A method for identifying abnormal hydropower station safety monitoring data

Abstract

The application discloses a kind of hydropower station safety monitoring data anomaly identification method, comprising the following steps: S1, the monitoring signal of multiclass sensor acquisition is standardized, and standardization monitoring signal is obtained;S2, EMD mode decomposition is executed to standardization monitoring signal, and multiple order mode components are obtained.The application adopts EMD mode decomposition to standardization monitoring signal and carries out hierarchical processing, separates IMF component and residual trend remainder component, adapts the characteristics of hydropower station monitoring signal nonlinearity, non-stationary.Meanwhile according to local standard deviation calculation weight, the mode component of the same type different sensor is weighted and fused, can automatically reduce the weight of noise, local abnormal interference measuring point, strengthens the effect of effective monitoring data, both fully excavate the time sequence change rule of single group signal, and utilize the space and physical association between multiple measuring points, effectively suppress random noise, reduce single point data misjudgment problem.

Description

Technical Field

[0001] This invention relates to the field of hydropower station safety monitoring technology, specifically to a method for identifying anomalies in hydropower station safety monitoring data. Background Technology

[0002] Hydropower stations are important infrastructure in the water conservancy project system. The operation status of their core structures, such as dams, slopes, and powerhouses, directly affects the safety of the overall project. In order to achieve all-weather safety management, various monitoring sensors such as water level gauges, strain gauges, vibration sensors, and flow meters are usually deployed on site to continuously collect operational data such as structural deformation, vibration, water level, and flow rate, and to complete the safety status assessment based on the monitoring data.

[0003] In existing technologies, the industry generally adopts a processing mode that combines fixed statistical threshold method with basic signal analysis for the identification of anomalies in hydropower station safety monitoring data. This type of method first performs noise reduction and standardization preprocessing on the raw monitoring data, and then sets a fixed alarm threshold for a single measuring point based on engineering experience or normal distribution law, such as the three-standard-deviation discrimination criterion. When the real-time monitoring data of a single measuring point continuously exceeds the preset threshold, the monitoring data is judged to be abnormal and an early warning is triggered.

[0004] However, the operating environment of hydropower stations is complex, and the monitoring signals collected by various sensors are nonlinear and non-stationary signals formed by multi-physics field coupling. There are spatial correlations and physical coupling characteristics between the monitoring data of different locations and different types of sensors. The existing fixed threshold discrimination method only judges the signal amplitude of a single measuring point from exceeding the limit, which has obvious technical defects. First, this method does not perform layered analysis on the original signal, making it impossible to distinguish between high-frequency fluctuation components and low-frequency trend components in the signal. It is difficult to uncover the inherent temporal variation patterns of the monitoring data and has insufficient ability to identify problems such as short-term sudden anomalies and long-term trend drift. Second, this method analyzes data from a single sensor in isolation, ignoring the collaborative correlation characteristics between sensors of the same type and different measuring points. It is prone to misjudgment due to single-point noise and local interference, and it also cannot identify hidden anomalies caused by multi-point linkage. Third, traditional signal processing methods have a single feature extraction dimension, relying solely on the amplitude of raw data for analysis, failing to fully explore the multi-scale features of the signal, ultimately resulting in a low overall anomaly identification accuracy, which is difficult to meet the high-precision and high-reliability safety monitoring requirements of hydropower stations. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a method for identifying anomalies in hydropower station safety monitoring data, which solves the problem of low accuracy in anomaly identification in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for identifying anomalies in safety monitoring data of hydropower stations, comprising the following steps: S1. Standardize the monitoring signals collected by multiple types of sensors to obtain standardized monitoring signals; S2. Perform EMD mode decomposition on the standardized monitoring signal to obtain multi-order modal components; perform weighted processing on the modal components of different sensors of the same type to obtain multi-order fused modal components; the multi-order fused modal components include multi-order IMF fused components and residual trend term fused components; S3. Extract IMF feature vectors from each sub-segment of each IMF fusion component, and superimpose the IMF feature vectors corresponding to the same type of sensor to construct an IMF aggregated feature matrix. S4. Extract trend feature vectors from each sub-segment of the residual trend term fusion component and construct a trend feature matrix; S5. Concatenate the IMF aggregated feature matrix and the trend feature matrix corresponding to the same type of sensor to obtain the full-dimensional feature matrix; S6. An anomaly recognition neural network is used to process the full-dimensional feature matrices of various sensors and output the anomaly diagnosis results of hydropower station safety monitoring.

[0007] Further, step S2 includes the following sub-steps: S21. Perform EMD mode decomposition on the standardized monitoring signal to obtain multi-order modal components; the multi-order modal components include multi-order IMF components and residual trend component; S22. Taking each time point of each modal component as the center, extract the central neighborhood range and calculate the standard deviation to obtain the local standard deviation corresponding to each time point. S23. Calculate the weighted weights for each time point based on the local standard deviations at each time point; S24. Based on the weighting weights, perform weighted operations on the same-order modal components of different sensors of the same type to obtain the fused modal components of the corresponding order.

[0008] Further, the specific process of step S23 is as follows: using an exponential function to map the local standard deviation of each modal component and each time step in the standardized monitoring signal of each sensor to the component stability; summing the component stability of all sensors of the same type, modal components of the same order, and corresponding components at the same time step to obtain the total component stability; using the ratio of the component stability at a single time step to the total component stability as the weighting weight at that time step; the specific process of step S24 is as follows: multiplying the weighting weight of each modal component of a single sensor at each time step with the amplitude of the modal component of the same order at the corresponding time step, and then accumulating the multiplication results of all sensors of the same type to obtain the fused modal component of the corresponding order.

[0009] Further, step S3 includes the following sub-steps: S31. Divide each IMF fusion component into several IMF component segments; S32. For each IMF component segment, extract three types of feature elements: local energy, peak value, and maximum value, and construct the IMF feature vector. S33. Superimpose the IMF feature vectors corresponding to the same type of sensor, the same order of IMF fusion component, and the same IMF component sub-segment to obtain the IMF aggregated feature vector of the sub-segment. S34. Arrange all IMF aggregated feature vectors corresponding to the same type of sensor in chronological order to generate an IMF aggregated feature matrix.

[0010] Furthermore, the method for obtaining each feature parameter in step S32 is as follows: Within a single IMF component segment, the instantaneous energy is obtained by squaring the amplitude at each time step, and the local energy of the segment is obtained by summing all the instantaneous energies. The average energy is obtained by taking the average of all instantaneous energies within the IMF component segment, and the ratio of the maximum instantaneous energy within the segment to the average energy is taken as the peak value. The maximum instantaneous energy within the IMF component segment is taken as the maximum value.

[0011] Further, step S4 includes the following sub-steps: S41. Divide the residual trend term fusion component into several trend component segments; S42. Squaring the amplitude at each time step within each trend component segment yields the instantaneous energy at the corresponding time step. S43. Accumulate all instantaneous energies within a single trend component segment to obtain the local energy of that segment; S44. Calculate the local mean of instantaneous energy within a single trend component segment and the global mean of all instantaneous energy of the residual trend term fusion component, respectively. S45. Subtract the local mean from the local energy of the trend component segment to obtain the local deviation. S46. Subtract the global mean from the local energy of the trend component segment to obtain the global deviation. S47. Construct a trend feature vector using the local energy, local deviation, and global deviation of the same trend component segments as feature elements. S48. Arrange all trend feature vectors corresponding to the same type of sensor in chronological order to generate a trend feature matrix.

[0012] Furthermore, the processing procedure of the anomaly detection neural network in step S6 includes the following sub-steps: S61. The feature compression enhancement unit processes the full-dimensional feature matrix corresponding to each type of sensor to obtain compressed features; S62. The joint weight generation unit is used to calculate all compressed features to obtain the joint weight; S63. The joint weights are multiplied element-wise by the multiplier to obtain the joint enhanced compression features; S64. Multi-scale residual units are used to process each joint enhancement compression feature separately to obtain multi-scale fused enhancement features; S65. All multi-scale fusion enhancement features are added element-wise by adder A1 to obtain multi-type fusion features; S66. The identification output unit processes multiple types of fused features and outputs the abnormal diagnosis results of hydropower station safety monitoring.

[0013] Furthermore, the feature compression enhancement unit includes a first convolutional layer, an average pooling layer, a max pooling layer, an adder A2, and a second convolutional layer; The input of the first convolutional layer is the input of the feature compression enhancement unit, and the output of the first convolutional layer is connected to the input of the average pooling layer and the input of the max pooling layer, respectively. The input terminals of adder A2 are connected to the output terminals of the average pooling layer and the max pooling layer, respectively, and the output terminal of adder A2 is connected to the input terminal of the second convolutional layer. The output of the second convolutional layer is the output of the feature compression enhancement unit.

[0014] Furthermore, the operational expression of the joint weight generation unit is: , In the formula, The weights are joint weights, and sigmoid is the activation function. For 1×1 convolution, This is the first compression feature. For the first One compression feature, For the first One compression feature, It is a positive integer. This is the number of the compression feature.

[0015] Furthermore, the multi-scale residual unit includes a first-scale convolutional layer, a second-scale convolutional layer, a third-scale convolutional layer, a Concat layer, a third convolutional layer, an adder A3, and a fourth convolutional layer; The input terminals of the first-scale convolutional layer, the second-scale convolutional layer, the third-scale convolutional layer, and the first input terminal of adder A3 are interconnected and serve as the input terminals of the multi-scale residual unit. The input of the Concat layer is connected to the output of the first-scale convolutional layer, the second-scale convolutional layer, and the third-scale convolutional layer, respectively, and the output of the Concat layer is connected to the input of the third convolutional layer. The second input of adder A3 is connected to the output of the third convolutional layer, and the output of adder A3 is connected to the input of the fourth convolutional layer. The output of the fourth convolutional layer is the output of the multi-scale residual unit.

[0016] The beneficial effects of this invention are as follows: 1. This invention employs EMD (Electronic Mode Decomposition) to perform hierarchical processing on standardized monitoring signals, separating IMF (Integrated Motion Flow) components and residual trend components, thus adapting to the nonlinear and non-stationary characteristics of hydropower station monitoring signals. Simultaneously, based on local standard deviations, weights are calculated to perform weighted fusion of modal components from different sensors of the same type. This automatically reduces the weight of measurement points affected by noise and local anomalies, enhancing the effectiveness of monitoring data. It not only fully explores the temporal variation patterns of individual signals but also utilizes the spatial and physical relationships between multiple measurement points to effectively suppress random noise and reduce misjudgments of single-point data.

[0017] 2. This invention extracts differentiated features from the IMF fusion component and the residual trend remainder fusion component, respectively, and constructs an IMF aggregated feature matrix and a trend feature matrix, which are then concatenated into a full-dimensional feature matrix. On the one hand, it captures short-term high-frequency fluctuation features of the signal through parameters such as local energy, peak value, and maximum value to identify sudden anomalies; on the other hand, it uses indicators such as local deviation and global deviation to characterize long-term trend changes of the signal and monitor slow drift-type hidden dangers, achieving multi-dimensional, all-time feature extraction and fully covering different types of anomaly manifestations in hydropower stations.

[0018] 3. This invention constructs a dedicated anomaly detection neural network, relying on feature compression enhancement units, joint weight generation units, and multi-scale residual units to complete feature optimization and deep fusion. The network can adaptively allocate feature weights from multiple types of sensors, capturing different detailed features through a multi-scale structure, while the residual structure ensures the training effect of the deep network. Compared with traditional threshold discrimination methods, this solution significantly improves the sensitivity and reliability of anomaly detection, enabling accurate comprehensive analysis of multi-source monitoring data and meeting the actual needs of high-precision safety monitoring in hydropower stations. Attached Figure Description

[0019] Figure 1 A flowchart of a method for identifying anomalies in safety monitoring data of a hydropower station; Figure 2 This is a schematic diagram of the structure of an anomaly detection neural network; Figure 3 This is a schematic diagram of the feature compression enhancement unit. Figure 4This is a schematic diagram of the structure of a multi-scale residual unit. Detailed Implementation

[0020] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0021] like Figure 1 As shown, a method for identifying anomalies in safety monitoring data of a hydropower station includes the following steps: S1. Standardize the monitoring signals collected by multiple types of sensors to obtain standardized monitoring signals; S2. Perform EMD mode decomposition on the standardized monitoring signal to obtain multi-order modal components; perform weighted processing on the modal components of different sensors of the same type to obtain multi-order fused modal components; the multi-order fused modal components include multi-order IMF fused components and residual trend term fused components; S3. Extract IMF feature vectors from each sub-segment of each IMF fusion component, and superimpose the IMF feature vectors corresponding to the same type of sensor to construct an IMF aggregated feature matrix. S4. Extract trend feature vectors from each sub-segment of the residual trend term fusion component and construct a trend feature matrix; S5. Concatenate the IMF aggregated feature matrix and the trend feature matrix corresponding to the same type of sensor to obtain the full-dimensional feature matrix; S6. An anomaly recognition neural network is used to process the full-dimensional feature matrices of various sensors and output the anomaly diagnosis results of hydropower station safety monitoring.

[0022] This embodiment uses four types of sensors: strain gauges, vibration sensors, water level gauges, and flow meters, to monitor dam strain, equipment vibration, water level changes, and fluid flow, respectively. Depending on the actual operating conditions and monitoring needs of the hydropower station, additional equipment such as piezometers, multi-point displacement gauges, weir gauges, and strong-motion meters can be added for monitoring indicators such as dam foundation uplift pressure, slope deformation, seepage, and seismic response, adapting to different engineering geological conditions and safety control requirements.

[0023] In this embodiment, step S1 uses Z-score normalization to complete the monitoring signal preprocessing.

[0024] In this embodiment, S2 includes the following sub-steps: S21. Perform EMD mode decomposition on the standardized monitoring signal to obtain multi-order modal components, which are divided into multi-order IMF components and residual trend component; S22. Extract a neighborhood centered on the data points of each modal component at each time point and calculate the local standard deviation; S23. Calculate the weighted weights for each time step based on the local standard deviation; S24. Combining the weighted weights, perform weighted calculations on the same-order modal components of different sensors of the same type to obtain the fused modal components.

[0025] In this embodiment, the monitoring signal duration is 1 hour, the sampling frequency is 1 time / 10 seconds, and there are a total of 360 sampling points for a single signal; 31 sampling points are selected for the neighborhood window to calculate the local standard deviation.

[0026] In this invention, except for the last order which is the residual trend remainder component, all others are IMF components; correspondingly, the IMF components of the same order of the same type of sensor are weighted to obtain the IMF fusion component, and the residual trend remainder components are weighted to obtain the residual trend remainder fusion component.

[0027] In this embodiment, in step S23, the local standard deviation is mapped to component stability using an exponential function, and then the weighted weights are solved by normalization. The calculation formula is as follows: , In the formula, For the first The first sensor standardized monitoring signal The first modal component Weighted weights at each moment For the first The first sensor standardized monitoring signal The first modal component Local standard deviation at time, For the first The first sensor standardized monitoring signal The first modal component Local standard deviation at time, It is a natural constant. For the sensor number, Number the modal component order. The time number, For cumulative variables, The number of sensors of the same type For component stability.

[0028] In this embodiment, in step S24, modal component fusion is performed based on the weights, and the calculation formula is as follows: , In the formula, For the first The first fusion modal component Amplitude at time, For the first The first sensor standardized monitoring signal The first modal component Amplitude at time, For the first The first sensor standardized monitoring signal The first modal component Weighted weights at each moment.

[0029] The above weighting method can achieve adaptive adjustment: the greater the fluctuation of the sensor signal and the higher the local standard deviation, the lower the component stability and corresponding weight, effectively suppressing the interference caused by noise and local anomalies; the sensor with stable signal has a higher weight, and finally a fusion component with better signal-to-noise ratio is obtained, reducing false alarms and false alarms caused by single-point monitoring.

[0030] In this embodiment, S3 includes the following sub-steps: S31. Divide each order of IMF fusion component into several IMF component segments; S32. Extract local energy, peak value, and maximum value segment by segment and combine them into an IMF feature vector; S33. Fuse the IMF feature vectors of the same sub-segment of sensors of the same type to obtain the IMF aggregated feature vector; S34. Arrange all aggregated feature vectors in chronological order to generate the IMF aggregated feature matrix.

[0031] In this embodiment, the duration of a single IMF component segment is 3 minutes, and a single IMF fusion component is divided into 20 segments.

[0032] The calculation methods for each feature are as follows: Local energy: The instantaneous energy is obtained by squaring all the amplitudes within the sub-segment, and the sum of all the instantaneous energies is the local energy of the sub-segment; Peak value: The average instantaneous energy of a segment is calculated, and the ratio of the maximum instantaneous energy of the segment to the average value is taken as the peak value; Maximum value: The maximum instantaneous energy within the sub-segment.

[0033] The expression for the IMF eigenvector is: In the formula, For the first of its kind of sensor The first IMF fusion component IMF feature vectors of each IMF component segment For the first of its kind of sensor The first IMF fusion component The first feature element in the IMF feature vector of each IMF component segment. For the first of its kind of sensor The first IMF fusion component The second feature element in the IMF feature vector of each IMF component segment. For the first of its kind of sensor The first IMF fusion component The third feature element in the IMF feature vector of each IMF component segment. For the sub-segment number, This is for the transpose operation.

[0034] IMF aggregated feature vector expression: In the formula, For the first of its kind of sensor IMF aggregated feature vector of each IMF component segment This represents the number of IMF fusion components corresponding to the same type of sensor.

[0035] This invention achieves cross-scale feature fusion by superimposing multi-level features, comprehensively covering signal fluctuation features at different frequencies and avoiding the loss of features at a single scale.

[0036] In this embodiment, a 3-row × 20-column IMF aggregated feature matrix is ​​finally obtained.

[0037] In this embodiment, S4 includes the following sub-steps: S41. Divide the residual trend component into several trend component segments, and the segmentation rules are consistent with the IMF fusion components. S42. Squaring the amplitude within each segment yields the instantaneous energy at each time step. S43. Accumulate all instantaneous energies within the sub-segment to obtain the local energy; S44. Calculate the local mean of the instantaneous energy of a single sub-segment and the global mean of the fused component of the residual trend term of the entire segment, respectively. S45. Subtract the local mean from the local energy to obtain the local deviation. S46. Subtract the global mean from the local energy to obtain the global deviation. S47. Construct a trend feature vector using local energy, local deviation, and global deviation. S48. Arrange all trend feature vectors in chronological order to generate a trend feature matrix.

[0038] Trend feature vector expression: In the formula, For the residual trend remainder of the same type of sensor, the fusion component is the first... Trend feature vector of each trend component segment For the residual trend remainder of the same type of sensor, the fusion component is the first... The first feature element in the trend feature vector of each trend component segment. For the residual trend remainder of the same type of sensor, the fusion component is the first... The second feature element in the trend feature vector of each trend component segment. For the residual trend remainder of the same type of sensor, the fusion component is the first... The third feature element in the trend feature vector of each trend component segment.

[0039] The method for dividing the residual trend component in S41 is consistent with the method for dividing it in S31.

[0040] The local deviation of this invention can reflect the fluctuation deviation of the signal in a short period of time, while the global deviation is used to determine whether the overall trend of the signal exceeds the normal range. The combination of the two can accurately identify slow drift and continuous abnormality.

[0041] In this embodiment, the trend feature matrix has a size of 3 rows × 20 columns.

[0042] In this embodiment, the IMF aggregated feature matrix and the trend feature matrix of the same type of sensor are concatenated to obtain a full-dimensional feature matrix. In this embodiment, the matrix size is 6 rows × 20 columns.

[0043] like Figure 2 As shown, the anomaly detection neural network consists of a feature compression and enhancement unit, a joint weight generation unit, a multiplier M, a multi-scale residual unit, an adder A1, and a detection output unit. In this embodiment, four types of sensors were selected, corresponding to a unit quantity L=4.

[0044] In this embodiment, the processing procedure of the anomaly detection neural network in S6 includes: S61. The full-dimensional feature matrix of each type of sensor is input into a feature compression and enhancement unit, and the compressed features are output. S62. The joint weight generation unit integrates all compressed features and calculates the joint weight; S63. The multiplier multiplies the joint weights element by element with each compression feature to obtain the joint enhanced compression feature; S64. Multi-scale residual units process the joint enhancement compression features and output multi-scale fused enhancement features. S65. Adder A1 adds all multi-scale fusion enhancement features element by element to obtain multi-type fusion features; S66. The identification output unit analyzes the multi-type fusion features and outputs the final anomaly diagnosis result.

[0045] This invention first processes the full-dimensional feature matrix of each type of sensor through a feature compression enhancement unit, achieving feature compression while retaining core features, reducing redundant information interference, and improving feature processing efficiency. Then, a joint weight generation unit performs collaborative analysis on the compressed features of each type of sensor to generate joint weights, highlighting the importance of features from different sensor types. Subsequently, a multiplier M multiplies the joint weights element-wise with the corresponding compressed features to construct jointly enhanced compressed features, strengthening effective features and suppressing useless features. The application of a multi-scale residual unit accurately captures feature information at different scales, avoiding feature gradient vanishing, improving the depth and completeness of feature extraction, and obtaining multi-scale fusion enhanced features. An adder A1 adds the multi-scale fusion enhanced features of each type of sensor element-wise, achieving deep fusion of features from multiple sensor types, fully exploring the potential correlations between different sensor types, and breaking the limitations of single-type sensor feature analysis. Finally, a recognition output unit processes the multi-type fusion features to achieve accurate output of anomaly diagnosis results.

[0046] like Figure 3 As shown, the feature compression and enhancement unit in S61 consists of a first convolutional layer, an average pooling layer, a max pooling layer, an adder A2, and a second convolutional layer. Data flow: The input signal is processed by the first convolutional layer to extract shallow features, which are then fed into the average pooling layer and the max pooling layer respectively. The two pooling results are fused by the adder A2, and then processed by the second convolutional layer to complete secondary feature extraction and output.

[0047] In this embodiment, the first convolutional kernel is 3×1 with 16 output channels; the average pooling and max pooling windows are both 2×1 with a stride of 2; the second convolutional kernel is 3×1 with 16 output channels; and the unit output feature size is 6×10×16. This structure simultaneously preserves mean and extreme value features, taking into account both the overall state and local anomaly details.

[0048] This invention first extracts shallow effective features from the full-dimensional feature matrix through the first convolutional layer, then extracts global mean features through the average pooling layer and captures local extreme value anomalies through the max pooling layer. The two pooling results are fused and complemented by the adder A2, and finally the second convolutional layer completes the secondary feature extraction, effectively preserving the details of fluctuations and anomalies in the monitoring data.

[0049] In this embodiment, the expression for the joint weight generation unit in S62 is: , In the formula, The weights are joint weights, and sigmoid is the activation function. For 1×1 convolution, This is the first compression feature. For the first One compression feature, For the first One compression feature, It is a positive integer. This equals the number of compressed features, the number of sensor types, and the number of input full-dimensional feature matrices. The number of the compression feature. To add L compressed features element by element.

[0050] This unit adaptively assigns weights to features from different sensors through feature aggregation and activation functions, strengthening effective features and weakening interfering features, thereby achieving dynamic allocation of multi-source features.

[0051] like Figure 4 As shown, the multi-scale residual unit in S64 includes: a first-scale convolutional layer, a second-scale convolutional layer, a third-scale convolutional layer, a Concat layer, a third convolutional layer, an adder A3, and a fourth convolutional layer. Data flow: The input signal is simultaneously fed into three convolutional layers of different scales. The convolution results are concatenated by the Concat layer, dimensionality reduced by the third convolutional layer, and then added to the original input as residuals. Finally, the result is output by the fourth convolutional layer.

[0052] In this embodiment, the three convolutional kernels are 3×1, 5×1, and 7×1 respectively, with 16 output channels each; the third and fourth convolutional kernels are 1×1 and 3×1 respectively, with 16 and 8 output channels respectively; the unit output feature size is 6×10×8. The multi-scale structure can capture feature information of different granularities, and the residual connections effectively alleviate the gradient vanishing problem in deep networks, improving model stability.

[0053] The recognition output unit consists of a series of flattened and fully connected layers: the multi-scale fusion enhancement features of size 6×10×8 are flattened into a 480-dimensional one-dimensional vector; the fully connected layer is set with 5 neurons, and the Softmax activation function is used to output the probability of each category, and the maximum probability is taken as the final diagnosis result.

[0054] This invention employs three parallel convolutional layers of different scales to simultaneously extract features from different receptive fields. A concat layer then fuses and complements these multi-scale features, followed by dimensionality reduction and fusion by a third convolutional layer. Finally, an adder (A3) performs residual concatenation between the processed multi-scale features and the original input features, before outputting the final result through a fourth convolutional layer. This approach simultaneously captures minute local anomalies and global trend changes in hydropower station monitoring data, avoiding feature omissions caused by single-scale convolution. Furthermore, the residual concatenation effectively mitigates the gradient vanishing problem during deep network training, enhancing the integrity of feature propagation and the stability of network training, ultimately outputting enhanced multi-scale fused features.

[0055] In this embodiment, the diagnostic results are divided into five warning levels, as defined below: Level 1 Warning: Monitoring data is normal, with no abnormalities. Level 2 warning: Slight data fluctuations exist; it is recommended to strengthen daily monitoring. Level 3 warning: Significant anomalies are observed in the data, requiring intensified monitoring and on-site inspections. Level 4 Warning: The data anomaly is severe. Report immediately and initiate emergency procedures. Level 5 warning: There is a major safety hazard that endangers the safety of structures and personnel. Emergency response and personnel evacuation operations should be carried out.

[0056] In this embodiment, the training process of the anomaly recognition neural network is as follows: First, historical monitoring data of the hydropower station is collected, and sample labeling is completed by combining equipment operation records and on-site inspection results to construct a training dataset containing five-level warning labels; the full-dimensional feature matrix extracted by the process of this invention S1~S5 is used as the network input, and the labeled warning level is used as the supervision label. The cross-entropy loss function is selected to calculate the deviation between the predicted value and the true value; gradient descent optimization algorithm is used for iterative training until the loss function converges and the accuracy of the validation set no longer improves; the optimal model parameters are saved to complete the network training for real-time anomaly recognition on site.

[0057] This invention relies on a scheme that combines EMD mode decomposition, multi-sensor weighted fusion, multi-dimensional feature extraction and multi-scale neural network to overcome the limitations of the traditional fixed threshold method. It fully explores the temporal patterns of monitoring signals and the correlation features of multiple sensors, effectively improving the accuracy and reliability of anomaly identification in hydropower station safety monitoring.

[0058] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying anomalies in safety monitoring data of a hydropower station, characterized in that, Includes the following steps: S1. Standardize the monitoring signals collected by multiple types of sensors to obtain standardized monitoring signals; S2. Perform EMD mode decomposition on the standardized monitoring signal to obtain multi-order modal components; perform weighted processing on the modal components of different sensors of the same type to obtain multi-order fused modal components; the multi-order fused modal components include multi-order IMF fused components and residual trend term fused components; S3. Extract IMF feature vectors from each sub-segment of each IMF fusion component, and superimpose the IMF feature vectors corresponding to the same type of sensor to construct an IMF aggregated feature matrix. S4. Extract trend feature vectors from each sub-segment of the residual trend term fusion component and construct a trend feature matrix; S5. Concatenate the IMF aggregated feature matrix and the trend feature matrix corresponding to the same type of sensor to obtain the full-dimensional feature matrix; S6. An anomaly detection neural network is used to process the full-dimensional feature matrices of various sensors and output the anomaly diagnosis results for hydropower station safety monitoring. Step S2 includes the following sub-steps: S21. Perform EMD mode decomposition on the standardized monitoring signal to obtain multi-order modal components; the multi-order modal components include multi-order IMF components and residual trend component; S22. Taking each time point of each modal component as the center, extract the central neighborhood range and calculate the standard deviation to obtain the local standard deviation corresponding to each time point. S23. Calculate the weighted weights for each time point based on the local standard deviations at each time point; S24. Based on the weighting weights, perform weighted operations on the same-order modal components of different sensors of the same type to obtain the fused modal components of the corresponding order. The specific process of step S23 is as follows: using an exponential function to map the local standard deviation of each modal component and each moment in the standardized monitoring signal of each sensor to the component stability; summing the component stability of all sensors of the same type, modal components of the same order, and corresponding components at the same moment to obtain the total component stability; and using the ratio of the component stability at a single moment to the total component stability as the weighting weight at that moment. The specific process of step S24 is as follows: multiplying the weighting weight of each modal component of a single sensor at each moment with the amplitude of the modal component of the same order of the sensor at the corresponding moment, and then accumulating the multiplication results of all sensors of the same type to obtain the fused modal component of the corresponding order.

2. The method for identifying anomalies in hydropower station safety monitoring data according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Divide each IMF fusion component into several IMF component segments; S32. For each IMF component segment, extract three types of feature elements: local energy, peak value, and maximum value, and construct the IMF feature vector. S33. Superimpose the IMF feature vectors corresponding to the same type of sensor, the same order of IMF fusion component, and the same IMF component sub-segment to obtain the IMF aggregated feature vector of the sub-segment. S34. Arrange all IMF aggregated feature vectors corresponding to the same type of sensor in chronological order to generate an IMF aggregated feature matrix.

3. The method for identifying anomalies in hydropower station safety monitoring data according to claim 2, characterized in that, The method for obtaining each feature parameter in step S32 is as follows: Within a single IMF component segment, the instantaneous energy is obtained by squaring the amplitude at each time step, and the local energy of the segment is obtained by summing all the instantaneous energies. The average energy is obtained by taking the average of all instantaneous energies within the IMF component segment, and the ratio of the maximum instantaneous energy within the segment to the average energy is taken as the peak value. The maximum instantaneous energy within the IMF component segment is taken as the maximum value.

4. The method for identifying anomalies in hydropower station safety monitoring data according to claim 1, characterized in that, Step S4 includes the following sub-steps: S41. Divide the residual trend term fusion component into several trend component segments; S42. Squaring the amplitude at each time step within each trend component segment yields the instantaneous energy at the corresponding time step. S43. Accumulate all instantaneous energies within a single trend component segment to obtain the local energy of that segment; S44. Calculate the local mean of instantaneous energy within a single trend component segment and the global mean of all instantaneous energy of the residual trend term fusion component, respectively. S45. Subtract the local mean from the local energy of the trend component segment to obtain the local deviation. S46. Subtract the global mean from the local energy of the trend component segment to obtain the global deviation. S47. Construct a trend feature vector using the local energy, local deviation, and global deviation of the same trend component segments as feature elements. S48. Arrange all trend feature vectors corresponding to the same type of sensor in chronological order to generate a trend feature matrix.

5. The method for identifying anomalies in hydropower station safety monitoring data according to claim 1, characterized in that, The processing procedure of the anomaly detection neural network in step S6 includes the following sub-steps: S61. The feature compression enhancement unit processes the full-dimensional feature matrix corresponding to each type of sensor to obtain compressed features; S62. The joint weight generation unit is used to calculate all compressed features to obtain the joint weight; S63. The joint weights are multiplied element-wise by the multiplier to obtain the joint enhanced compression features; S64. Multi-scale residual units are used to process each joint enhancement compression feature separately to obtain multi-scale fused enhancement features; S65. All multi-scale fusion enhancement features are added element-wise by adder A1 to obtain multi-type fusion features; S66. The identification output unit processes multiple types of fused features and outputs the abnormal diagnosis results of hydropower station safety monitoring.

6. The method for identifying anomalies in hydropower station safety monitoring data according to claim 5, characterized in that, The feature compression enhancement unit includes a first convolutional layer, an average pooling layer, a max pooling layer, an adder A2, and a second convolutional layer; The input of the first convolutional layer is the input of the feature compression enhancement unit, and the output of the first convolutional layer is connected to the input of the average pooling layer and the input of the max pooling layer, respectively. The input terminals of adder A2 are connected to the output terminals of the average pooling layer and the max pooling layer, respectively, and the output terminal of adder A2 is connected to the input terminal of the second convolutional layer. The output of the second convolutional layer is the output of the feature compression enhancement unit.

7. The method for identifying anomalies in hydropower station safety monitoring data according to claim 5, characterized in that, The computational expression for the joint weight generation unit is: ， In the formula, The weights are joint weights, and sigmoid is the activation function. For 1×1 convolution, This is the first compression feature. For the first One compression feature, For the first One compression feature, It is a positive integer. This is the number of the compression feature.

8. The method for identifying anomalies in hydropower station safety monitoring data according to claim 5, characterized in that, The multi-scale residual unit includes a first-scale convolutional layer, a second-scale convolutional layer, a third-scale convolutional layer, a Concat layer, a third convolutional layer, an adder A3, and a fourth convolutional layer; The input terminals of the first-scale convolutional layer, the second-scale convolutional layer, the third-scale convolutional layer, and the first input terminal of adder A3 are interconnected and serve as the input terminals of the multi-scale residual unit. The input of the Concat layer is connected to the output of the first-scale convolutional layer, the second-scale convolutional layer, and the third-scale convolutional layer, respectively, and the output of the Concat layer is connected to the input of the third convolutional layer. The second input of adder A3 is connected to the output of the third convolutional layer, and the output of adder A3 is connected to the input of the fourth convolutional layer. The output of the fourth convolutional layer is the output of the multi-scale residual unit.

Images