A unit state intelligent identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization
By employing a spatiotemporal holographic panoramic frequency domain statistical fusion visualization method, combined with multi-point data and a support vector machine model, the accuracy and reliability issues of hydropower unit status identification were resolved, achieving accurate reflection and stable identification of unit status, and adapting to complex operating conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA INST OF WATER RESOURCES & HYDROPOWER RES
- Filing Date
- 2026-05-21
- Publication Date
- 2026-07-24
AI Technical Summary
Existing hydropower unit condition identification methods fail to effectively integrate multi-source heterogeneous measurement point information, lack deep coupling correlation analysis and effective Riemannian manifold feature processing, resulting in insufficient identification accuracy and poor reliability, making it difficult to adapt to complex coupled operating conditions.
By employing a spatiotemporal holographic panoramic frequency domain statistical fusion visualization method, and through kernel density estimation of multi-measurement data, gradient vector field calculation, regularized covariance matrix mapping, and support vector machine multi-classification model, we can achieve deep coupling analysis of multi-dimensional operating characteristics of the unit and accurate reflection of its state characteristics.
It significantly improves the accuracy and stability of unit status identification, adapts to complex coupled operating conditions, reduces operation and maintenance costs, reduces the risk of downtime due to failure, and provides reliable technical support.
Smart Images
Figure CN122451607A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of unit condition detection technology, specifically a method for intelligent identification of unit condition based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization. Background Technology
[0002] As core equipment in energy systems, the operating status of hydropower units directly affects the stability and security of energy supply. Due to the strong coupling between water, machinery, and electricity during operation, the operating status of hydropower units is complex and variable, significantly increasing the difficulty of status identification. Current mainstream hydropower unit status identification methods have two major limitations: First, they rely heavily on data from single or similar measuring points, failing to integrate information from different categories of measuring points. This makes it impossible to comprehensively capture the multidimensional operating characteristics of the unit under the coupling of water, machinery, and electricity, and thus difficult to accurately reflect the true operating status of the unit, resulting in insufficient accuracy in identifying complex unit states. Second, traditional status identification and classification are performed in Euclidean space, while the covariance matrix of an image patch actually lies on a Riemannian manifold composed of symmetric positive definite matrices (SPDs). Directly performing classification operations on the Riemannian manifold is difficult and prone to feature distortion and classification bias, further affecting the reliability of status identification.
[0003] Furthermore, while some existing methods attempt to fuse data from multiple measurement points, they are mostly simple superpositions, failing to achieve deep coupling and correlation analysis of data from different categories of measurement points and thus unable to fully explore the inherent correlation characteristics between measurement points. Simultaneously, the lack of effective Riemannian manifold feature processing strategies makes it difficult to solve the adaptation problem between manifold space and Euclidean space, failing to meet the high-precision state identification requirements of hydropower units under complex coupled operating conditions. Therefore, there is an urgent need for an intelligent method that can fuse information from multiple heterogeneous measurement points, adapt to Riemannian manifold feature processing, and improve the accuracy of complex state identification of units, providing reliable technical support for the safe and efficient operation of hydropower units. Summary of the Invention
[0004] To address the aforementioned problems in existing technologies, this application provides a method for intelligent identification of unit status based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization. By fusing key measurement point data from multiple different categories of the unit and performing deep coupling correlation analysis, this method comprehensively captures the multidimensional operating characteristics of the unit, accurately reflects the true operating status of the unit, and significantly improves the accuracy of identifying the complex state of the unit.
[0005] To achieve the above objectives, this application adopts the following technical solution: a method for intelligent identification of unit status based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization, comprising the following steps: The system acquires operating signals from multiple key measurement points of the unit, extracts the amplitude spectrum of the operating signals, performs kernel density estimation on the amplitude spectrum and draws a kernel density estimation map, and combines the kernel density estimation maps of all measurement points into a preset array to generate a multi-measurement point frequency domain statistical map. Feature extraction is performed on the frequency domain statistical map of multiple measurement points, the gradient vector fields in the horizontal and vertical directions of the image are calculated, the image is then divided into multiple non-overlapping local regions, and pixel feature vectors are constructed in each local region to form a data matrix. Calculate the original covariance matrix of each data matrix, perform regularization on the original covariance matrix to obtain a symmetric positive definite covariance matrix, transform the symmetric positive definite covariance matrix to the tangent space through logarithmic mapping, perform semi-vectorization on the tangent space matrix, and concatenate the feature vectors of all local regions to obtain the unit state feature vector. The unit's state feature vector is input into a preset support vector machine multi-classification model, and the type of structural damage is determined based on the output.
[0006] Key measuring points include upper guide X-swivel, lower guide X-swivel, water guide X-swivel, upper frame -X horizontal vibration velocity, lower frame +X horizontal vibration velocity, top cover -X horizontal vibration velocity, bladeless zone pressure, top cover pressure, and volute door noise.
[0007] The gradient vector field is calculated based on the following formula: ; ; in, I ( x , y (x, y) is the pixel gray value of the image at coordinates (x, y); G x It is the gradient matrix in the horizontal direction; G y It is the gradient matrix in the vertical direction.
[0008] Within each local region, a low-dimensional feature vector is constructed for each pixel, and a 3-dimensional feature vector is constructed for the k-th pixel within the region. ; The data matrix Z = [z1, z2, ..., zM]T is formed. Where k is the number of pixels and M is the total number of pixels in each sub-region; It is the magnitude of the gradient.
[0009] The original covariance matrix is: ; After regularization, we get: C =Σraw + I3; in, C is the mean of the eigenvectors within this region; C is a 3×3 symmetric positive definite covariance matrix located on the Riemannian manifold. superior; =0.01 is the regularization coefficient; I3 is the 3×3 identity matrix.
[0010] The use of a log-Euclidean metric framework is equivalent to mapping points on the manifold to the tangent space of the identity matrix I, achieved by calculating the matrix logarithm. T =log m ( C ); Among them, log m Let be the logarithm of a matrix, if C=UΛU T If it is an eigenvalue decomposition, then: T=U diag(ln(λ1),ln(λ2),ln(λ3)) U T ; Where T is the mapped tangent space matrix. λ i These are the eigenvalues of the covariance matrix C.
[0011] After performing semi-vectorization on the tangent space matrix, we obtain: v block =[T 11 ,T 12 ,T 13 ,T 22 ,T 23 ,T 33 ] T ; The unit state feature vector is obtained by concatenating the feature vectors of all local regions: Among them, v block It is a 6-dimensional local descriptor for a single block; F is the final output image feature vector.
[0012] The beneficial effects of this application are: This application provides a method for intelligent identification of unit status based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization. By fusing key measurement point data from multiple different categories of the unit and performing deep coupling correlation analysis, it comprehensively captures the multi-dimensional operating characteristics of the unit, accurately reflecting the true operating status of the unit and effectively solving the problem of difficult status identification caused by the strong coupling of water, machinery, and electricity in hydropower units. An innovative "rolling tangent space" mapping strategy is proposed, effectively solving the problems of high difficulty and feature distortion in direct classification on Riemannian manifolds, thus improving the stability and accuracy of status identification. The method combines a support vector machine multi-classification model with a complete performance evaluation process to ensure the stability and reliability of unit status identification. The multi-source information collaborative processing method effectively improves the method's anti-interference ability and generalization ability, enabling it to adapt to the complex coupled operating conditions of hydropower units, providing strong technical support for the safe and efficient operation of the unit, reducing operation and maintenance costs, and minimizing the risk of downtime due to failures. Attached Figure Description
[0013] Figure 1 This is a flowchart illustrating the intelligent identification method for unit status in this application; Figure 2 A schematic diagram of the field test of the measurement point signal under 150MW power generation operating conditions; Figure 3 A schematic diagram of the field test of the measurement point signal under 300MW power generation operating conditions; Figure 4 A schematic diagram of the field test of the measuring point signal under pumping conditions; Figure 5 A schematic diagram of the field test of the measuring point signal under the pumping phase adjustment condition; Figure 6 Generate image illustrations for each of the four operating conditions. Detailed Implementation
[0014] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.
[0015] like Figure 1 The method for intelligent identification of unit status based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization, as shown, includes the following steps: S1. Acquire the operating signals of multiple key measurement points of the unit, extract the amplitude spectrum of the operating signals, perform kernel density estimation on the amplitude spectrum and draw the kernel density estimation map, combine the kernel density estimation maps of all measurement points according to the preset array, and generate a multi-measurement point frequency domain statistical map.
[0016] Nine measuring points were selected, including the upper guide X swing, lower guide X swing, water guide X swing, upper frame -X horizontal vibration velocity, lower frame +X horizontal vibration velocity, top cover -X horizontal vibration velocity, bladeless area pressure, top cover pressure, and volute door noise. A Fast Fourier Transform was performed on each of the points' signals x(t) to obtain its complex form. Then, the modulus y(i) of the complex number was taken, i.e., the amplitude spectrum was calculated. Kernel density estimation was then performed on y(i) to obtain the probability density function f and the corresponding coordinate xi. The kernel density estimation results were plotted using a bar chart to obtain the kernel density estimation graph. These nine graphs were then combined in a 3×3 manner to generate a total multi-point frequency domain statistical graph.
[0017] S2. Extract features from the multi-point frequency domain statistical graph, calculate the gradient vector field in the horizontal and vertical directions of the image, divide the image into multiple non-overlapping local regions, construct pixel feature vectors in each local region and form a data matrix.
[0018] S21, Image Gradient Field Calculation Calculate the gradients of the input image I in the horizontal and vertical directions. The gradient reflects the rate of change of image gray levels and is key information describing texture edges and details. For image I(x,y), calculate its gradient vector field: ; ; in, I ( x , y (x, y) is the pixel gray value of the image at coordinates (x, y); G x It is the gradient matrix in the horizontal direction; G y It is the gradient matrix in the vertical direction.
[0019] S22, Image Segmentation and Feature Mapping To preserve the spatial distribution information of the image, it was divided into 4×4=16 non-overlapping local regions (patches). Within each region... B i,j Within the vector, for each pixel k, a low-dimensional feature vector z is constructed. k .
[0020] For the k-th pixel within the region, construct a 3D feature vector: ; The data matrix Z = [z1, z2, ..., zM]T is formed. Where k is the number of pixels and M is the total number of pixels in each sub-region; It is the magnitude of the gradient, which enhances robustness against edge direction flipping.
[0021] S3. Calculate the original covariance matrix of each data matrix, perform regularization on the original covariance matrix to obtain a symmetric positive definite covariance matrix, transform the symmetric positive definite covariance matrix to the tangent space through logarithmic mapping, perform semi-vectorization on the tangent space matrix, and concatenate the feature vectors of all local regions to obtain the unit state feature vector.
[0022] S31, Covariance Matrix Construction and Regularization The covariance matrix of each data block is calculated, which incorporates the correlations between different features (e.g., the correlation between gray values and gradients). Since the calculated covariance matrix may become a singular matrix (non-invertible) due to insufficient data or flat regions, it is regularized by adding a small identity matrix to ensure that it becomes a strictly symmetric positive definite matrix (SPD).
[0023] 1) Original covariance: ; 2) Regularization: C =Σ raw + I3; in, C is the mean of the eigenvectors within this region; C is a 3×3 symmetric positive definite covariance matrix located on the Riemannian manifold. superior; =0.01 is the regularization coefficient, which prevents matrix singularity and ensures the positive definiteness of eigenvalues; I3 is a 3×3 identity matrix.
[0024] S32, Logarithmic mapping from Riemannian manifolds to tangent spaces The SPD matrix space is a curved Riemannian manifold, where Euclidean distance (e.g., direct subtraction) fails. To use linear classifiers (e.g., SVM), it is necessary to map points C on the manifold to the tangent space. This method employs the Log-Euclidean metric framework, equivalent to mapping points on the manifold to the tangent space at the identity matrix I, achieved by calculating the matrix logarithm. T =log m ( C ); Among them, log m The matrix logarithm is different from the element-wise logarithm. C=UΛU T If it is an eigenvalue decomposition, then: T=U diag(ln(λ1),ln(λ2),ln(λ3)) U T ; Where T is the mapped tangent space matrix, which is a symmetric matrix but no longer subject to positive definiteness constraints and lies in Euclidean space. λ i These are the eigenvalues of the covariance matrix C.
[0025] S33, Vectorization and Feature Concatenation Since the tangent space matrix T is a symmetric matrix ( T ij =T ji The independent elements of a 3×3 matrix consist only of the upper triangular (or lower triangular) portion. For a 3×3 matrix, there are 3(3+1) / 2=6 independent coefficients. These independent coefficients are flattened into a vector, and the features of all 16 blocks are concatenated.
[0026] 1) Semi-vectorization: v block =[T 11 ,T 12 ,T 13 ,T 22 ,T 23 ,T 33 ] T ; 2) Cascading: Cascading the feature vectors of all local regions yields the unit state feature vector: Among them, v block It is a 6-dimensional local descriptor for a single block; F is the final output image feature vector.
[0027] S4. Input the unit state feature vector into the preset support vector machine multi-classification model, and determine the type of structural damage based on the output.
[0028] S41. Pre-set Support Vector Machine Multi-Classification Model Training After obtaining the feature matrix of the training data, a classification model is built using a Support Vector Machine (SVM). Detailed steps are as follows: 1. Define SVM template 2. Construct a multi-class classification model (ECOC) After the model training is complete, load the test set data for validation. Detailed steps: 1. Extract test set features: Perform the same preprocessing and feature extraction steps on the images in the Test folder as on the training set.
[0029] 2. Prediction: The feature vector of the test sample is input into the trained classifier, and the type of structural damage is determined by the output of the classifier.
[0030] Call the classification model trained in step 3, input the test features, and the model outputs the predicted labels (1, 2, 3...).
[0031] 3. Calculation accuracy: Accuracy = (Number of correctly predicted samples / Total number of test samples) × 100%
[0032] Example: The main parameters of a power plant unit are as follows: runner diameter 4.158m, rated head 430m, rated output 306.1MW, rated speed 428.6r / min, and runner blade count 9; generator / motor rated capacity: 333.3 / 325MVA / MW. Using the field test data of Unit 4 of this power plant as a sample, this study verifies the effectiveness of the unit operation status identification method based on the coupled images of nine measuring points: upper guide X-axis swing, lower guide X-axis swing, water guide X-axis swing, upper frame -X horizontal vibration velocity, lower frame +X horizontal vibration velocity, top cover -X horizontal vibration velocity, bladeless zone pressure, top cover pressure, and volute door noise. Specifically, this includes: (1) Collect signals from 9 measuring points; During normal unit operation, nine sensors are deployed. After stabilization under four operating conditions—150MW power generation, 300MW power generation, pumping operation, and pumping phase regulation—signals are collected from the nine measuring points. The sampling rate is 1000Hz, and the number of sampling points is 4096. The nine signals under the four operating conditions are as follows: Figures 2-5 As shown in the figure, due to the influence of hydraulic, electromagnetic, and mechanical factors, sound signals are very complex and it is difficult to directly identify their state from the figure.
[0033] (2) Generate an image.
[0034] right Figures 2-5 The waveform shown is plotted using bar graphs to estimate the kernel density of the single-sided amplitude spectrum of the signal, resulting in images for four states, as shown below. Figure 6 As shown in the figure, the coupled images differ significantly for different states, and these images can be used to accurately identify the operating status of the unit.
[0035] Images were generated from the acoustic signals of the tailrace inlet valve during four operating conditions: 150MW power generation, 300MW power generation, pumping operation, and pumping phase regulation. From the images generated using 240 sets of acoustic signal data covering the four different operating conditions, 120 sets (30 sets for each condition) were randomly selected as training samples and input into an SVM for learning. The remaining 120 sets of data were then used as test samples for validation. The results show that all test samples were correctly identified, thus proving that it is feasible to determine the operating status of hydropower units using acoustic signal images.
[0036] Finally, it should be noted that in this document, relationships such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.
[0037] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A method for intelligent identification of unit status based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization, characterized in that, Includes the following steps: The operating signals of multiple key measuring points of the unit are acquired, the amplitude spectrum of the operating signals is extracted, the kernel density of the amplitude spectrum is estimated and a kernel density estimation map is drawn, and the kernel density estimation maps of all measuring points are combined according to a preset array to generate a multi-measuring-point frequency domain statistical map. Feature extraction is performed on the multi-point frequency domain statistical map, the gradient vector fields in the horizontal and vertical directions of the image are calculated, and the image is divided into multiple non-overlapping local regions. Pixel feature vectors are constructed in each local region to form a data matrix. Calculate the original covariance matrix of each data matrix, perform regularization on the original covariance matrix to obtain a symmetric positive definite covariance matrix, transform the symmetric positive definite covariance matrix to the tangent space through logarithmic mapping, perform semi-vectorization on the tangent space matrix, and concatenate the feature vectors of all local regions to obtain the unit state feature vector. The unit's state feature vector is input into a preset support vector machine multi-classification model, and the type of structural damage is determined based on the output.
2. The intelligent unit status identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization as described in claim 1, characterized in that, The key measuring points include the upper guide X-axis swing, the lower guide X-axis swing, the water guide X-axis swing, the upper frame -X horizontal vibration velocity, the lower frame +X horizontal vibration velocity, the top cover -X horizontal vibration velocity, the bladeless zone pressure, the top cover pressure, and the volute door noise.
3. The intelligent unit status identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization as described in claim 1, characterized in that, The gradient vector field is calculated based on the following formula: ； ； in, I ( x , y (x, y) is the pixel gray value of the image at coordinates (x, y); G x It is the gradient matrix in the horizontal direction; G y It is the gradient matrix in the vertical direction.
4. The intelligent unit status identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization as described in claim 1, characterized in that, Within each of the aforementioned local regions, for each pixel, a low-dimensional feature vector is constructed; for the k-th pixel within the region, a 3-dimensional feature vector is constructed. ； The data matrix Z = [z1, z2, ..., zM]T is formed. Where k is the number of pixels and M is the total number of pixels in each sub-region; It is the magnitude of the gradient.
5. The intelligent unit status identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization as described in claim 1, characterized in that, The original covariance matrix is: ； After regularization, we get: C =Σ raw + I3; in, C is the mean of the eigenvectors within this region; C is a 3×3 symmetric positive definite covariance matrix located on the Riemannian manifold. superior; =0.01 is the regularization coefficient; I3 is the 3×3 identity matrix.
6. The intelligent unit status identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization as described in claim 1, characterized in that, Using a log-Euclidean metric framework is equivalent to mapping points on a manifold to the tangent space at the identity matrix I, achieved by calculating the matrix logarithm: T =log m ( C ); Among them, log m Let be the logarithm of a matrix, if C=UΛU T If it is an eigenvalue decomposition, then: T=U diag(ln(λ1),ln(λ2),ln(λ3)) U T ; Where T is the mapped tangent space matrix. λ i These are the eigenvalues of the covariance matrix C.
7. The intelligent unit status identification method based on spatiotemporal holographic panoramic frequency domain statistical fusion visualization as described in claim 1, characterized in that, After performing semi-vectorization on the tangent space matrix, we obtain: v block =[T 11 ,T 12 ,T 13 ,T 22 ,T 23 ,T 33 ] T ; The unit state feature vector is obtained by concatenating the feature vectors of all local regions: ; Among them, v block It is a 6-dimensional local descriptor for a single block; F is the final output image feature vector.