Thermal power plant measuring point time sequence anomaly analysis method based on multi-modal

Abstract

The application provides a kind of based on multimodal power plant measuring point time sequence anomaly analysis method, it is related to abnormal detection field, for the multimodal power plant measuring point data that changes with time, RDG-PACNet prediction model is proposed, it is composed of residual center module, scale fusion module and anomaly detection module, wherein residual center module is used to fuse neighborhood residual and update center set;Scale fusion module completes cross-scale alignment and feature fusion;Anomaly detection module summarizes scale anomaly probability, determines abnormal analysis result.The application has the beneficial effects that: by fusing adjacency matrix, neighborhood weighted and secondary centering processing, the abnormal change deviating from neighborhood center can be highlighted, by combining probability matrix, sample weight and binary coding matrix to update the center set, the concentration of feature distribution under the same center is improved by using residual variance constraint loss, so as to enhance the separability of normal segment and abnormal segment.

Description

Technical Field

[0001] This invention belongs to the field of anomaly detection, specifically relating to a method for analyzing time-series anomalies at measurement points in thermal power plants based on multimodal methods. Background Technology

[0002] With the development of smart power plant construction, a large number of monitoring points are deployed in the production process of thermal power plants to collect various types of operational data such as temperature, pressure, flow, and load, providing a data foundation for equipment condition monitoring. Existing time series anomaly analysis methods for monitoring points mostly model single data sources, making it difficult to fully utilize the correlation characteristics between different monitoring points and the complementary relationships between multimodal information. At the same time, the operation process of thermal power units has strong coupling, nonlinearity, and dynamic change characteristics. Traditional threshold-based or statistical analysis methods are difficult to accurately distinguish between normal fluctuations and abnormal states, which can easily lead to false alarms and missed alarms. Therefore, it is very important to integrate multimodal time series data of monitoring points, explore the changing patterns of equipment operating status, and achieve accurate analysis of monitoring point anomalies in thermal power plants.

[0003] In the existing technology, there are some invention patents that apply statistical analysis and data modeling methods to the detection of equipment operation anomalies. For example, patent publication number CN122347275A discloses a statistical method and system for analyzing fishery ecological health anomalies. By dividing the target fishery waters into multiple continuous water layers and collecting multi-factor time series data such as temperature, dissolved oxygen, pH, and nutrient concentration in each water layer, an ecological factor data synergy matrix is ​​constructed. The mean difference within the layer, the matrix structure difference value, and the cross-water layer ecological anomaly index are calculated to achieve fishery ecological health anomaly analysis. This scheme can identify ecological structure anomalies through multi-factor synergy relationships and is suitable for vertical ecological status monitoring scenarios in fishery waters. However, its detection object is aquatic ecological environment data, focusing on changes in the synergy relationships of ecological factors between different water layers and health status evaluation. It does not model the multi-variable, multi-scale operational status changes, inter-point correlation characteristics, and abnormal evolution patterns in the long-term operation process of thermal power plant measurement point time series data. Therefore, it cannot be directly applied to the scenario of analyzing the abnormal operation status of thermal power plant equipment.

[0004] Another patent, CN122241535A, discloses a fracturing pump hydraulic end operation anomaly diagnosis system. It mainly collects multi-source operation monitoring data and reference phase data of the fracturing pump hydraulic end, and combines it with cycle splitting, operating condition normalization, multi-physics field positive coupling simulation model, health residual matrix and trend residual matrix construction to realize fracturing pump hydraulic end anomaly diagnosis and fault location. This solution can identify equipment operation anomalies using equipment mechanism model and residual analysis method, and is suitable for fracturing pump hydraulic end fault diagnosis scenarios. However, it relies on specific equipment structural parameters, operating cycle phase information and physical simulation model. It mainly solves the problem of mechanical equipment fault deterioration diagnosis. It does not address the characteristic changes at different time scales in multi-measuring point time series data of thermal power plants, the implicit correlation between measuring points, and the problem of cross-scale anomaly information fusion. Therefore, it is difficult to meet the needs of complex thermal power plant measuring point time series anomaly detection. Summary of the Invention

[0005] This invention provides a method for analyzing time-series anomalies at thermal power plant measurement points based on multimodal data. For multimodal thermal power plant measurement point data that varies over time, an RDG-PACNet prediction model is proposed, which consists of a residual center module, a scale fusion module, and an anomaly detection module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps: Collect operating data of individual thermal power units to construct data from measuring points in thermal power plants; After the thermal power plant measurement data is averaged and pooled at different scales, it is reversibly normalized, divided into segments, and mapped to obtain the thermal power plant measurement data embedding. The measurement points of the thermal power plant are embedded and flattened into a two-dimensional feature representation matrix. The feature constraint time adjacency matrix and the feature similarity adjacency matrix are calculated, and the feature similarity neighborhood center residual matrix and the time neighborhood center residual matrix are calculated. The mean and variance of the feature-constrained temporal adjacency matrix, the feature-similar adjacency matrix, the feature-similar neighborhood center residual matrix, and the temporal neighborhood center residual matrix are calculated and normalized to obtain the feature-similar neighborhood residual weight and the temporal neighborhood residual weight. The residual weights of the feature similarity neighborhood and the residual weights of the time neighborhood are weighted based on the residual vectors in the residual matrix of the feature similarity neighborhood center and the residual matrix of the time neighborhood center, and the feature similarity adjacency matrix and the feature constraint time adjacency matrix are fused to calculate the two-dimensional feature representation matrix after quadratic centering. The two-dimensional feature representation matrix after quadratic centering is mapped to a feature representation, and the similarity value is calculated by combining the center set. After forming a probability matrix, the sample weights are calculated. The probability matrix is ​​sampled to generate a binary encoding matrix, the center set is updated to calculate the residual variance constraint loss, and the residual center module is constructed. The enhanced representation is calculated using the probability matrix and the updated center set. After mapping and interpolation, the relation transfer matrix and the previous probability matrix are obtained. The aligned representation is calculated by the probability matrix and the previous probability matrix. Combined with the enhanced representation, a fused feature representation is formed, and a scale fusion module is constructed. The fused feature representation is mapped to the scale anomaly probability, and the comprehensive anomaly probability is calculated. The time series anomaly analysis results of the thermal power plant measurement points are compared with the anomaly judgment threshold. The learnable parameters are updated based on the residual variance constraint loss to construct an anomaly detection module.

[0007] Preferably, relevant data from measuring points in the thermal power plant are collected, including main steam temperature data, main steam pressure data, feedwater flow data, fuel flow data, turbine vibration amplitude data, generator active power data, and unit load rate data. The original dataset is constructed and preprocessed, and the dataset is divided into two parts in a ratio of 7:2:1.

[0008] Preferably, the thermal power plant measurement point embedding is obtained by: performing average pooling on the thermal power plant measurement point data at different scales to generate thermal power plant measurement point time series at multiple scales; dividing the thermal power plant measurement point time series at each scale into multiple segments through reversible instance normalization; and then mapping each segment to a specified feature space to obtain the thermal power plant measurement point embedding composed of the number of segments, the segment embedding dimension, and the number of thermal power plant measurement point variables.

[0009] Preferably, the feature-constrained temporal adjacency matrix and the feature-similar adjacency matrix are calculated, and the feature-similar neighborhood center residual matrix and the temporal neighborhood center residual matrix are also calculated. Specifically, the embeddings of thermal power plant measurement points are merged according to the segment embedding dimension and the number of thermal power plant measurement point variables, so that the thermal power plant measurement point embedding corresponding to each segment is converted into a one-dimensional feature vector, and arranged in chronological order to obtain a two-dimensional feature representation matrix. The one-dimensional feature vector of each segment is multiplied by the query feature mapping parameter and the key feature mapping parameter, respectively. After scale normalization and exponential normalization, the feature-similar adjacency weights are obtained. A feature similarity adjacency matrix is ​​formed. The temporal distance is calculated based on the sequence number difference between segments, and the difference between one-dimensional feature vectors is calculated. The initial weights of temporal adjacency are obtained by combining the temporal distance decay parameter, the feature difference decay parameter, and the flattened feature dimension. After normalization, a feature-constrained temporal adjacency matrix is ​​formed. The feature similarity adjacency matrix, the feature-constrained temporal adjacency matrix, and the two-dimensional feature representation matrix are multiplied respectively. The result of multiplying the two-dimensional feature representation matrix with the corresponding matrix is ​​subtracted at corresponding positions to obtain the feature similarity neighborhood center residual matrix and the temporal neighborhood center residual matrix.

[0010] Preferably, the residual weights of the feature similarity neighborhood and the residual weights of the time neighborhood are obtained as follows: The weights in the feature similarity adjacency matrix and the corresponding residual vectors in the feature similarity neighborhood center residual matrix are multiplied and summed to obtain the mean of the feature similarity neighborhood residuals. Then, the difference between each residual vector and the mean of the feature similarity neighborhood residuals is calculated. This difference is then weighted and summed using the flattened feature dimension and the weights in the feature similarity adjacency matrix to obtain the variance of the feature similarity neighborhood residuals. Finally, the weights in the feature constraint time adjacency matrix and the corresponding residual vectors in the time neighborhood center residual matrix are multiplied and summed to obtain the variance of the feature similarity neighborhood residuals. The residual vectors are multiplied and summed to obtain the mean of the time neighborhood residuals. Then, the differences between each residual vector and the mean of the time neighborhood residuals are calculated. The weighted sum is then performed by combining the flattened feature dimensions and the weights in the feature constraint time adjacency matrix to obtain the variance of the time neighborhood residuals. The variances of the feature similarity neighborhood residuals and the time neighborhood residuals are processed separately by combining the learnable scale parameter. The results of the processing of the variances of the feature similarity neighborhood residuals and the time neighborhood residuals are negativeized and normalized to obtain the weights of the feature similarity neighborhood residuals and the weights of the time neighborhood residuals.

[0011] Preferably, the calculation of the two-dimensional feature representation matrix after quadratic centering specifically involves: multiplying the residual weights of the feature similarity neighborhood residuals with the residual vectors of the corresponding segments in the feature similarity neighborhood center residual matrix; simultaneously multiplying the residual weights of the time neighborhood residuals with the residual vectors of the corresponding segments in the time neighborhood center residual matrix; and adding the results of multiplying the residual weights of the time neighborhood residuals with the residual vectors of the corresponding segments in the time neighborhood center residual matrix and the results of multiplying the residual weights of the feature similarity neighborhood residuals with the residual vectors of the corresponding segments in the feature similarity neighborhood center residual matrix to obtain the residual matrix; and then constructing the feature similarity neighborhood residuals and the time neighborhood residuals according to the segment order. Construct a diagonal weight matrix, multiply it by the feature similarity adjacency matrix and the feature constraint temporal adjacency matrix respectively, and then add the results of the diagonal weight matrix multiplied by the feature similarity adjacency matrix and the feature constraint temporal adjacency matrix respectively. After row normalization, a fused adjacency matrix is ​​obtained. Based on the fused adjacency matrix, the residual matrix is ​​subjected to neighborhood weighting to obtain the fused residual neighborhood center matrix. The centering learnable parameters are input into the activation function to obtain the centering coefficients. The fused residual neighborhood center matrix is ​​then multiplied by the centering coefficients, and the residual matrix is ​​subtracted from the multiplication result according to the corresponding positions to obtain the two-dimensional feature representation matrix after secondary centering.

[0012] Preferably, the sample weights are calculated as follows: the two-dimensional feature representation matrix after quadratic centering is input into the linear layer to obtain the feature representation, and a center set containing multiple center vectors at the current scale is constructed. The inner product of each feature representation variable in the feature representation with each center vector in the center set is calculated. At the same time, the L2 norm of the feature representation vector and the center vector is calculated to obtain the similarity value. The similarity value is combined with the temperature coefficient and subjected to exponential normalization to obtain the probability value. All segments at the current scale are traversed to form a probability matrix. Then, the weighted result of the residual variance of the feature similarity neighborhood and the residual variance of the time neighborhood is calculated, and the weighted result of all segments is averaged to obtain the average residual variance value. Combined with a small constant, the residual variance value of each segment is exponentially operated with the average residual variance value to obtain the sample weights.

[0013] Preferably, the residual variance constraint loss is calculated as follows: Bernoulli sampling is performed on the correspondence between each segment and the center set based on the probability matrix, and a binary encoding matrix is ​​generated by combining Gumbel-Softmax sampling. The center set is input into the query mapping layer, and the feature representation is input into the key mapping layer and the value mapping layer to obtain the query representation, key representation, and value representation, respectively. Then, the query representation and key representation are processed by correlation calculation, scale normalization, and exponential operation. The correlation calculation result is multiplied at corresponding positions by the transpose of the binary encoding matrix and the transpose of the sample weights, respectively. After normalization, it is multiplied by the value representation to obtain the updated center set. Then, the residual variance constraint loss is constructed. The probability value is combined with the two-dimensional feature representation matrix after quadratic centering to obtain the residual centering mean. Then, the difference between the two-dimensional feature representation matrix after quadratic centering and the residual centering mean is calculated. The residual variance constraint loss is obtained by summing the probability value, the sum of all probability values, the small constant, and the number of center vectors of the updated center set.

[0014] Preferably, the relation transfer matrix and the previous probability matrix are obtained as follows: the probability matrix is ​​multiplied by the updated center set to obtain a weighted representation, and the weighted representation is input into the feature mapping layer and a bias parameter is added to obtain the feature matrix. The two-dimensional feature representation matrix after quadratic centering is added to the feature matrix at corresponding positions and layer normalization is performed to obtain the enhanced representation. Then, the updated center set and the adjacent updated center sets are input into the same cross-scale center mapping layer to obtain the current center mapping representation and the previous center mapping representation. The correlation between the two is calculated, scale is normalized, and row exponent is normalized to obtain the relation transfer matrix. The adjacent probability matrices are interpolated and adjusted so that the number of rows is consistent with the number of segments at the current scale. The previous probability matrix after interpolation is multiplied by the transpose of the relation transfer matrix to obtain the previous probability matrix after transformation to the current scale.

[0015] Preferably, the formation of the fusion feature representation specifically involves: subtracting the probability matrix from the previous probability matrix after transformation to the current scale at corresponding positions; calculating the squared L2 norm of the resulting difference vector; and performing exponential decay processing in conjunction with the difference adjustment parameter to obtain the cross-scale consistency value; multiplying the probability values ​​in the probability matrix with their logarithmic values ​​and summing the results; normalizing and numerically shifting the result according to the logarithmic value of the number of center vectors in the updated center set to obtain the concentration value; then forming a three-dimensional input vector from the sample weights, the cross-scale consistency value, and the concentration value; performing linear calculation with the fusion weight calculation parameter and the bias parameter; and processing the result through an activation function to obtain the fragment fusion weights; multiplying the previous probability matrix after transformation to the current scale with the updated center set to obtain a weighted representation; and obtaining an aligned representation through linear layer mapping; and then fusing the enhanced representation and the aligned representation according to the fragment fusion weights to obtain the fusion feature vector; and finally, traversing all fragments at the current scale to obtain the fusion feature representation.

[0016] Preferably, the analysis results of time-series anomalies at thermal power plant measurement points are compared with the anomaly determination threshold. The learnable parameters are updated based on the residual variance constraint loss. Specifically, the fused feature representation is input into a linear layer for linear mapping and processed by an activation function to obtain the scale anomaly probability. The scale with the most segments is selected as the benchmark scale. According to the time arrangement of each segment, the scale anomaly probabilities at other scales are interpolated and adjusted to the number of segments at the benchmark scale to obtain the anomaly probability. The anomaly probabilities at the same segment position at all scales are averaged to obtain the comprehensive anomaly probability. The comprehensive anomaly probability is compared with the anomaly determination threshold. When the comprehensive anomaly probability is greater than the anomaly determination threshold, the corresponding segment is marked as an anomaly segment; otherwise, it is marked as a normal segment, thus obtaining the analysis results of time-series anomalies at thermal power plant measurement points. In the subsequent model training process, the residual variance constraint loss at each scale is averaged to obtain the model training loss. The learnable parameters of the model are iteratively updated based on the model training loss.

[0017] The advantages of the technical effects provided by the present invention in the above technical solution are as follows:

[0018] The residual center module of this invention simultaneously characterizes the feature correlation and temporal continuity between segments through a feature similarity adjacency matrix and a feature-constrained temporal adjacency matrix. It adaptively generates residual weights based on the residual variances of feature similarity neighborhoods and temporal neighborhoods, which can reduce the interference of single neighborhood relationships and local noise on feature representation. By fusing the adjacency matrix, neighborhood weighting, and secondary centering, it can highlight abnormal changes that deviate from the neighborhood center. By combining the probability matrix, sample weights, and binary encoding matrix to update the center set, it uses residual variance constraint loss to improve the concentration of feature distribution under the same center, thereby enhancing the separability of normal and abnormal segments.

[0019] The scale fusion module of this invention generates an enhanced representation by using a probability matrix and an updated set of centers. It then utilizes cross-scale center mapping, relation transfer matrix, and probability matrix interpolation to transform the center relations and fragment probabilities of adjacent scales to the current scale. This reduces the information offset caused by differences in the number of fragments and center distribution at different scales. Based on sample weights, cross-scale consistency values, and concentration values, the module adaptively calculates fragment fusion weights, enabling the enhanced representation of the current scale to be dynamically fused with the aligned representation of the previous scale on a fragment-by-fragment basis. This preserves the local features of the current scale, introduces supplementary information from adjacent scales, and improves the stability, cross-scale consistency, and anomaly feature identification capabilities of the fused feature representation.

[0020] The anomaly detection module of this invention maps fused feature representations to scale anomaly probabilities and selects the scale with the most segments as the benchmark scale. It then interpolates and adjusts the scale anomaly probabilities at other scales to unify the number and time position of segments at different scales, avoiding the influence of scale differences on anomaly judgment. By averaging the anomaly probabilities at the same segment position across all scales, a comprehensive anomaly probability is obtained. This allows for the comprehensive utilization of anomaly information at different scales, reducing misjudgments caused by fluctuations at a single scale. The comprehensive anomaly probability is then compared with the anomaly judgment threshold to output abnormal and normal segments. Simultaneously, the model's learnable parameters are updated using residual variance constraint loss, which helps improve the stability and accuracy of time-series anomaly analysis results at thermal power plant measurement points. Attached Figure Description

[0021] Figure 1 This is a method for analyzing timing anomalies at measurement points in thermal power plants based on multimodal methods.

[0022] Figure 2 This is a diagram of the residual center module.

[0023] Figure 3 This is a diagram of the scale fusion module.

[0024] Figure 4 This is an error distribution diagram of the RDG-PACNet anomaly detection model.

[0025] Figure 5 The image shows the performance of the RDG-PACNet anomaly detection model.

[0026] Figure 6 Histogram of the distribution of comprehensive judgment indexes for the RDG-PACNet anomaly detection model. Detailed Implementation

[0027] This invention proposes a multimodal time-series anomaly analysis method for measuring points in thermal power plants, the steps of which are as follows: Figure 1As shown, for multimodal thermal power plant measurement data that varies over time, an RDG-PACNet anomaly detection model is proposed, consisting of a residual center module, a scale fusion module, and an anomaly detection module. The residual center module is used to fuse neighborhood residuals and update the center set; the scale fusion module completes cross-scale alignment and feature fusion; and the anomaly detection module summarizes the scale anomaly probabilities and determines the anomaly analysis results.

[0028] Collect operating data from individual thermal power units to construct monitoring data for thermal power plants.

[0029] Furthermore, relevant data were collected from monitoring points in the thermal power plant. Taking a single thermal power unit as the target, data was continuously collected for 60 days under a unified clock in the DCS, DEH, and TSI systems, covering start-up, shutdown, load increase, load decrease, steady-state operation, and abnormal equipment conditions. Main steam temperature data was collected using armored K-type thermocouples, with a range of 0–650℃, an accuracy of ±1.5℃, and a sampling frequency of 1Hz. Main steam pressure data was collected using intelligent pressure transmitters, with a range of 0–35MPa, an accuracy of ±0.1%FS, and a sampling frequency of 1Hz. Feedwater flow rate data was collected using differential pressure flow meters, with a range of 0–2500t / h, an accuracy of ±0.5%FS, and a sampling frequency of 1Hz. Fuel flow rate data was collected using electronic belt scales and coal feeder weighing sensors, with a range of 0–350t / h and an accuracy of ±0.5%FS. The turbine vibration amplitude data was collected using an eddy current displacement sensor with a range of 0–300 μm, an accuracy of ±1%FS, and an original sampling frequency of 2 kHz. Peak-to-peak values ​​were calculated in 1-second windows. Generator active power data was collected using a power transmitter with a range of 0–700 MW, an accuracy class of 0.2, and a sampling frequency of 1 Hz. Unit load rate data was calculated based on the ratio of generator active power to rated power, with a range of 0–110% and a resolution of 0.1%. All data were aligned at 1-second time intervals, with time synchronization errors controlled within 100 ms. Unit operating conditions, alarm status, and maintenance records were recorded synchronously as anomaly markers. A thermal power plant measurement point dataset was constructed and preprocessed, with the dataset divided in a 7:2:1 ratio.

[0030] After the thermal power plant measurement data is averaged and pooled at different scales, it is then processed by reversible instance normalization, divided into segments, and mapped to obtain the thermal power plant measurement data embedding.

[0031] Furthermore, input the measurement data from the thermal power plant. Average pooling at different scales is performed to generate time series data of thermal power plant measurement points at multiple scales. The time series data of thermal power plant monitoring points at each scale were divided into multiple segments through reversible instance normalization, among which... Map the segments to the number of time points contained in a single segment. In the feature space, the embedding of the measuring points in the thermal power plant is obtained. ,in, For the number of segments, Embed dimensions for fragments, This represents the number of measurement points and variables at the thermal power plant.

[0032] The mean and variance of the feature-constrained temporal adjacency matrix, the feature-similar adjacency matrix, the feature-similar neighborhood center residual matrix, and the temporal neighborhood center residual matrix are calculated and normalized to obtain the feature-similar neighborhood residual weights and the temporal neighborhood residual weights.

[0033] Furthermore, the input of thermal power plant measurement points is embedded... According to the fragment embedding dimension Number of measurement points and variables in thermal power plants The data is merged so that the embedded data of the thermal power plant measurement points corresponding to each segment is converted into flattened feature dimensions. One-dimensional feature vector Arranging the one-dimensional feature vectors in chronological order yields a two-dimensional feature representation matrix. ,in, To find the number of segments, calculate the feature similarity adjacency matrix, and then... The one-dimensional feature vector of the first segment is multiplied by the query feature mapping parameters, and the first segment... The one-dimensional feature vector of each segment is multiplied by the key feature mapping parameters, and then scaled to obtain the feature correlation value. Using a segment as a baseline, its correlation values ​​with all features at the current scale are exponentially normalized to obtain feature similarity adjacency weights, forming a feature similarity adjacency matrix. The specific mathematical model is as follows: ; In the formula, , For the respective The segment and the first One-dimensional feature vectors of each segment To query feature mapping parameters, For key feature mapping parameters, This is the scale normalization factor in the calculation of feature correlation. It is an exponential function. Number the segments. The first is the feature similarity adjacency matrix, followed by the feature-constrained time adjacency matrix, based on the first... The segment and the first The sequence number difference of the segments is used to calculate the absolute value of the temporal distance between them, and then processed based on the temporal distance decay parameter to calculate the... The one-dimensional feature vector of the first segment and the first segment The difference between the one-dimensional feature vectors of each segment is analyzed, and the results are processed based on the feature difference decay parameter and the flattened feature dimension. An exponential operation is then performed on the results to obtain the initial weights for temporal adjacency. Using a given segment as a baseline, calculate its initial temporal adjacency weights with all segments at the current scale, and normalize all initial temporal adjacency weights to obtain the feature-constrained temporal adjacency weights, forming a feature-constrained temporal adjacency matrix. The specific mathematical model is as follows: ; In the formula, For time-distance attenuation parameters, For characteristic difference attenuation parameter, The first is the feature-constrained temporal adjacency matrix. Then, the neighborhood center residuals are calculated based on the two adjacency matrices. After multiplying the feature-similar adjacency matrix with the two-dimensional feature representation matrix, the result of multiplying the two-dimensional feature representation matrix with the matrix is ​​subtracted from the result by corresponding positions to obtain the feature-similar neighborhood center residual matrix. The specific mathematical model is as follows: ; In the formula, The residual matrix of the feature similarity neighborhood center is obtained by multiplying the feature constraint temporal adjacency matrix and the two-dimensional feature representation matrix, and then subtracting the result of the matrix multiplication of the two-dimensional feature representation matrix from the matrix multiplication by corresponding positions. .

[0034] In this embodiment, the time distance attenuation parameter is set to a value range of 2 to 10, specifically 5; the feature difference attenuation parameter is set to a value range of 0.5 to 2.0, specifically 1.0; the query feature mapping parameter and the key feature mapping parameter are learnable parameters, independently initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02, and the random seed is set to 2026. When the time distance attenuation parameter is set to 5, the temporal adjacency weights of adjacent segments can be largely preserved, while the temporal relationships of segments more than 5 segments apart gradually attenuate, which can balance the local temporal continuity of the thermal power plant measurement data and reduce the interference of distant segments. When the feature difference attenuation parameter is set to 1.0, based on the flattening and normalization of the feature dimensions, segments with large feature differences can be moderately suppressed, while retaining the association between segments with similar feature states. This avoids the adjacency relationship being too sparse due to the parameter being too small or the different operating states being incorrectly associated due to the parameter being too large, thereby improving the stability of the feature-constrained temporal adjacency matrix and the ability to identify abnormal changes.

[0035] The mean and variance of the feature-constrained temporal adjacency matrix, the feature-similar adjacency matrix, the feature-similar neighborhood center residual matrix, and the temporal neighborhood center residual matrix are calculated and normalized to obtain the feature-similar neighborhood residual weights and the temporal neighborhood residual weights.

[0036] Furthermore, for feature-similar neighborhoods, the feature-similar adjacency matrix is... The In row weights, the first Each weight and feature-similar neighborhood center residual matrix The Perform residual vector multiplication and sum the results to obtain the mean residual of the feature-similar neighborhood. The specific mathematical model is as follows: ; In the formula, The mean of residuals in the similar neighborhood of features. For the number of segments, For the characteristic similarity adjacency matrix, the i-th... In row weights, the first Each weight, The first residual matrix of the characteristic similarity neighborhood center The residual vector is calculated, then the variance of the residuals in the feature similarity neighborhood is calculated, and the residual vector is subtracted from the mean of the residuals in the feature similarity neighborhood. The values ​​of each dimension in the subtraction result are squared and summed, and then combined with the flattened feature dimensions. The first adjacency matrix with similar features Row weight, the first row in the row The variance of the residuals in the feature similarity neighborhood is obtained by multiplying the weights and then summing them. The specific mathematical model is as follows: ; In the formula, The residual variance is used for the feature similarity neighborhood, and for the temporal neighborhood, the feature constrains the temporal adjacency matrix. The Middle In row weights, the first Each weight and the residual matrix of the time neighborhood center The Perform residual vector multiplication and sum the results to obtain the mean residual in the time neighborhood. The residual matrix at the center of the time neighborhood is subtracted from the mean of the time neighborhood residuals. The values ​​of each dimension in the subtraction result are squared and summed, and then combined with the flattened feature dimensions. The first adjacency matrix of the feature constraint time Row weight, the first row in the row The variance of the time neighborhood residuals is obtained by multiplying the weights together and then summing them. A learnable scale parameter is introduced to calculate the residual variances of the feature similarity neighborhood and the time neighborhood, respectively. The two calculation results are then negativeized and normalized to obtain the feature similarity neighborhood residual weights and the time neighborhood residual weights. The specific mathematical model is as follows: ; In the formula, For the residual weights of the similar neighborhood, For time neighborhood residual weights, , For learnable scale parameters, This is the exponential normalization function.

[0037] In this embodiment, the learnable scale parameter is independently initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02, and the random seed is set to 2026.

[0038] The residual weights of the feature similarity neighborhood and the time neighborhood are weighted based on the residual vectors in the residual matrix of the feature similarity neighborhood center and the residual matrix of the time neighborhood center, and the feature similarity adjacency matrix and the feature constraint time adjacency matrix are fused to calculate the two-dimensional feature representation matrix after quadratic centering.

[0039] Furthermore, the residual weights of feature-similar neighborhoods are... Residual matrix of neighborhood center with similar features The Middle Multiply the residual vectors corresponding to each segment, and simultaneously weight the residuals in the temporal neighborhood. residual matrix with time neighborhood center The Middle The residual vectors corresponding to each segment are multiplied, and the two multiplication results are summed to obtain the residual matrix. The specific mathematical model is as follows: ; In the formula, The residual matrix is... The th element in the residual matrix of the characteristic similar neighborhood center The residual vector corresponding to each segment The th in the residual matrix of the time neighborhood center The residual vectors corresponding to each segment are used to construct a fused adjacency matrix. Then, the residual weights of the feature-similar neighborhood and the temporal neighborhood are used to construct diagonal weight matrices according to the segment order. These matrices are then multiplied by the feature-similar adjacency matrix and the feature-constrained temporal adjacency matrix, respectively. The two multiplication results are summed, and each row of the resulting matrix is ​​normalized to obtain the fused adjacency matrix. The specific mathematical model is as follows: ; In the formula, To merge the adjacency matrix, For the construction of diagonal matrices, To perform row-wise normalization, the residual matrix is ​​weighted by neighborhood based on the fused adjacency matrix to obtain the fused residual neighborhood center matrix. A learnable centering parameter is then introduced into the activation function to obtain centering coefficients ranging from 0 to 1. The specific mathematical model is as follows: ; In the formula, For centralized learnable parameters, For activation function, The centering coefficients are multiplied by the fused residual neighborhood center matrix, and the residual matrix is ​​subtracted from the multiplication result at corresponding positions to obtain the two-dimensional feature representation matrix j after secondary centering. The specific mathematical model is as follows: ; In the formula, This is the two-dimensional feature representation matrix after quadratic centering. To merge the residual neighborhood center matrix.

[0040] In this embodiment, a normal distribution with a mean of 0 and a standard deviation of 0.02 is used for independent initialization, and the random seed is set to 2026.

[0041] The two-dimensional feature representation matrix after quadratic centering is mapped to a feature representation. The similarity value is calculated by combining the center set and forming a probability matrix, and then the sample weights are calculated.

[0042] Furthermore, the two-dimensional feature representation matrix after quadratic centering is... The input linear layer yields the feature representation. , construct the first Central set at each scale It includes There are central vectors with dimension . The first feature in the representation The characteristic variable is respectively related to the first feature variable. The inner product is calculated using the nth center vector, and the feature representation vector and the nth center vector are calculated simultaneously. The similarity value is obtained by calculating the L2 norm of each central vector and combining the inner product result with the L2 norm result. The specific mathematical model is as follows: ; In the formula, The similarity value is... For the first There are 1 central vector, which are trainable vectors. For the first Each feature represents a variable. The similarity value is combined with a temperature coefficient, and the resulting values ​​are exponentially normalized to obtain probability values. The specific mathematical model is as follows: ; In the formula, This is a probability value. Given the temperature coefficient, we iterate through all segments at the current scale to obtain the probability matrix. The weighted result of the residual variance of the feature similarity neighborhood and the residual variance of the time neighborhood is calculated, the two are added together, and the average of the sum is obtained to obtain the average residual variance value. The specific mathematical model is as follows: ; In the formula, For the number of segments, This represents the average residual variance. For the residual weights of the similar neighborhood, For time neighborhood residual weights, For the residual variance of the characteristic similar neighborhood, Let the variance of the residuals in the time neighborhood be the first... The residual variance and mean residual variance of each segment are calculated and then exponentially combined with a small constant to obtain the sample weights. The specific mathematical model is as follows: ; In the formula, For sample weights, It is a small constant.

[0043] In this embodiment, the temperature coefficient is set to a range of 0.05 to 0.50, specifically 0.10. A temperature coefficient of 0.10 appropriately amplifies the differences in similarity values ​​between different center vectors, allowing the probability matrix to more clearly reflect the correspondence between segments and center sets. This avoids excessive probability concentration and increased gradient fluctuations due to a too-small temperature coefficient, or an overly large temperature coefficient causing an overly flat probability distribution and decreased center discrimination ability. The small constant is set to a range of 10⁻. 8 ~10⁻ 4 Specifically set to 10⁻ 6 This value is relatively small compared to the average residual variance, so it will not significantly change the calculation result of the sample weights. At the same time, it can prevent the denominator from being 0 or the value from being excessively amplified when the average residual variance is close to 0, thereby improving the numerical stability of the sample weight calculation and model training process.

[0044] The probability matrix is ​​sampled to generate a binary encoding matrix, the center set is updated, the residual variance constraint loss is calculated, and the residual center module is constructed, such as... Figure 2 As shown.

[0045] Furthermore, based on the probability matrix For each segment and the central set The correspondence between them is subjected to Bernoulli sampling, and combined with Gumbel-Softmax sampling to generate a binary encoding matrix. The number of rows in the binary encoding matrix represents the number of segments at the current scale, and the number of columns represents the center vector of the center set. The quantity, the central set is input into the query mapping layer, and the feature representation is... Inputting the key mapping layer and value mapping layer yields the query representation, key representation, and value representation, respectively. The query representation and key representation are then correlated, and scaled and exponentially processed. The correlation results are then compared with the transpose of the binary encoding matrix and the sample weights, respectively. The transpose of the result is multiplied at corresponding positions, the result is normalized, and then multiplied with the value representation to obtain the updated central set. The specific mathematical model is as follows: ; In the formula, For the updated central set, To query the mapping parameters, For key mapping parameters, For value mapping parameters, The scale normalization factor in the calculation of center and fragment correlation. For query representation, As a key, Represented by value, For the normalization operation, a residual variance-constrained loss is constructed, and the probability values ​​are... The two-dimensional feature representation matrix after quadratic centering Multiply the results, sum the results, and combine the sum of all probability values ​​with the sum of the small constants to obtain the centralized mean of the residuals. The specific mathematical model is as follows: ; In the formula, The mean of the residuals is the centralized mean. The second feature representation matrix after quadratic centering is the first... The two-dimensional feature vector of the row, For the number of segments, The constant is used. Then, the difference between the two-dimensional feature representation matrix after quadratic centering and the mean of the residual centering is processed, and the difference results are summed. This summation is then multiplied by the probability values, and the sum of these multiplications is also summed. Finally, the sum of all probability values ​​is added to the constant, and this process is repeated for all updated center sets to obtain the residual variance constraint loss. The specific mathematical model is as follows: ; In the formula, For residual variance constraint loss, This represents the number of center vectors in the updated center set.

[0046] In this embodiment, the query mapping parameters, key mapping parameters, and value mapping parameters are learnable parameters, independently initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02, and the random seed is set to 2026; the range of the small constant is set to 10⁻. 8 ~10⁻ 4 Specifically set to 10⁻ 6 .

[0047] The enhanced representation is calculated using the probability matrix and the updated center set, and the relational transit matrix and the previous probability matrix are obtained through mapping and interpolation.

[0048] Furthermore, the probability matrix With the updated central set Multiplying yields a weighted representation, which is then input into the feature mapping layer and supplemented with bias parameters to obtain the feature matrix. The specific mathematical model is as follows: ; In the formula, The characteristic matrix, For feature mapping parameters, The bias parameter is the two-dimensional feature representation matrix after quadratic centering. The enhanced representation is obtained by adding it to the feature matrix at corresponding positions and then performing layer normalization. The specific mathematical model is as follows: ; In the formula, For layer normalization processing, To enhance the representation, the updated central set will be... With the neighboring updated center set Inputting the same cross-scale center mapping layer yields the current center mapping representation and the previous center mapping representation. The transpose of the current and previous center mapping representations is then used to calculate the correlation, and scale normalization is performed according to the cross-scale relationship calculation dimension. The normalized correlation results are then subjected to exponential normalization row-wise to obtain the relationship transitivity matrix. The specific mathematical model is as follows: ; In the formula, For relational transitive matrix, To perform index normalization on a row-by-row basis, For cross-scale center mapping parameters, This is the scale normalization factor used in calculating cross-scale central correlation. This represents the current central mapping. For the previous center mapping representation, the adjacent probability matrix is... Perform interpolation adjustments to make its row count match the number of segments at the current scale. To obtain the previous probability matrix after transformation to the current scale, multiply the interpolated and adjusted previous probability matrix by the transpose of the relation transfer matrix. .

[0049] In this embodiment, the feature mapping parameters and cross-scale center mapping parameters are learnable parameters, which are independently initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02. The random seed is set to 2026, and the initial value of the bias parameter is set to 0.

[0050] Alignment representations are calculated by comparing the probability matrix with the previous probability matrix, and then combined with the enhanced representation to form a fused feature representation. A scale fusion module is then constructed, such as... Figure 3 As shown.

[0051] Furthermore, the probability matrix The probability values ​​and the previous probability matrix after transformation to the current scale. The probability values ​​are subtracted at corresponding positions, and the square of the L2 norm is calculated for the resulting difference vector. Then, exponential decay is applied according to the difference adjustment parameter to obtain the cross-scale consistent value. The specific mathematical model is as follows: ; In the formula, This is a cross-scale consistent value. For difference adjustment parameters, The first probability matrix after transformation to the current scale is the... A probability value, The first in the probability matrix The probability value, the nth probability value in the probability matrix Each vector is multiplied by its logarithm, and the product results are summed. The summation is then calculated based on the number of center vectors in the updated center set. The numerical values ​​are normalized, and then the values ​​are shifted based on the normalized results to obtain the concentration values. The specific mathematical model is as follows: ; In the formula, This is the concentration value. The first in the probability matrix The probability value and the first An updated central set The center vector, The sample weights are small constants, and then the sample weights are... The cross-scale consistency value and concentration value form a three-dimensional input vector, which is linearly calculated with the fusion weight calculation parameters. This vector, combined with the bias parameters, is then processed by an activation function to obtain the fragment fusion weights. The specific mathematical model is as follows: ; In the formula, For fragment fusion weights, To integrate the weight calculation parameters, The bias parameter is the previous probability matrix after transformation to the current scale and the updated center set. Perform multiplication to obtain a weighted representation, and then obtain an aligned representation through linear layer mapping. This will enhance the representation The Middle The eigenvectors and alignment representations of the row The feature vector fusion segment fusion weights are calculated, and the two calculation results are added at corresponding positions to obtain the fused feature vector. The specific mathematical model is as follows: ; In the formula, To fuse feature vectors, To enhance the representation of the first The eigenvectors of the row, To align the representation of the first The feature vectors of the rows are used to traverse all segments at the current scale to obtain the fused feature representation. .

[0052] In this embodiment, the fusion weight calculation parameter is a learnable parameter, which is independently initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02. The random seed is set to 2026, and the initial value of the bias parameter is set to 0. The value range of the difference adjustment parameter is set to 0.5 to 2.0, specifically 1.0. Since the probability vectors in the probability matrix are all normalized, their L2 squared differences are usually between 0 and 2. When the difference adjustment parameter is 1.0, the cross-scale consistency value can be smoothly reduced from 1 to about 0.135 as the probability difference decreases. This can not only significantly suppress segments with large cross-scale differences, but also avoid losing effective information from adjacent scales due to excessively rapid decay.

[0053] The fused feature representation is mapped to the scale anomaly probability, and the comprehensive anomaly probability is calculated. The time series anomaly analysis results of the thermal power plant measurement points are compared with the anomaly judgment threshold. The learnable parameters are updated based on the residual variance constraint loss to construct an anomaly detection module.

[0054] Furthermore, an anomaly detection module is constructed, with the input fused feature representation. ,in, To flatten the feature dimensions, Given the number of fragments, the fused feature representation is linearly mapped into the linear layer and then processed by an activation function to obtain the scale anomaly probability. The scale with the most segments was selected as the baseline scale. Based on the temporal order of the segments, the scale anomaly probabilities at other scales were determined by the number of segments. Number of segments interpolated to the reference scale Obtain the abnormal probability The average anomaly probability at the same segment location across all scales is calculated to obtain the comprehensive anomaly probability. The specific mathematical model is as follows: ; In the formula, The total number at different scales. To synthesize the overall anomaly probability, the overall anomaly probability is compared with the anomaly determination threshold. By comparing the results, when the overall anomaly probability exceeds the anomaly detection threshold, the corresponding segment is marked as an anomalous segment; otherwise, it is marked as a normal segment. This yields the time-series anomaly analysis results for the thermal power plant's measurement points. During model training, residual variance is constrained at each scale. The mean is calculated to obtain the model training loss, and the learnable parameters of the model are iteratively updated based on the model training loss.

[0055] In this embodiment, the anomaly detection threshold is set to 0.50. When the anomaly detection threshold is set to 0.50, segments with an abnormal tendency higher than the normal tendency can be identified as abnormal segments. The judgment meaning is intuitive and will not be biased towards normal or abnormal segments. If the threshold is lower than 0.40, normal operation fluctuations are easily marked as abnormal segments, increasing false alarms. If the threshold is higher than 0.70, some minor anomalies or early abnormal segments may be identified as normal segments, increasing false negatives. Therefore, setting it to 0.50 can achieve a balance between anomaly detection sensitivity and judgment stability.

[0056] Furthermore, the RDG-PACNet anomaly detection model was written in Python, and the experiments were run on a Windows operating system. PyTorch was used as the framework in the CUDA 11.27 environment, and training was performed on a GeForce RTX 3090. The optimizer used was Adam, the initial learning rate was set to 0.001, the training batch size was set to 64, and the training period was set to 100. The dataset consisted of 60 days of data related to measurement points in thermal power plants, which was preprocessed and then input into the RDG-PACNet anomaly detection model.

[0057] Furthermore, the RDG-PACNet anomaly detection model error distribution diagram, anomaly detection model performance diagram, and anomaly detection model comprehensive judgment index distribution histogram are shown below. Figure 4 , Figure 5 , Figure 6As shown, the model can effectively track the normal trend of measurement data from thermal power plants and accurately identify abnormal segments that deviate from normal operating conditions. During normal operation, the fused measurement time series maintains a high degree of consistency with the normal operating reference trend, with only small-range random fluctuations, indicating that the model can effectively learn the temporal variation patterns of thermal power plant measurement data. When the measurement data shows significant deviations in time segments of approximately 70–80, 145–155, and 225–235, RDG-PACNet can identify the corresponding locations as outliers, and these outliers are concentrated in abrupt change regions, indicating that the model can effectively capture outliers by utilizing multi-scale feature fusion and residual center constraints. The local anomaly changes in the measured data improve the anomaly localization capability. The error distribution map shows that the model prediction error is generally centered around 0, with most errors concentrated in a small range, indicating that the RDG-PACNet output results have a high consistency with the normal operation trend. The histogram of the comprehensive judgment index distribution shows that the comprehensive anomaly probability covers the range of 0 to 1, and normal samples are mainly concentrated in the low probability area, while the comprehensive anomaly probability corresponding to abnormal samples is significantly higher, with some detection results close to 1, which can be clearly distinguished from the set anomaly judgment threshold of 0.50, indicating that the comprehensive anomaly probability generated by the model has a good anomaly discrimination capability.

Claims

1. A method for analyzing time-series anomalies at measurement points in thermal power plants based on multimodal methods, characterized in that, Includes the following steps: Collect operating data of a single thermal power unit to construct thermal power plant measurement point data. The thermal power plant measurement point data includes main steam temperature data, main steam pressure data, feedwater flow data, fuel flow data, turbine vibration amplitude data, generator active power data, and unit load rate data covering thermal power unit start-up, shutdown, load increase, load decrease, steady-state operation, and abnormal equipment conditions. After the thermal power plant measurement data is averaged and pooled at different scales, it is reversibly normalized, divided into segments, and mapped to obtain the thermal power plant measurement data embedding. The measurement points of the thermal power plant are embedded and flattened into a two-dimensional feature representation matrix. The feature constraint time adjacency matrix and the feature similarity adjacency matrix are calculated, and the feature similarity neighborhood center residual matrix and the time neighborhood center residual matrix are calculated. The mean and variance of the feature-constrained temporal adjacency matrix, the feature-similar adjacency matrix, the feature-similar neighborhood center residual matrix, and the temporal neighborhood center residual matrix are calculated and normalized to obtain the feature-similar neighborhood residual weight and the temporal neighborhood residual weight. The residual weights of the feature similarity neighborhood and the residual weights of the time neighborhood are weighted based on the residual vectors in the residual matrix of the feature similarity neighborhood center and the residual matrix of the time neighborhood center, and the feature similarity adjacency matrix and the feature constraint time adjacency matrix are fused to calculate the two-dimensional feature representation matrix after quadratic centering. The two-dimensional feature representation matrix after quadratic centering is mapped to a feature representation, and the similarity value is calculated by combining the center set. After forming a probability matrix, the sample weights are calculated. The probability matrix is ​​sampled to generate a binary encoding matrix, the center set is updated to calculate the residual variance constraint loss, and the residual center module is constructed. The enhanced representation is calculated using the probability matrix and the updated center set. The relation transfer matrix and the previous probability matrix are obtained through mapping and interpolation. The previous probability matrix is ​​the probability matrix corresponding to the previous scale adjacent to the current scale. The alignment representation is calculated by comparing the probability matrix with the previous probability matrix, and then combined with the enhanced representation to form a fused feature representation, thus constructing a scale fusion module. The fused feature representation is mapped to the scale anomaly probability, and the comprehensive anomaly probability is calculated. The time series anomaly analysis results of the thermal power plant measurement points are compared with the anomaly judgment threshold. The learnable parameters are updated based on the residual variance constraint loss to construct an anomaly detection module.

2. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 1, characterized in that, The process of obtaining the thermal power plant measurement point embedding is as follows: the thermal power plant measurement point data is subjected to average pooling at different scales to generate thermal power plant measurement point time series at multiple scales. The thermal power plant measurement point time series at each scale is divided into multiple segments through reversible instance normalization. Then, each segment is mapped to a specified feature space to obtain the thermal power plant measurement point embedding composed of the number of segments, the segment embedding dimension, and the number of thermal power plant measurement point variables.

3. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 2, characterized in that, The calculation of the feature-constrained temporal adjacency matrix, the feature-similar adjacency matrix, the feature-similar neighborhood center residual matrix, and the temporal neighborhood center residual matrix involves the following steps: The embedded data of the thermal power plant measurement points are merged according to the segment embedding dimension and the number of thermal power plant measurement point variables, converting the thermal power plant measurement point embedding corresponding to each segment into a one-dimensional feature vector. These vectors are then arranged in chronological order to obtain a two-dimensional feature representation matrix. The one-dimensional feature vector of each segment is multiplied by the query feature mapping parameter and the key feature mapping parameter, respectively. After scale normalization and exponential normalization, the feature-similar adjacency weights are obtained and formed into... A feature similarity adjacency matrix is ​​formed. The temporal position distance is calculated based on the sequence number difference between segments, and the difference between one-dimensional feature vectors is calculated. The initial weights of temporal adjacency are obtained by combining the temporal distance decay parameter, the feature difference decay parameter, and the flattened feature dimension. After normalization, a feature-constrained temporal adjacency matrix is ​​formed. The feature similarity adjacency matrix, the feature-constrained temporal adjacency matrix, and the two-dimensional feature representation matrix are multiplied respectively. The result of multiplying the two-dimensional feature representation matrix with the corresponding matrix is ​​subtracted according to the corresponding position to obtain the feature similarity neighborhood center residual matrix and the temporal neighborhood center residual matrix.

4. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 3, characterized in that, The residual weights of the feature similarity neighborhood and the time neighborhood are obtained by multiplying each weight in the feature similarity adjacency matrix with the corresponding residual vector in the feature similarity neighborhood center residual matrix and summing the results to obtain the mean of the feature similarity neighborhood residuals. Then, the difference between each residual vector and the mean of the feature similarity neighborhood residuals is calculated. This difference is then weighted and summed using the flattened feature dimension and the weights in the feature similarity adjacency matrix to obtain the variance of the feature similarity neighborhood residuals. Finally, the weights in the feature constraint time neighborhood matrix and the corresponding residual vectors in the time neighborhood center residual matrix are multiplied and summed to obtain the variance of the feature similarity neighborhood residuals. The residual vectors are multiplied and summed to obtain the mean of the time neighborhood residuals. Then, the differences between each residual vector and the mean of the time neighborhood residuals are calculated. The weighted sum is then performed by combining the flattened feature dimensions and the weights in the feature constraint time adjacency matrix to obtain the variance of the time neighborhood residuals. The variances of the feature similarity neighborhood residuals and the time neighborhood residuals are processed separately by combining the learnable scale parameter. The processed variances of the feature similarity neighborhood residuals and the time neighborhood residuals are negativeized and normalized to obtain the weights of the feature similarity neighborhood residuals and the time neighborhood residuals.

5. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 4, characterized in that, The calculation of the two-dimensional feature representation matrix after quadratic centering is specifically as follows: The residual weights of the feature similarity neighborhood are multiplied by the residual vectors of the corresponding segments in the feature similarity neighborhood center residual matrix; simultaneously, the residual weights of the time neighborhood are multiplied by the residual vectors of the corresponding segments in the time neighborhood center residual matrix; and the results of multiplying the residual weights of the time neighborhood and the residual vectors of the corresponding segments in the time neighborhood center residual matrix are added together to obtain the residual matrix. The residual weights of the feature similarity neighborhood and the time neighborhood are then used to construct pairs according to the segment order. The diagonal weight matrix is ​​obtained by multiplying it by the feature similarity adjacency matrix and the feature constraint temporal adjacency matrix, respectively, and then summing the results. Row-wise normalization is then performed to obtain the fused adjacency matrix. Based on the fused adjacency matrix, the residual matrix is ​​subjected to neighborhood weighting to obtain the fused residual neighborhood center matrix. The centering learnable parameters are input into the activation function to obtain the centering coefficients. The fused residual neighborhood center matrix is ​​then multiplied by the centering coefficients, and the residual matrix is ​​subtracted from the multiplication result at corresponding positions to obtain the two-dimensional feature representation matrix after secondary centering.

6. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 5, characterized in that, The sample weights are calculated as follows: the two-dimensional feature representation matrix after quadratic centering is input into the linear layer to obtain the feature representation, and a center set containing multiple center vectors at the current scale is constructed. The inner product of each feature representation variable in the feature representation with each center vector in the center set is calculated. At the same time, the L2 norm of the feature representation vector and the center vector is calculated to obtain the similarity value. The similarity value is combined with the temperature coefficient and subjected to exponential normalization to obtain the probability value. All segments at the current scale are traversed to form a probability matrix. Then, the weighted result of the residual variance of the feature similarity neighborhood and the residual variance of the time neighborhood is calculated, and the weighted result of all segments is averaged to obtain the average residual variance value. Combined with a small constant, the residual variance value of each segment is exponentially operated with the average residual variance value to obtain the sample weights.

7. The method for analyzing time-series anomalies at measurement points in thermal power plants based on multimodal methods according to claim 6, characterized in that, The residual variance constraint loss is calculated as follows: Bernoulli sampling is performed on the correspondence between each segment and the center set based on the probability matrix, and Gumbel-Softmax sampling is combined to generate a binary encoding matrix. The center set is input into the query mapping layer, and the feature representation is input into the key mapping layer and the value mapping layer to obtain the query representation, key representation, and value representation, respectively. Then, the query representation and key representation are processed by correlation calculation, scale normalization, and exponential operation. The correlation calculation result is multiplied at corresponding positions by the transpose of the binary encoding matrix and the transpose of the sample weights, respectively. After normalization, it is multiplied by the value representation to obtain the updated center set. Then, the residual variance constraint loss is constructed. The probability value is combined with the two-dimensional feature representation matrix after quadratic centering to obtain the residual centering mean. The difference between the two-dimensional feature representation matrix after quadratic centering and the residual centering mean is calculated. The residual variance constraint loss is obtained by summing the probability value, the sum of all probability values, the small constant, and the number of center vectors of the updated center set.

8. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 7, characterized in that, The relation transfer matrix and the previous probability matrix are obtained as follows: the probability matrix is ​​multiplied by the updated center set to obtain a weighted representation, and the weighted representation is input into the feature mapping layer and a bias parameter is added to obtain the feature matrix. The two-dimensional feature representation matrix after quadratic centering is added to the feature matrix at corresponding positions and layer normalization is performed to obtain the enhanced representation. Then, the updated center set and the adjacent updated center sets are input into the same cross-scale center mapping layer to obtain the current center mapping representation and the previous center mapping representation. The correlation of the two is calculated, scale normalized, and row exponent normalized to obtain the relation transfer matrix. The adjacent probability matrices are interpolated to make their number of rows consistent with the number of segments at the current scale. The interpolated previous probability matrix is ​​multiplied by the transpose of the relation transfer matrix to obtain the previous probability matrix after transformation to the current scale.

9. The method for analyzing time-series anomalies at measuring points in thermal power plants based on multimodal methods according to claim 8, characterized in that, The formation of the fusion feature representation is specifically as follows: the probability matrix is ​​subtracted from the previous probability matrix after transformation to the current scale at corresponding positions, the square of the second norm of the resulting difference vector is calculated, and exponential decay processing is performed in combination with the difference adjustment parameter to obtain the cross-scale consistency value. The probability values ​​in the probability matrix are multiplied by their logarithms and summed. The logarithm of the number of center vectors in the updated center set is normalized and numerically shifted to obtain the concentration value. Then, the sample weights, cross-scale consistency value, and concentration value are combined to form a three-dimensional input vector, which is linearly calculated with the fusion weight calculation parameter and bias parameter and processed by the activation function to obtain the fragment fusion weight. The previous probability matrix after transformation to the current scale is multiplied by the updated center set to obtain the weighted representation, and the alignment representation is obtained through linear layer mapping. Then, the enhanced representation and the alignment representation are fused according to the fragment fusion weight to obtain the fusion feature vector. The fusion feature representation is obtained by traversing all fragments at the current scale.

10. A method for analyzing time-series anomalies at measurement points in a thermal power plant based on multimodal methods, as described in claim 9, is characterized in that... The analysis results of time-series anomalies at thermal power plant measurement points are compared with the anomaly detection threshold. The learnable parameters are updated based on the residual variance constraint loss. Specifically, the fused feature representation is input into a linear layer for linear mapping and processed by an activation function to obtain the scale anomaly probability. The scale with the most segments is selected as the baseline scale. According to the time arrangement of each segment, the scale anomaly probabilities at other scales are interpolated and adjusted to the number of segments at the baseline scale to obtain the anomaly probability. The anomaly probabilities at the same segment position at all scales are averaged to obtain the comprehensive anomaly probability. The comprehensive anomaly probability is compared with the anomaly detection threshold. When the comprehensive anomaly probability is greater than the anomaly detection threshold, the corresponding segment is marked as an anomalous segment; otherwise, it is marked as a normal segment, thus obtaining the time-series anomaly analysis results at thermal power plant measurement points. In the subsequent model training process, the residual variance constraint loss at each scale is averaged to obtain the model training loss. The learnable parameters of the model are iteratively updated based on the model training loss.

Images