Hydroelectric generating set axis trajectory intelligent identification method fusing multivariate throw signals and nonlinear dynamic indexes

Abstract

The invention discloses a hydroelectric generating set axis trajectory intelligent identification method fusing a multivariate throw signal and a non-linear dynamic index, and relates to the field of intelligent operation and maintenance of a hydroelectric generation system, and an axis trajectory feature quantification tool fusing a symbolization theory and the non-linear dynamic index is developed by establishing an axis trajectory timing mechanism. And a hydroelectric generating set shafting state identification model integrating signal acquisition, feature extraction and mode identification is constructed, and accurate identification of the equipment shafting abnormal state is realized. Rich state information contained in the axis track of the hydroelectric generating set can be deeply excavated, and on the premise that normal production and power generation of the hydroelectric generating set are not affected, efficient recognition of a hydropower station on the early-stage abnormal state of the shaft system is improved, so that the shutdown and maintenance cost caused by shaft system faults is reduced, and the working efficiency is improved. And an effective technical means is provided for improving the operation reliability of the shaft system and ensuring the safe and stable operation of the hydroelectric generating set.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent operation and maintenance of hydropower generation systems, and in particular relates to an intelligent identification method for the axis trajectory of a hydropower unit integrating multivariate swing signals and nonlinear dynamic indicators. Background Art

[0002] With the development of pumped storage technology, hydropower has transformed from a single "energy supplier" to an "energy supplier and regulator", taking on important tasks such as power generation, peak load regulation, and frequency regulation. Frequent dispatching tasks have aggravated the aging of hydropower equipment, and frequent operation in low-load vibration areas has also increased the possibility of equipment failure, especially in the case of high head and large capacity, the consequences of failure are more serious. Therefore, timely diagnosis of hydropower unit failures has become a core issue of intelligent operation and maintenance of hydropower stations.

[0003] The axis trajectory is an important indicator of the operating status of the shaft system. Through the displacement changes in different directions, the operating status of the main shaft of the hydropower unit can be dynamically displayed. Under normal circumstances, the axis trajectory is approximately circular, but in the case of a fault, it may present different shapes such as "8-shaped" or "petal-shaped". Therefore, identifying the operating status of the unit through the shape of the axis trajectory and quickly detecting abnormalities is an important research direction for fault diagnosis of hydropower station equipment.

[0004] Existing shaft trajectory fault detection methods are mainly divided into two categories: 1) mechanism modeling method; 2) image recognition method. The mechanism modeling method simulates the operating state of the turbine through mathematical equations, analyzes the changes of indicators such as the shaft trajectory, and studies the movement mode of the turbine under different fault conditions. However, due to the simplification of the turbine model, the combined effects of fluid dynamics, dynamic pressure and electromagnetic force are not fully considered. The simulated shaft trajectory may not be consistent with the actual situation, and some fault mechanisms have not been studied in depth, which limits the wide application of this method. Image recognition methods use computer vision algorithms, feature detection technology and artificial intelligence to identify different shapes of shaft trajectories to detect shaft system faults. These methods mainly capture the morphological features of the shaft trajectory and combine classification algorithms to achieve automatic recognition. At present, the shaft trajectory image recognition methods are roughly divided into two categories: 1) Artificial feature extraction to identify the type of shaft trajectory. In recent years, with the development of artificial intelligence technology, deep learning algorithms have been widely used in shaft trajectory recognition. These algorithms use end-to-end recognition methods to directly input images to recognize trajectory shapes. Although these methods improve the efficiency of fault detection, they still have the following limitations: 1) Insufficient tolerance to environmental interference. The complex operating environment causes a lot of noise in the axis trajectory. Although signal noise reduction can be performed, it will reduce the recognition efficiency. 2) Weak feature interpretability. After converting the original signal into an image, the interpretability of the features and results is weakened. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent identification method for the shaft center trajectory of a hydropower unit that integrates multiple swing signals and non-linear dynamic indicators, so as to improve the efficient identification of the early abnormal state of the shafting system by the hydropower station without affecting the normal production and power generation of the hydropower unit, and to ensure the safe and stable operation of the power station.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is as follows: An intelligent identification method for the shaft center trajectory of a hydropower unit that integrates multiple swing signals and non-linear dynamic indicators, the steps are as follows: Step 1: Collect and store multiple swing signals characterizing the shaft center trajectory of the hydropower unit, and construct a database of the shaft center trajectory of the hydropower unit; Step 2: Integrate the improved coarse-graining theory, multi-dimensional embedding theory and symbolic theory to develop RCM v MDSE to realize multi-scale extraction of the characteristics of multiple swing signals of the hydropower unit; Step 3: Establish an intelligent identification model for the shaft center trajectory of the hydropower unit based on RCM v MDSE and integrated learning technology; Step 4: Construct and train an intelligent identification model based on the identified swing signal data to obtain a fully trained intelligent identification model for the shaft center trajectory of the hydropower unit; Step 5: Use the intelligent identification model of the shaft center trajectory of the hydropower unit to monitor the swing signal of the unit in real time, and make a real-time judgment on the operating state of the shaft center trajectory of the unit.

[0007] Furthermore, the change of the shafting swing of the hydropower unit over time is sensed by swing sensors installed perpendicular to each other on the water guide, and the current signal is connected to the data acquisition instrument in the form of hard wiring for calibration and converted into a swing signal; the swing signal characterizing the shaft center trajectory of the hydropower unit is transmitted to the host computer system through wired communication, and the data is stored in real time based on a relational database and a time series database; the shaft center trajectory in the shaft center trajectory database is manually identified to construct a shaft center trajectory database of the hydropower unit containing shaft center trajectory - swing signal - operating state, which is used as the data acquisition module of the intelligent identification model.

[0008] Furthermore, the improved coarse-graining theory, multi-dimensional embedding theory and symbolic theory are integrated to develop RCMvMDSE to extract the characteristics of multiple swing signals. It should be noted that the coarse-graining method used in RCMvMDSE combines refined and composite processing techniques, effectively solving the disadvantage of unstable entropy value of the traditional coarse-graining method at high scales.

[0009] RCM v The feature extraction process of MDSE is as follows: First, the coarse-grained time signals of different channels of the multi-variable swing signal are obtained respectively according to the improved coarse-graining method, and the specific expressions are as follows:

[0010] Among them, and are the sampling points and the number of channels respectively; represents the th coarse-grained sequence, and represent the number of indices in the coarse-grained time signal and the original signal, represents the scale factor; Then, through the normal cumulative distribution function (NCDF) and the rounding function, the coarse-grained time signal is converted into a multi-dimensional symbolic signal . The conversion process is as follows:

[0011]

[0012] Among them, , and its value is between 0 and 1; is the p th sample point in the k th coarse-grained sequence of the i th channel time series; and are the standard deviation and the mean value of the th coarse-grained sequence of the k th channel time series respectively; is the number of categories of the symbol sequence .

[0013] Secondly, the symbol sequence is reconstructed into the corresponding symbol delay vector by using the multi-dimensional embedding theory, and the calculation formula is as follows:

[0014] Among them, and are the embedding dimension vector and the time delay vector, and the expressions are and respectively; ranges from 1 to , .

[0015] At the same time, the with different scale factors is calculated:

[0016] Among them, denotes the similarity probability between any two vectors and in, denotes the sample points of. The expression of is as follows:

[0017] In summary, calculate the mean values of all and under the scale from 1 to below, and use them to find at scale and : below RCM v MSDE :

[0018] Finally, repeat the above process until reaches the maximum scale factor At this time, RCM v MSDE the set of is characterized by this multivariate swing:

[0019] Use RCM v MSDE to extract the characteristics of the multivariate swing signal of the hydro-generator unit's shaft center trajectory, and use this as the feature extraction module of the intelligent recognition model.

[0020] Furthermore, on the basis of the signal acquisition and feature extraction module, introduce the random forest (RF), an ensemble learning technique, to build an intelligent recognition model for the hydro-generator unit's shaft center trajectory that integrates signal acquisition - feature extraction - pattern recognition, and realize the automatic recognition and diagnosis of the unit's shaft center trajectory state. The specific training and testing process of RF is as follows: First, adopt the bootstrap sampling technique to randomly select L samples from the original feature dataset to create a sub-training set. This process is repeated T times to generate T different sub-training sets. In each sub-training set, about 2 / 3 of the samples are called in-bag samples, which are used to train the CART model, and the remaining 1 / 3 of the samples are called out-of-bag samples (OOB samples), which are reserved for evaluating the classification error rate of the generated tree model.

[0021] Then, build and train the CART model on each unique in-bag dataset without any pruning. Note that during the model creation process, the features of each node are randomly selected from the available features.

[0022] Finally, the test data is passed to the trained CART model to generate T The final prediction result is determined by majority voting, i.e. T The prediction with the highest frequency among the results is taken as the final result.

[0023] Furthermore, the massive labeled historical axis trajectory swing data in the axis trajectory database of the hydropower unit are input into the feature extraction model to extract key information, and these features are input into the RF model for fitting training to explore the complex nonlinear relationship between the axis trajectory characteristics and the state, and obtain a fully trained axis trajectory intelligent recognition model.

[0024] Furthermore, the intelligent recognition model of the axis trajectory of the hydropower unit is used to determine the operating status of the axis trajectory of the unit in real time, and to issue real-time alarms for abnormal conditions of the unit. Indicators such as accuracy, precision, recall rate, and F1-score are used to evaluate the performance of the intelligent recognition model.

[0025] The present invention can achieve the following beneficial effects: (1) The present invention discloses a refined composite multivariate multiscale discrete sample entropy algorithm that can effectively comprehensively consider the correlation between two swing signals and quantitatively evaluate the coupling information of multidimensional time series. This method overcomes the technical difficulties of extracting complex multi-sensor feature information at multiple time scales and provides a new technical means for multi-sensor state fusion identification of hydropower units.

[0026] (2) The intelligent identification method of the axis trajectory of a hydropower unit that integrates multivariate swing signals and nonlinear dynamic indicators disclosed in the present invention can directly mine feature information from the original waveform of the axis trajectory and automatically determine the operating status of the unit. The processed data are all original data, with extremely high feasibility and good engineering application value.

[0027] (3) Compared with the existing image-based axis trajectory recognition technology for hydropower units, the axis trajectory recognition technology based on multivariate swing signals disclosed in the present invention has better environmental interference resistance and feature interpretation performance, which is of great significance for the current intelligent operation and maintenance of hydropower units. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a schematic diagram of the process of the present invention.

[0029] Figure 2 It is a structural block diagram of the fine composite multivariate multiscale dispersion sample entropy of the present invention.

[0030] Figure 3It is the axial center locus and multi - variable swing waveform diagram of a certain hydropower unit.

[0031] Figure 4 It is the feature distribution diagram extracted by refined composite multi - variable multi - scale dispersion sample entropy.

[0032] Figure 5 It is the low - dimensional feature distribution diagram extracted by refined composite multi - variable multi - scale dispersion sample entropy.

[0033] Figure 6 It is the recognition result diagram of the intelligent recognition model of the present invention on the training set samples.

[0034] Figure 7 It is the recognition result diagram of the intelligent recognition model of the present invention on the test set samples.

[0035] Figure 8 It is the recognition result diagram of the intelligent recognition model of the present invention on the test set samples.

[0036] Figure 9 It is the result comparison diagram between the present invention and the image recognition model. Detailed implementation manners

[0037] Aiming at the deficiencies of the existing technology, the present invention studies the essence of the regression axial center locus and proposes an intelligent recognition method for the axial center locus of hydropower units that integrates multi - variable swing signals and non - linear dynamics indicators. First, eddy current sensors installed on the turbine shafting are used to collect the operation data of the axial center locus, and a set of orthogonal swing signals are obtained. Secondly, based on the multi - dimensional embedding theory, a new non - linear dynamics method - refined composite multi - variable multi - scale dispersion sample entropy (RCM v MDSE) is proposed. Finally, combining random forest and RCM v MDSE realizes the intelligent recognition of the turbine axial center locus. It provides an effective technical means to reduce the possibility of accidental shutdown of the hydropower system and significantly improve the efficient conversion of hydropower energy. The specific steps are as follows: Step 1: The swing sensors installed perpendicular to each other on the water guide are used to sense the change of the swing of the hydropower unit shafting over time, and the current signal is connected to the data acquisition instrument in the form of hard wiring for calibration and converted into a swing signal; the swing signal representing the axial center locus of the hydropower unit is transmitted to the upper computer system through wired communication, and the data is stored in real - time based on the relational database and the time - series database; the axial center locus in the axial center locus database is manually marked to construct a hydropower unit axial center locus database containing axial center locus - swing signal - operation status, which is used as the data acquisition module of the intelligent recognition model. Step 2: Integrate and improve the coarse - graining theory, multi - dimensional embedding theory and symbolization theory to develop RCM vMDSE realizes the multi-scale extraction of the multivariate swing signal features of the hydropower unit; as shown in the flow chart Figure 2 shown, the RCM v The specific process of MDSE is as follows: Step 2.1: Obtain the coarse-grained time signals of different channels of the multivariate swing signal according to the improved coarse-graining method, and the specific expression is as follows:

[0038] Among them, and are the sampling points and the number of channels respectively; represents the th coarse-grained sequence, represents the scale factor; and represent the number of indices in the coarse-grained time signal and the original signal; Step 2.2: Through the normal cumulative distribution function (NCDF) and the rounding function, convert the coarse-grained time signal into a multi-dimensional symbolic signal . The conversion process is:

[0039]

[0040] Among them, , and its value is between 0 and 1; is the p th sample point in the k th coarse-grained sequence of the i th channel time series; and are the standard deviation and the average value of the th coarse-grained sequence of the k th channel time series respectively; is the number of categories of the symbol sequence .

[0041] Step 2.3: Use the multi-dimensional embedding theory to reconstruct the symbol sequence into the corresponding symbol delay vector , and the calculation formula is as follows:

[0042] Among them, and are the embedding dimension vector and the time delay vector, and the expressions are and respectively; ranges from 1 to , .

[0043] Step 2.4: Calculate the different scaling factors :

[0044] in, express The similarity probability of any two vectors in is, express of sample points. The expression is as follows:

[0045] Step 2.5: Calculate the value from scale 1 to Download all and and use them to find the scale Next RCM v MSDE :

[0046] Step 2.6: Repeat the above process until Reaching the maximum scale factor ,at this time RCM v MSDE The collection of is characterized by the multivariate swing signal:

[0047] Step 3: Based on the signal acquisition and feature extraction modules, random forest (RF), an integrated learning technology, is introduced to build an intelligent recognition model for the axis trajectory of the hydropower unit that integrates signal acquisition, feature extraction, and pattern recognition, so as to realize the automatic recognition and diagnosis of the axis trajectory status of the unit.

[0048] Step 4: Input the massive labeled historical axis trajectory swing data in the axis trajectory database of the hydropower unit into the feature extraction model to extract key information, and input these features into the RF model for fitting training to explore the complex nonlinear relationship between the axis trajectory characteristics and the state, and obtain a fully trained axis trajectory intelligent recognition model.

[0049] Step 5: Use the intelligent recognition model of the axis trajectory of the hydropower unit to determine the operating status of the axis trajectory of the unit in real time, issue a real-time alarm for abnormal status of the unit, and use indicators such as the accuracy, precision, recall rate and F1-score of the tested samples to evaluate the performance of the intelligent recognition model.

[0050] Example: The WT100 type deflection sensors installed perpendicular to each other on the water guide of the water turbine are used to sense the change of the shaft system swing of the hydropower unit over time, and the current signal is connected to the data acquisition instrument in the form of hard wiring for calibration and converted into a deflection signal; the deflection signal representing the shaft center locus of the hydropower unit is transmitted to the upper computer system through the TCP finite communication method, and the data is stored in real time based on the relational database MySQL and the time series database Influxdb; the shaft center locus in the shaft center locus database is artificially identified, and a total of six states of the shaft center locus and the corresponding multi-component deflection signals, such as normal, misalignment, imbalance, rubbing, micro-crack and oil film whirl (in the example, they are respectively called states 1-6), are obtained, forming a shaft center locus database of the hydropower unit.

[0051] Figure 3 The shaft center locus of the shaft system of the hydropower unit in different states and its corresponding swing signal waveform diagram are shown. As Figure 3 shown, the shaft center locus in different states shows significant differences. For example, the shaft center locus of state 6 is significantly different from that of other states, while the shaft center locus of state 4 forms a typical figure-eight pattern, which is the characteristic of a serious friction and impact fault. In contrast, the shaft center loci in states 1-3 and state 5 are more similar, so it is very difficult to visually distinguish the corresponding states, which brings great difficulties and potential hazards to the artificial discrimination of the shaft center locus state. The performance of the swing signal is different from that of the shaft center locus. First, the amplitude of the swing signal in state 1 is smaller than that of other states, indicating that the shaft system is more stable under normal conditions. Secondly, the swing signal in state 6 is more complex, reflecting the complex signal characteristics related to extreme dangerous situations such as oil film whirl. Finally, there are also some changes in the swing signals in states 2, 3 and 5. For example, the amplitude in the X direction in state 2 remains within a very small range, while the amplitudes in the X and Y directions in state 5 are basically similar, forming a sharp contrast with the different amplitudes in state 3. By analyzing these swing signals, the feasibility of identifying the shaft center locus based on the swing signal is further proved.

[0052] A total of 300 groups of data were randomly selected from the shaft center locus database of the hydropower unit for example analysis, with 50 groups of data for each state. Using RCM v MDSE to extract the swing signal characteristics corresponding to the shaft center locus of the hydropower unit in different operating states. Figure 4 The characteristic distributions of these shaft center loci in six states are shown. As Figure 4 shown, the shaft center locus characteristics in the same state generally remain smooth, indicating that RCM v MDSE shows good stability in feature extraction. In addition, RCM vMDSE can also effectively capture the motion characteristics in different states during the feature extraction process. For example, the eigenvalue of state 4 and state 6 is significantly higher than that of the other four states. This is mainly because rubbing and oil film whirl represent strong abnormal working conditions, resulting in more fault information in their shaft center orbits, thus generating larger features. The features of the shaft center orbit of state 1 show an obvious upward trend at scales 10, 15, and 17, while states 2, 3, and 5 do not show this trend. Although the features of state 3 and state 5 are visually indistinguishable, the features of state 2 show an upward trend around scale 15, which is not seen in state 5.

[0053] The feature dataset is evenly divided into a training set and a test set, and each training set contains 175 samples. As Figures 5 - 7 shown, the RF model is trained using the features in the training set, and then the RF model is used to identify the unknown samples in the test set. Figure 5 shows the result of feature visualization using t-distributed stochastic neighbor embedding (TSNE). Obviously, there is almost no obvious mixing between the shaft center orbit features of the six states, and only a small part of state 4 overlaps with state 2. Generally speaking, the shaft center orbits of different states can be clearly distinguished by the naked eye. Figure 6 shows the recognition performance of the trained RF model on the training set, and there is no misclassification, indicating that the model has been fully trained. Subsequently, the trained model is applied to the unknown samples in the test set, and the specific recognition results are as Figure 7 shown. The results show that the model achieves perfect recognition on the test set and there is no misjudgment for the shaft system orbits in any state. These results confirm that the proposed model shows strong recognition performance in this case, further verifying its effectiveness in the shaft center orbit recognition task.

[0054] The process image recognition algorithm is introduced to conduct a comparative analysis with the shaft center orbit recognition technology involved in the present invention to verify the superiority of the invention. The involved image models include Alexnet, Resnet, and Hu invariant moment. In addition, referring to previous studies, the present invention selects four indicators to quantify the recognition performance of different models: accuracy rate (AR), precision rate (PR), recall rate (RR), and F1-score. The specific formulas are as follows:

[0055]

[0056]

[0057]

[0058] In the formula, TP , TN, FP and FN represent true positive, true negative, false positive, and false negative.

[0059] To better demonstrate the effectiveness of feature extraction by different models, the present invention adopts a method to reduce the dimension of these features, thereby obtaining a low-dimensional but highly informative representation. In this process, AlexNet and ResNet extract the data before the softmax layer as features. The feature visualization results of different models are shown in Figure 8 as shown. As Figure 8 shown, the features extracted by the three models of AlexNet, ResNet, and Hu invariant moment are all aliased to a certain extent, and the overall feature extraction effect is not as good as the method involved in the present invention.

[0060] The examples adopt three different ratios of training set to test set: 1:1, 2:1, and 3:1. Each case is tested 100 times, and the performance indicators of the four models can be seen in Figure 9 . As Figure 9 shown, the intelligent recognition model involved in the present invention has achieved the best recognition effect in all indicators. In the experiment with a ratio of training set to test set of 1:1, its AR, PR, RR, and F1-score reached 99.430%, 99.464%, 99.433%, and 99.432% respectively. In the experiment with a ratio of 2:1, these values were 99.603%, 99.605%, 99.600%, and 99.586% respectively. In the experiment with a ratio of 3:1, they were 99.660%, 99.671%, 99.663%, and 99.653% respectively. The image recognition models ResNet, AlexNet, and Hu invariant moment ranked 2nd, 3rd, and 4th respectively, and the overall ranking was inferior to the intelligent recognition model involved in the present invention, verifying the superiority of the method involved in the present invention.

[0061] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations on the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. An intelligent identification method for the axis trajectory of a hydropower unit integrating multivariate swing signals and nonlinear dynamic indicators, characterized in that The following steps are involved: Step 1: Collect and store the multivariate swing signal representing the axis trajectory of the hydropower unit, and build a database of the axis trajectory of the hydropower unit; Step 2: Integrate and improve the coarse-grained theory, multidimensional embedding theory, and symbolic theory to establish RCM v MDSE, RCM v MDSE is a fine composite multivariate multiscale dispersion sample entropy, which is used to extract the multivariate swing signal characteristics of hydropower units at multiple scales; Step 3: Build RCM v Intelligent identification model of the axis trajectory of hydropower units based on MDSE and ensemble learning methods; Step 4: Construct and train an intelligent recognition model based on the identified swing signal data to obtain a fully trained intelligent recognition model for the axis trajectory of the hydropower unit; Step 5: Use the intelligent recognition model of the hydropower unit axis trajectory to monitor the unit swing signal in real time and make real-time judgment on the unit axis trajectory operation status.

2. According to claim 1, a method for intelligently identifying the axis trajectory of a hydropower unit integrating multivariate swing signals and nonlinear dynamic indicators is characterized by: The specific method of step 1 is as follows: Step 1.1, using the swing sensors installed perpendicular to each other on the water guide to sense the change of the shaft swing of the hydropower unit over time, and connecting the current signal to the data acquisition instrument through hard wiring to calibrate and convert it into a swing signal; Step 1.2, transmitting the swing signal representing the axis trajectory of the hydropower unit to the host computer system through wired communication, and storing the data in real time based on the relational database and the time series database; Step 1.3, manually mark the axis trajectory in the axis trajectory database, and build a hydropower unit axis trajectory database including axis trajectory-swing signal-operating status. The hydropower unit axis trajectory database is used as a data acquisition module for the intelligent recognition model.

3. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 1 is characterized in that: RCM v The construction process of MDSE is as follows: Step 2.1, obtaining coarse-grained time signals of different channels of the multivariate swing signal respectively according to the improved coarse-grained method; Step 2.2, converting the coarse-grained time signal into a multi-dimensional symbol signal through a normal cumulative distribution function and a rounding function; Step 2.3, reconstruct the symbol sequence into the corresponding symbol delay vector using multi-dimensional embedding theory; Step 2.4: Repeat steps 2.1 to 2.3 until the scale factor Reaching the maximum scale factor , using RCM v MDSE extracts the characteristics of the multivariate swing signal of the axis trajectory of the hydropower unit and uses RCM v MDSE is used as the feature extraction module of the intelligent recognition model.

4. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 3 is characterized in that: In step 2.1, the coarse-grained time signals of different channels of the multivariate swing signal are obtained according to the improved coarse-grained method, and the expressions are as follows: ； in, and are the number of sampling points and channels respectively; Indicates A coarse-grained sequence, and represents the number of indices in the coarse-grained time signal and the original signal, Represents the scale factor.

5. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 4 is characterized in that: In step 2.2, the coarse-grained time signal is transformed into Converted into multi-dimensional symbolic signal , the conversion process is: ； ； in, , its value is between 0 and 1; For the p The first channel time series k The first i Sample points; and They are The first channel time series k The standard deviation and mean of the coarse-grained series; is a sequence of symbols The number of categories.

6. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 5 is characterized in that: In step 2.3, we use multidimensional embedding theory to embed the symbol sequence Reconstructed into the corresponding symbol delay vector , the calculation formula is as follows: ； in, and are the embedding dimension vector and the time delay vector, respectively. and ; The range is 1 to , ; At the same time, the different scaling factors are calculated : ； in, express The similarity probability of any two vectors in is, express The sample points; The expression is as follows: ； Calculate from scale 1 to Download all and and use them to find the scale Next RCM v MSDE : 。 7. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 6 is characterized in that: In step 2.4, repeat steps 2.1 to 2.3 until Reaching the maximum scale factor ,at this time RCM v MSDE The collection of is characterized by the multivariate swing: ； Leveraging RCM v MSDE extracts the characteristics of the multivariate swing signal of the axis trajectory of the hydropower unit and uses it as the feature extraction module of the intelligent recognition model.

8. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 1 is characterized in that: The method for step 3 is: Based on the signal acquisition and feature extraction modules, an integrated learning method called random forest is introduced to construct an intelligent recognition model of the axis trajectory of the hydropower unit integrating signal acquisition, feature extraction and pattern recognition, so as to realize the automatic recognition and diagnosis of the axis trajectory status of the unit.

9. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 1, characterized in that: The method for step 4 is: The massive labeled historical axis trajectory swing data in the axis trajectory database of the hydropower unit is input into the feature extraction model to extract key information, and these features are input into the RF model for fitting training to explore the nonlinear relationship between the axis trajectory characteristics and the state, and obtain a fully trained axis trajectory intelligent recognition model.

10. The method for intelligently identifying the axis trajectory of a hydropower unit by integrating multivariate swing signals and nonlinear dynamic indicators according to claim 1, characterized in that: The method of step 5 is: use the intelligent recognition model of the axis trajectory of the hydropower unit to determine the operating status of the axis trajectory of the unit in real time, issue a real-time alarm for abnormal status of the unit, and use indicators including accuracy, precision, recall rate and F1-score to evaluate the performance of the intelligent recognition model.

Images