting neighboring sensors for the high-voltage transformer phase a temperature sensor is shown. The darker colored sensors are the top 20 sensors that are strongly correlated with the high-voltage transformer phase a temperature sensor. Figure 4 This presentation visually illustrates the typical nonlinear correlation changes of the regularized kernel under healthy conditions of the temperature sensor in phase a of a high-voltage transformer. Data from 23:10:00 on August 1, 2017 to 0:40:00 on August 2, 2017 is used for the display. Darker colors in the correlation heatmap indicate stronger correlations. Figure 4 It can be seen that although the typical nonlinear correlation of the regularized kernel has changed slightly, the overall trend remains stable, indicating that no abnormal situation has occurred. Figure 5 This paper presents the normalized kernel typical nonlinear correlation dynamic network of the high-voltage transformer phase A temperature sensor at 13:10:00 on August 11, 2017, under healthy conditions. Orange nodes represent the target sensor node, i.e., the high-voltage transformer phase A temperature sensor, and the remaining nodes are neighboring sensor nodes. The strength of the connecting edges represents the correlation strength between the two nodes; the darker the color, the weaker the correlation between the two nodes, and vice versa. The causal relationship between each node is indicated by arrows. To visually represent the evolution of the Comprehensive State Index (SAMS) of the healthy sensor over time, the obtained SAMS of the high-voltage transformer phase A temperature sensor is projected onto a two-dimensional coordinate graph, as shown below. Figure 6 As shown. By Figure 6 It can be seen that during the period when the temperature sensor of phase a of the high-voltage transformer is in a healthy state, the comprehensive status index (SAMS) score of each collection point shows oscillation and fluctuation, but the overall trend remains stable and no significant abnormalities occur, which verifies the accuracy of the method in identifying the health sensor under normal operating conditions of wind turbine generator sets.

[0260] A sensor drift fault was actively introduced into the temperature sensor of phase A of the high-voltage transformer. Within a selected fault time window, a deviation value, initially 0, was superimposed onto the original sensor data, increasing linearly with the time step to a preset maximum value. This simulated a drift fault where the sensor readings gradually deviated from the normal value. The fault time point was set from 00:00:00 on August 11, 2017 to 00:00:00 on August 12, 2017 (136 data collection points). Figure 7This paper demonstrates the changes in regularization and kernel canonical nonlinear correlation of the temperature sensor of phase A of a high-voltage transformer under a drift fault from 21:10:00 to 23:50:00 on August 11, 2017. Figure 4 The comparison shows that the regularization and kernel typical nonlinear correlation changed significantly under sensor drift faults, resulting in obvious differences in the correlation coefficient matrix. The correlation between many sensors decreased significantly, which intuitively shows that when the target sensor fails, the correlation network of each sensor will change. Figure 8 This paper presents a regularized kernel-based typical nonlinear dynamic causal network of the high-voltage transformer's phase A temperature sensor at 13:10:00 on August 11, 2017, under a drift fault. This network can be intuitively compared with... Figure 5 A comparison of the regularized kernel typical nonlinear dynamic causal networks at 13:10:00 on August 11, 2017, clearly shows the differences between the regularized kernel typical nonlinear dynamic causal networks at the same time under drift faults. Many connection edges and causal directions in the network have changed, which once again confirms that when a sensor fault occurs, it will cause a sudden change in the connection between the networks, thus providing a basis for our fault analysis.

[0261] Based on the above results, the Comprehensive Status Index (SAMS) of the phase a temperature sensor of the high-voltage transformer is evaluated according to pre-set rules. Figure 9 The results show that before 00:00:00 on August 11, 2017, the Comprehensive Status Index (SAMS) of the high-voltage transformer phase a temperature sensor maintained a low score with small overall fluctuations, indicating that the sensor was healthy at that time. At 00:00:00 on August 11, the SAMS index underwent a significant change and subsequently oscillated continuously, meeting the alarm conditions. This confirms that when a drift fault is introduced in the high-voltage transformer phase a temperature sensor, the SAMS index will undergo a sudden change. Figure 10 The results show that after removing the high-voltage transformer phase a temperature sensor, during the period from 00:00:00 on August 11, 2017 to 00:00:00 on August 12, 2017 (136 data collection points), the comprehensive status index (SAMS) of each neighboring sensor tended to stabilize, the overall trend remained stable, and no significant abnormalities were found. Therefore, the high-voltage transformer phase a temperature sensor was identified as faulty.

[0262] See Figure 11 The diagram shown is a structural diagram of a device according to an embodiment of the present invention, comprising:

[0263] The sample acquisition module 1101 collects data of continuous monitoring variables collected by sensors in the power equipment monitoring system, selects data from the continuous sampling times of each sensor under the same operating condition of K power equipment units as experimental samples and preprocesses them.

[0264] The time window construction module 1102 sets the sliding window length, selects the first N data points of the preprocessed data as a reference data matrix, and constructs the time window data matrix at the current time by setting the sliding step size according to the same sliding window length.

[0265] The associated node selection module 1103 calculates the typical nonlinear correlation of the regularization kernel between each sensor based on the reference data matrix, and selects the associated node sensors of the target sensor based on the magnitude of the typical nonlinear correlation of the regularization kernel between the target sensor and other sensors; the target sensor and the associated node sensors of the target sensor constitute the target sensor network nodes;

[0266] The directed causal difference network construction module 1104 calculates the typical nonlinear correlation of the regularization kernel between the target sensor and the sensors associated with the target sensor in the time lag window, and determines the direction of the edges between the target sensor network nodes as causal directionality; it calculates and compares the typical nonlinear correlation of the regularization kernel between the target sensor and the sensors associated with the target sensor in the current time window with the typical nonlinear correlation of the regularization kernel in the reference time window to determine the strength of the edges between the target sensor network nodes; based on the direction and strength of the edges between the target sensor network nodes, it constructs a directed causal difference network for all time windows.

[0267] The node score calculation module 1105 evaluates the variation of each node in the directed causal differential network in the current time window based on the information transmission flow entropy, and calculates the comprehensive state index (SAMS) of the target sensor for each time window based on the information transmission entropy flow of the variation.

[0268] The fault detection module 1106 sets a threshold. When the SAMS of the target sensor within a certain time window is higher than the threshold, the target sensor is considered to be abnormal.

[0269] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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. A method for diagnosing sensor anomalies in a power equipment monitoring system, characterized in that, Includes the following steps: Data of continuously monitored variables collected by sensors in the power equipment monitoring system were collected. Data from the continuous sampling times of each sensor under the same operating condition of K power equipment units were selected as experimental samples and preprocessed. Set the sliding window length, select the first N data points of the preprocessed data as the reference data matrix, and construct the time window data matrix at the current time step by setting the sliding step size according to the same sliding window length. The typical nonlinear correlation of the regularization kernel between each sensor is calculated based on the reference data matrix. The target sensor associated node sensor is selected based on the magnitude of the typical nonlinear correlation of the regularization kernel between the target sensor and other sensors. The target sensor and the target sensor associated node sensor constitute the target sensor network node. Calculate the typical nonlinear correlation of the time lag window between the target sensor and the sensor associated with the target sensor node, and determine the direction of the edge between the target sensor network nodes as causal directionality. The strength of the edges between the target sensor network nodes is determined by comparing the typical nonlinear correlation of the regularization kernel between the target sensor and the sensor associated with the target sensor in the current time window with the typical nonlinear correlation of the regularization kernel in the reference time window. Based on the direction and strength of the edges between nodes in the target sensor network, a directed causal difference network is constructed for all time windows; The variation of each node in the directed causal differential network in the current time window is evaluated based on the information transmission flow entropy, and the comprehensive state index (SAMS) of the target sensor for each time window is calculated based on the information transmission entropy flow of the variation. A threshold is set, and if the SAMS of the target sensor exceeds the threshold within a certain time window, the target sensor is considered to be abnormal.

2. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 1, characterized in that, The system collects continuous monitoring data from sensors in the power equipment unit, processes all data, and assumes a total of... There are [number] sensors, and the number of time points for each sensor is [number]. , No. The sequence of sensors is denoted as: ， ； in, Indicates sensor In the Observations at a given time point; Normalization is represented as: ； in, For sensors exist The observed value at time, For sensors exist Normalized data of observations at time points; For sensors The average of the observations at all time points; For sensors The standard deviation of the observations at all time points.

3. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 2, characterized in that, The step of selecting the first N data points from the preprocessed data as a reference data matrix is ​​represented as follows: ； Among them, the reference data matrix elements in This represents preprocessed data representing observations from different sensors at different time points. Setting a time-lag sliding window is represented as follows: ； Among them, time-lag sliding window elements in This represents the values ​​of different variables from different sensors within the current time window; the last row of data represents the current... Data at any given time, number of columns Represents the number of sensors; Set the sliding window size to L and the sliding step size to 1. Combine the time-lag sliding window with the reference sample matrix to form the current time-time window matrix, represented as: ； in, for The window matrix at each time step; Repeat the above steps Next, obtain the window matrix that iterates through all time points.

4. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 2, characterized in that, The calculation of the typical nonlinear correlation of the regularized kernel includes the following steps: Window sample pair construction, in any time window Inside, take the sensor With sensors The normalized sequences are as follows: ； ； Where L represents the number of sampling points within the window; Using Gaussian kernel function to calculate time window sensor Gram matrix With sensors Gram matrix , is represented as: ； ； in, For sensors The value at time a, For sensors The value at time b; For sensors The value at time a, For sensors The value at time b; Building sensors kernel Gram matrix With sensors kernel Gram matrix : ； ； The kernel Gram matrix is ​​centered to obtain the sensor. Centralized kernel Gram matrix With sensors Centralized kernel Gram matrix , respectively represented as: ； ； Where H is the centering matrix, E is The identity matrix, where Q is all 1s and dimension is 1. Matrix; To obtain the principal components of the centered kernel Gram matrix, we solve for the eigenvalues ​​of the two centered kernel Gram matrices, which are expressed as follows: ； ； in, and They are respectively eigenvalues ​​and eigenvectors, and They are respectively eigenvalues ​​and eigenvectors; Before selecting eigenvalues Using 10 feature vectors as principal components to reduce dimensionality and avoid The numerical instability caused by singularities is represented as: ； ； Where m≤L; and Sensors With sensors Principal components; Computational Sensors With sensors The projection scores of time a along the directions of the c principal components are expressed as: ； ； in, Indicates sensor principal components Column c The p-th component, Indicates sensor The value in the a-th row and p-th column of the Gram matrix; Indicates sensor principal components Column c The One component; Indicates sensor The value in the a-th row and p-th column of the Gram matrix; Sensors respectively With sensors The m kernel principal component scores at time a are concatenated to obtain an m-dimensional score vector. and , is represented as: ； ； Stack the m-kernel principal component scores at each time step to construct an in-window sensor. Kernel principal component score matrix With sensors Kernel principal component score matrix , is represented as: ； ； Perform CCA in the kernel principal component space to construct the autocovariance matrix and crosscovariance matrix: ； ； ； ； in, and Sensors Autocovariance matrix and sensor The autocovariance matrix, Indicates from The amount to The cross-covariance matrix form of the components, Indicates from The amount to The cross-covariance matrix form of the components; To avoid the autocovariance matrix becoming singular or ill-conditioned due to the limited number of samples in the window, a regularization term is introduced, expressed as: ； ； in, Indicates sensor The introduction of the autocovariance matrix of the regularization term... Indicates sensor The introduction of the regularization term introduces the autocovariance matrix; and All are identity matrices. and Both represent regularization coefficients; Define the regularized kernel canonical nonlinear correlation function as follows: ； in, This represents the regularized kernel canonical correlation coefficient. Indicates sensor The corresponding weighted coefficient vector of the kernel principal component scores; Indicates sensor The corresponding weighted coefficient vector of the kernel principal component scores; Imposing constraints on the canonical nonlinear correlation function of the regularized kernel, expressed as: ； ； The eigenvalues ​​are solved using the following formula: ； ； Solving yields a set of eigenvalues its square root , , The coefficients are canonical correlation coefficients, where the largest is... Represented as the first kernel canonical regularized correlation coefficient; The first kernel canonical regularized correlation coefficient is used as a measure of the sensor. With sensors Indices of typical nonlinear correlation of the regularization kernel within the window. , is represented as: ； in, ; The larger the value, the stronger the nonlinear correlation between the two sensors in the kernel principal component space; Inter-sensor Calculate the correlation coefficient matrix in sequence. , is represented as: 。 5. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 1, characterized in that, The calculation of the regularized kernel typical nonlinear correlation of the time lag windows between the target sensor and the sensors associated with the target sensor, and the determination of the direction of the edges between the target sensor network nodes as causal directionality, includes the following steps: Set time lag The largest constant is With sensors As the target sensor, with sensor As the sensor associated with the target sensor, for each hysteresis... Computational Sensors and sensors The regularized kernel's typical nonlinear correlation is expressed as: ； ； in, This represents an index used to calculate the typical nonlinear correlation of the regularization kernel. Indicates sensor For sensors The hysteresis regularization kernel is an index of typical nonlinear correlation. Indicates sensor For sensors The hysteresis regularization kernel is an indicator of typical nonlinear correlation. This represents the kernel principal component score matrix of sensor i at time t. This represents the kernel principal component score matrix of sensor j at time t. This represents the kernel principal component score matrix of sensor i at time (th). Let represent the kernel principal component score matrix of sensor j at time (th); The decision quantile is calculated by selecting the index of the canonical nonlinear correlation of the largest hysteresis regularization kernel. , is represented as: ； ； ； in, Indicates the largest sensor The hysteresis regularization kernel is an index of typical nonlinear correlation. express The corresponding lag time constant; Indicates the largest sensor The hysteresis regularization kernel is an index of typical nonlinear correlation. express The corresponding lag time constant; Causal directionality is represented by the direction of the edges, if Then the direction of the edge is taken ;like Then the direction of the edge is taken ;like The direction is bidirectional; among them, It is an adjustable parameter, which is the threshold used to determine the pointer.

6. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 5, characterized in that, The strength of the edge is determined in the following way: When the direction of the edge is When the correlation strength is... ; When the direction is When the correlation strength is... ; When the direction is bidirectional, then and The correlation strength is: 。 7. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 1, characterized in that, The process of evaluating the variation of each node in the directed causal differential network within the current time window based on the information transfer flow entropy, and calculating the comprehensive state index (SAMS) of the target sensor for each time window based on the information transfer entropy flow of the variation, includes the following steps: For the target sensor, normal data from the previous N time steps are selected to form a reference time window, and sensor data from the current time step is added to form the current time window. The norm of the canonical nonlinear correlation matrix of the regularization kernel between the reference time window and the current time window is used as the difference between the directed regularization kernel canonical nonlinear correlation coefficient matrix. ； in, This is the regularized kernel of the window at the current time, which is a typical nonlinear correlation matrix. The regularized kernel of the reference time window is a typical nonlinear correlation matrix. The regularized kernel typical nonlinear correlation coefficient between sensor i and sensor j in the current time window; The regularized kernel typical nonlinear correlation coefficients of sensor i and sensor j within the reference time window; The sum of the differences in the directed regularized kernel canonical nonlinear correlation coefficient matrices at all time points is taken as the total anomaly difference. , is represented as: ； Probability of calculating information flow entropy for: ； Where K represents the total number of time windows; Normalizing the probability of the information flow entropy yields the normalized probability of the information flow entropy. ; For sensors Calculate the variance on its reference sample set. , is represented as: ； in, It is a sensor The reading at time t in the reference sample set; It is a sensor The mean of the reference sample set; N is the number of time points in the reference sample. For sensors Calculate its variance on the sample set at the current time. , is represented as: ； in, It is a sensor The reading at time t in the sample set at the current moment; It is a sensor The mean of the sample set at the current time; N is the number of reference sample time points, and Q is the number of samples added at the current time. Computational Sensors Its own fluctuations are represented as follows: ； in, Reflection sensor Abnormal changes in reading fluctuations; larger values ​​indicate greater sensor instability. (This is relevant to the sensor.) The self-fluctuation was normalized to obtain a normalized sensor. Reading fluctuation ; Calculate the abnormal differences of the sensor at the current moment. , is represented as: ； Computational Sensors The Comprehensive Status Index (SAMS) is expressed as: ； Among them, k and B are adjustable parameters that can be adjusted according to different operating conditions and different power equipment.

8. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 1, characterized in that, Also includes: target sensor The SAMS (Comprehensive Status Index) scores for all time windows are mapped to a line chart, including the following steps: A two-dimensional coordinate system is constructed, with the X-axis representing the time window and the Y-axis representing the Comprehensive State Index (SAMS) score. The SAMS scores calculated by the target sensor within each time window are placed into the two-dimensional coordinate system to visualize the abnormal degradation of the target sensor, resulting in a two-dimensional visualization of the entire process.

9. The method for diagnosing sensor anomalies in a power equipment monitoring system according to claim 1, characterized in that, Also includes: After removing the target sensor from the network, the remaining nodes are calculated. Fault analysis is then performed on the associated sensors within the target sensor to determine which sensors were removed. The SAMS score, the comprehensive status index of the associated nodes, is calculated after removing sensors. If the SAMS (Status Index) score of the associated nodes is normal, then the target sensor is verified to be faulty.

10. A sensor anomaly diagnosis device for a power equipment monitoring system, characterized in that, include: The sample acquisition module collects data of continuously monitored variables from sensors in the power equipment monitoring system. It selects data from continuous sampling times of each sensor under the same operating condition of K power equipment units as experimental samples and preprocesses them. The time window construction module sets the sliding window length, selects the first N data points after preprocessing as a reference data matrix, and constructs the current time window data matrix by setting the sliding step size based on the same sliding window length. The associated node selection module calculates the typical nonlinear correlation of the regularization kernel between each sensor based on the reference data matrix, and selects the associated node sensors of the target sensor based on the magnitude of the typical nonlinear correlation of the regularization kernel between the target sensor and other sensors; the target sensor and the associated node sensors of the target sensor constitute the target sensor network nodes; The directed causal difference network construction module calculates the regularized kernel typical nonlinear correlation between the time lag windows of the target sensor and the sensor associated nodes, and determines the direction of the edges between the target sensor network nodes as causal directionality. The strength of the edges between the target sensor network nodes is determined by comparing the typical nonlinear correlation of the regularization kernel between the target sensor and the sensor associated with the target sensor in the current time window with the typical nonlinear correlation of the regularization kernel in the reference time window. Based on the direction and strength of the edges between nodes in the target sensor network, a directed causal difference network is constructed for all time windows; The node score calculation module evaluates the variation of each node in the directed causal differential network in the current time window based on the information transmission flow entropy, and calculates the comprehensive state index (SAMS) of the target sensor for each time window based on the information transmission entropy flow of the variation. The fault detection module sets a threshold. If the SAMS of the target sensor within a certain time window is higher than the threshold, the target sensor is considered to have malfunctioned.

Images