Assessment method for health state of pumped storage generator motor
By combining multi-scale signal decomposition and deep learning models, the problem of traditional assessment methods failing to capture early faults under complex operating conditions is solved, enabling accurate assessment of the health status of pumped storage power generators and improving the accuracy and reliability of the assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-10
AI Technical Summary
Traditional health status assessment methods for pumped storage generator motors cannot effectively capture early degradation characteristics of latent faults in operating scenarios with multiple operating conditions, varying head, and frequent start-stop cycles, resulting in limited fault prediction lead time and frequent assessment errors.
The health parameters are decomposed into multiple scales using cascaded variational mode decomposition (CEEMDAN-VMD) with adaptive noise complete set empirical mode decomposition, the eigenvalues of each component are calculated and the eigenvectors are constructed. Combined with stacked denoising autoencoders and attention-based deep learning models (Transformers), the model is trained using particle swarm optimization algorithm, and finally the health status is assessed using the Monte Carlo method.
It enables accurate and efficient assessment of the health status of pumped storage generator motors, improving the accuracy and reliability of fault prediction.
Smart Images
Figure CN121637017A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for assessing the health status of a pumped storage power generator motor, belonging to the technical field of reliability analysis (G06F119 / 02) in digital data processing. Background Technology
[0002] With the high proportion of new energy sources connected to the grid and the improvement of energy electrification, the volatility and uncertainty of the power system have significantly increased. Pumped storage generators, with their large capacity, fast start-stop, and strong climbing characteristics, have become a core force for peak shaving, valley filling, frequency regulation, and inertia support in the power grid. However, because the motors need to frequently switch operating conditions, they are subjected to strong coupling effects of hydraulic, electromagnetic, and mechanical multi-physical fields, making them prone to accumulating latent faults such as stator insulation degradation and bearing wear. If the health status of the motors is not assessed in a timely manner, it will not only shorten the equipment life but also directly threaten the power grid's power supply security.
[0003] However, traditional methods for assessing the health status of pumped storage generator motors often rely on the time and frequency domain characteristics of single-mode signals such as vibration and temperature, combined with shallow model discrimination. This is only applicable to scenarios with stable operating conditions and low noise. When faced with operating scenarios involving multiple operating conditions, variable head, and frequent start-stop, traditional assessment methods may fail to capture early degradation characteristics of latent faults due to issues such as signal mode aliasing and feature redundancy. This results in limited fault prediction lead time and often leads to incorrect assessment of the motor's health status. Summary of the Invention
[0004] The technical problem to be solved by this invention is: how to achieve accurate assessment of the health status of pumped storage power generation motors.
[0005] The technical solution proposed by this invention to solve the above-mentioned technical problems is: a method for assessing the health status of a pumped-storage generator motor, comprising the following steps:
[0006] Step 1: Set the three health states of the pumped storage generator motor as normal state, alert state, and abnormal state;
[0007] Eight health parameters affecting the health status of pumped storage generator motors are set as follows: stator current, stator voltage, vibration state value, stator winding temperature, bearing temperature, output power, upstream and downstream water level difference, and pump guide vane opening.
[0008] Step 2: Obtain 8 historical health parameters of the pumped storage generator motor under three historical health states within a time period T. Use the sliding window method to divide the 8 historical health parameters under the three historical health states into M historical sliding windows. Each historical sliding window contains 8 historical health parameters under any historical health state within a time period T for a certain duration.
[0009] Step 3: Decompose all historical health parameters in each historical sliding window according to the following formula (1).
[0010] (1);
[0011] In equation (1), It is an adaptive noise complete set empirical mode decomposition cascade variational mode decomposition; It is the qth historical health parameter in the m-th historical sliding window; It is the k-th historical modal component decomposed from the q-th historical health parameter in the m-th historical sliding window;
[0012] Step 4: Calculate the first historical eigenvalue of each historical modal component according to the following formula (2). Second historical eigenvalue and the third historical eigenvalue
[0013] (2);
[0014] In equation (2), yes The total number of sampling points within time T; yes The amplitude at the i-th sampling point; yes The average amplitude; yes The total frequency; yes The j-th frequency, yes The amplitude at the j-th frequency;
[0015] Step 4.1: Calculate the fourth historical eigenvalue of each historical modal component according to the following formula (3).
[0016] (3);
[0017] In equation (3), yes The fusion vector, where T is the transpose; It is the mean vector of the fusion vector of the historical modal components decomposed from the historical health parameters under the historical normal state in all historical sliding windows; It is the covariance matrix of the fusion vector of the historical modal components decomposed from the historical health parameters under the historical normal state in all historical sliding windows;
[0018] Step 4.2: Use the sample entropy of each historical mode component as its own fifth historical feature value. ;
[0019] Using Equation (4), the t-th historical eigenvalues of the k-th historical modal component derived from the q-th historical health parameter in all historical sliding windows are respectively constructed into historical feature vectors.
[0020] (4);
[0021] In equation (4), It is the t-th historical feature value of the k-th historical modal component decomposed from the q-th historical health parameter in the m-th historical sliding window;
[0022] Step 4.3: Calculate the historical screening coefficient for each historical feature vector using the following formula (5).
[0023] (5);
[0024] In equation (5); Yes, it is the mean of the t-th historical eigenvalues of the k-th historical modal component derived from the q-th historical health parameter decomposition in all historical sliding windows; It is the historical output power corresponding to the m-th historical sliding window; It is the average of the historical output power corresponding to all historical sliding windows;
[0025] Step 4.4: Sort all historical screening coefficients from smallest to largest, and calculate the first historical threshold using the following formula (6).
[0026] (6);
[0027] In equation (6), It is a quantile function; It is the set of all historical screening coefficients; b is the threshold coefficient, 85≤b≤95;
[0028] First historical threshold The value after +0.01 is used as the second historical threshold, and the value after +0.01 of the second historical threshold is used as the third historical threshold, and so on to obtain the sixth historical threshold.
[0029] The feature vectors corresponding to the historical screening coefficients that are greater than the first historical threshold are used to form the first historical feature vector matrix, the feature vectors corresponding to the historical screening coefficients that are greater than the second historical threshold are used to form the second historical feature vector matrix, and so on to obtain the sixth historical feature vector matrix.
[0030] Step 4.5: Establish 6 historical power prediction models. Input the first historical eigenvector matrix and historical output power into the first historical power prediction model for training to obtain the first historical prediction mean square error of the trained first historical power prediction model. Then input the second historical eigenvector matrix and historical output power into the second historical power prediction model for training to obtain the second historical prediction mean square error of the trained second historical power prediction model. And so on to obtain the sixth historical mean square error.
[0031] The historical threshold corresponding to the smallest historical mean square error among the six historical mean square errors is taken as the final historical threshold, and the historical feature vectors corresponding to the historical screening coefficients that are greater than the final historical threshold are used to form the final historical feature vector matrix.
[0032] Step 5: Input the final historical feature vector matrix into the stacked denoising autoencoder for denoising and dimensionality reduction to obtain the historical fusion feature vector matrix;
[0033] Step 6: Establish a health parameter prediction model. Input the historical fusion feature vector matrix and historical output power into the health parameter prediction model for training to obtain the trained health parameter prediction model. The input of the health parameter prediction model is set as the fusion feature vector matrix, and the output is set as a health parameter prediction vector composed of 8 health parameter prediction values.
[0034] Step 7: Collect real-time data from the pumped-storage generator motor. Eight real-time health parameters were analyzed using a sliding window method, resulting in a total of [number missing] parameters. Several real-time sliding windows, each containing... Eight real-time health parameters over a period of time;
[0035] Step 7.1: Decompose all real-time health parameters in each real-time sliding window using the same method as in Step 3 for decomposing all historical health parameters in each historical sliding window, to obtain the... The first real-time sliding window Decomposition of real-time health parameters One real-time modal component ;
[0036] Step 7.2: Process each real-time modal component in the same way as in Steps 4-4.3 to obtain the first eigenvalue of each real-time modal component. Second eigenvalue Third eigenvalue Fourth eigenvalue and the fifth eigenvalue and real-time feature vectors and each real-time feature vector Real-time filtering coefficient ;
[0037] Step 7.3 Process the real-time filtering coefficients in the same way as in Step 4.4 to obtain 6 real-time thresholds and 6 real-time feature vector matrices.
[0038] Step 7.4: Process the six real-time feature vector matrices in the same way as the six historical feature vector matrices in Step 4.5 to obtain six real-time mean square errors and the final real-time feature vector matrix;
[0039] Step 8: Input the final real-time feature vector matrix into the stacked denoising autoencoder for denoising and dimensionality reduction to obtain the real-time fused feature vector matrix;
[0040] Step 9: Input the real-time fused feature vector matrix into the trained health parameter prediction model and combine it with the Monte Carlo method for prediction, outputting B health parameter prediction value vectors;
[0041] The real-time health value L of the pumped storage generator motor is calculated according to formula (7).
[0042] (7);
[0043] In equation (7), Y1 is the b-th prediction residual vector; Y1 is a real-time health parameter vector composed of 8 real-time health parameters. It is the vector of predicted values for the b-th health parameter; It is the average vector of B predicted residual vectors; It is the first predicted residual vector;
[0044] If 0 < L ≤ 1, the health status of the pumped storage generator motor is assessed as normal.
[0045] If 1 < L ≤ 2, the health status of the pumped storage generator motor is assessed as a state of concern.
[0046] If L > 2, the health status of the pumped storage generator motor is assessed as abnormal.
[0047] Furthermore, the health parameter prediction model in step 6 is built using a deep learning model based on an attention mechanism, called transformer.
[0048] Furthermore, in step 6, the health parameter prediction model is trained using a particle swarm optimization algorithm.
[0049] The beneficial effects of this invention are as follows: By performing multi-scale modal decomposition on health parameters to obtain modal components, calculating five eigenvalues for each component and constructing eigenvectors for each component, and using precise thresholding to select effective eigenvectors to form a matrix for model training, the health status is finally assessed based on the prediction residuals. Compared to the shallow discrimination mode of traditional single-mode signals that relies only on time-domain and frequency-domain features, this invention can achieve accurate and efficient assessment of the health status of pumped storage power generation motors. Attached Figure Description
[0050] The present invention will be further described below with reference to the accompanying drawings.
[0051] Figure 1 This is a schematic diagram of the health parameter monitoring and sensing layout in the health status assessment method for pumped storage power generators in the embodiment.
[0052] Figure 2 This is a flowchart of the CEEMD-VMD breakdown in the health status assessment method for pumped storage generator motors in the embodiment. Detailed Implementation
[0053] Example
[0054] This embodiment of a method for assessing the health status of a pumped-storage generator motor includes the following steps:
[0055] Step 1: Set the three health states of the pumped storage generator motor as normal state, alert state, and abnormal state;
[0056] Eight health parameters affecting the health status of pumped storage generator motors are defined as stator current, stator voltage, vibration state value, stator winding temperature, bearing temperature, output power, upstream and downstream water level difference, and pump guide vane opening.
[0057] Step 2: Obtain 8 historical health parameters for time period T under three historical health states of the pumped storage power generator. Use the sliding window method to divide the 8 historical health parameters under the three historical health states into M historical sliding windows. Each historical sliding window contains 8 historical health parameters for a period of time T under any historical health state.
[0058] In this embodiment, the sliding window method is an existing technology for processing time series data. Its core principle is to slide a window of fixed length L across continuous time series data at preset step sizes, extracting local subsequences as input samples for the model. For example, eight historical health parameters within time T under normal historical conditions will be divided into multiple windows after the sliding window is applied, and each sliding window contains eight historical health parameters for a continuous duration within time T under normal historical conditions.
[0059] Step 3: Decompose all historical health parameters in each historical sliding window according to the following formula (1).
[0060] (1);
[0061] In equation (1), It is an adaptive noise complete set empirical mode decomposition cascade variational mode decomposition; It is the qth historical health parameter in the m-th historical sliding window; It is the k-th historical modal component derived from the q-th historical health parameter in the m-th historical sliding window.
[0062] In this embodiment, the Cascaded Variational Mode Decomposition (CEEMDAN-VMD) method, which integrates two advanced signal decomposition techniques, is used in series to solve the multi-scale feature extraction problem of non-stationary, noisy, and complex signals. Through a two-stage process of coarse and fine subdivision, the purity and effectiveness of signal decomposition are significantly improved. The decomposition process is as follows: Figure 2 As shown.
[0063] Step 4: Calculate the first historical eigenvalue of each historical modal component according to the following formula (2). Second historical eigenvalue and the third historical eigenvalue
[0064] (2)
[0065] In equation (2), yes The total number of sampling points within time T; yes The amplitude at the i-th sampling point; yes The average amplitude; yes The total frequency; yes The j-th frequency, yes The amplitude at the j-th frequency.
[0066] Step 4.1: Calculate the fourth historical eigenvalue of each historical modal component according to the following formula (3).
[0067] (3);
[0068] In equation (3), yes The fusion vector, where T is the transpose; It is the mean vector of the fusion vector of the historical modal components decomposed from the historical health parameters under the historical normal state in all historical sliding windows; It is the covariance matrix of the fusion vector of the historical modal components decomposed from the historical health parameters under the historical normal state in all historical sliding windows.
[0069] Step 4.2: Use the sample entropy of each historical mode component as its own fifth historical feature value. ;
[0070] In this embodiment, sample entropy is used to measure the complexity and irregularity of the sequence. That is, the more irregular and complex the historical modal components are, the larger the sample entropy value is usually. The calculation process is the existing technology.
[0071] Using Equation (4), the t-th historical eigenvalues of the k-th historical modal component derived from the q-th historical health parameter in all historical sliding windows are respectively constructed into historical feature vectors.
[0072] (4);
[0073] In equation (4), It is the t-th historical feature value of the k-th historical modal component decomposed from the q-th historical health parameter in the m-th historical sliding window.
[0074] In this embodiment, the historical feature vector That is, the historical feature vector constructed from the first historical eigenvalue of the first historical modal component derived from the first historical health parameter decomposition in all historical sliding windows. It is the historical feature vector constructed from the second historical feature value of the first historical modal component decomposed from the first historical health parameter in all historical sliding windows. A large number of feature vectors constructed in this way can cover multi-dimensional time series features, accurately capture device status details, and effectively improve the analysis accuracy and reliability of subsequent health parameter prediction models.
[0075] Step 4.3: Calculate the historical screening coefficient for each historical feature vector using the following formula (5).
[0076] (5);
[0077] In equation (5); Yes, it is the mean of the t-th historical eigenvalues of the k-th historical modal component derived from the q-th historical health parameter decomposition in all historical sliding windows; It is the historical output power corresponding to the m-th historical sliding window; It is the average of the historical output power corresponding to all historical sliding windows;
[0078] Step 4.4: Sort all historical screening coefficients from smallest to largest, and calculate the first historical threshold using the following formula (6).
[0079] (6);
[0080] In equation (6), It is a quantile function; is the set of all historical screening coefficients; b is the threshold coefficient, 85≤b≤95; in this embodiment, b is 90.
[0081] First historical threshold The value after +0.01 is used as the second historical threshold, and the value after +0.01 of the second historical threshold is used as the third historical threshold, and so on to obtain the sixth historical threshold.
[0082] The historical feature vectors corresponding to the historical screening coefficients that are greater than the first historical threshold are used to form the first historical feature vector matrix, the historical feature vectors corresponding to the historical screening coefficients that are greater than the second historical threshold are used to form the second historical feature vector matrix, and so on to obtain the sixth historical feature vector matrix.
[0083] Step 4.5: Establish 6 historical power prediction models. Input the first historical eigenvector matrix and historical output power into the first historical power prediction model for training to obtain the first historical prediction mean square error of the trained first historical power prediction model. Then input the second historical eigenvector matrix and historical output power into the second historical power prediction model for training to obtain the second historical prediction mean square error of the trained second historical power prediction model. And so on to obtain the sixth historical mean square error.
[0084] The historical threshold corresponding to the smallest historical mean square error among the six historical mean square errors is taken as the final historical threshold, and the historical feature vectors corresponding to the historical screening coefficients that are greater than the final historical threshold are used to form the final historical feature vector matrix.
[0085] Mean squared error is a commonly used indicator to measure the deviation between model predictions and actual values. In this embodiment, mean squared error is used to find the optimal threshold, thereby filtering out effective historical feature vectors, which improves the accuracy of power prediction and enhances the reliability of health status assessment.
[0086] Step 5: Input the final historical feature vector matrix into the stacked denoising autoencoder for denoising and dimensionality reduction to obtain the historical fusion feature vector matrix;
[0087] The stacked noise reduction autoencoder (SDAE) used in this embodiment is composed of multiple noise reduction autoencoders stacked layer by layer. For the final historical feature vector matrix of the input, it learns robust and effective features by adding noise to the input features and then reconstructing them. It simultaneously completes the noise reduction and dimensionality reduction of the matrix. The resulting historical fusion feature vector matrix can be directly used for subsequent model training. This is an existing technology for complex industrial data representation learning.
[0088] Step 6: Establish a health parameter prediction model. Input the historical fusion feature vector matrix and historical output power into the health parameter prediction model for training to obtain the trained health parameter prediction model. The input of the health parameter prediction model is set as the fusion feature vector matrix, and the output is set as the health parameter prediction vector composed of 8 health parameter prediction values.
[0089] The health parameter prediction model in this embodiment is constructed using the Transformer deep learning model based on the attention mechanism, and the Particle Swarm Optimization (PSO) algorithm is used to optimize the hyperparameters of the Transformer model during training.
[0090] Step 7: Collect real-time data from the pumped-storage generator motor. Eight real-time health parameters were analyzed using a sliding window method, resulting in a total of [number missing] parameters. Several real-time sliding windows, each containing... Eight real-time health parameters over a period of time;
[0091] In this embodiment, by means of... Figure 1 The sensor layout shown collects real-time health parameters.
[0092] Step 7.1: Decompose all real-time health parameters in each real-time sliding window using the same method as in Step 3 for decomposing all historical health parameters in each historical sliding window, to obtain the... The first real-time sliding window Decomposition of real-time health parameters One real-time modal component ;
[0093] Step 7.2: Process each real-time modal component in the same way as in Steps 4-4.3 to obtain the first eigenvalue of each real-time modal component. Second eigenvalue Third eigenvalue Fourth eigenvalue and the fifth eigenvalue and real-time feature vectors and each real-time feature vector Real-time filtering coefficient ;
[0094] The calculation method for each real-time modal component is the same as that for each historical modal component; simply replace the historical modal component in the corresponding formula with the real-time modal component. However, when calculating the fourth eigenvalue of each real-time modal component according to formula (4), the formula in (4) is different. and It remains unchanged because the health status has not yet been predicted; it is based on historical data under normal conditions. and To calculate the fourth eigenvalue of each real-time modal component.
[0095] Step 7.3 Process the real-time filtering coefficients in the same way as in Step 4.4 to obtain 6 real-time thresholds and 6 real-time feature vector matrices.
[0096] Specifically, the real-time feature vectors corresponding to real-time filtering coefficients greater than the first real-time threshold are used to form the first real-time feature vector matrix, the real-time feature vectors corresponding to real-time filtering coefficients greater than the second real-time threshold are used to form the second real-time feature vector matrix, and so on to obtain the sixth real-time historical feature vector matrix.
[0097] Step 7.4: Process the six real-time feature vector matrices in the same way as the six historical feature vector matrices in Step 4.5 to obtain six real-time mean square errors and the final real-time feature vector matrix;
[0098] Specifically, the real-time threshold corresponding to the smallest real-time mean square error among the six real-time mean square errors is taken as the final real-time threshold, and the real-time feature vectors corresponding to the real-time filtering coefficients that are greater than the final real-time threshold are used to form the final real-time feature vector matrix.
[0099] Step 8: Input the final real-time feature vector matrix into the stacked denoising autoencoder for denoising and dimensionality reduction to obtain the real-time fused feature vector matrix;
[0100] Step 9: Input the real-time fused feature vector matrix into the trained health parameter prediction model and combine it with the Monte Carlo method for prediction, outputting B health parameter prediction value vectors;
[0101] The Monte Carlo method (MC-Dropout) used in this embodiment enables Dropout layers (randomly discarding neurons) multiple times during the model prediction phase. It uses the statistical characteristics of multiple sampling results to quantify the model's prediction uncertainty. It belongs to the existing statistical simulation technology based on random sampling. In this embodiment, it is used in conjunction with the prediction model to generate B vectors of predicted health parameters.
[0102] The real-time health value L of the pumped storage generator motor is calculated according to formula (7).
[0103] (7);
[0104] In equation (7), Y1 is the b-th prediction residual vector; Y1 is a real-time health parameter vector composed of 8 real-time health parameters. It is the vector of predicted values for the b-th health parameter; It is the average vector of B predicted residual vectors; It is the first predicted residual vector;
[0105] If 0 < L ≤ 1, the health status of the pumped storage generator motor is assessed as normal.
[0106] If 1 < L ≤ 2, the health status of the pumped storage generator motor is assessed as a state of concern.
[0107] If L > 2, the health status of the pumped storage generator motor is assessed as abnormal.
[0108] In this embodiment, the real-time health value L=0.5 of the pumped storage generator motor is calculated according to formula (7), so the health status of the pumped storage generator motor is assessed as normal.
[0109] The above description is only a preferred embodiment of the present invention, but the present invention is not limited thereto. All equivalent substitutions or modifications made to the concepts and technical solutions of the present invention should be covered within the protection scope of the present invention.
Claims
1. A method of assessing the health state of a pumped hydro power generator motor, characterized by, The method comprises the following steps: Step 1: setting three health states of the pumped storage power generation motor as normal state, attention state and abnormal state respectively; Setting 8 health parameters affecting the health state of the pumped storage power generation motor as the stator current, stator voltage, vibration state value, stator winding temperature, bearing temperature, output power, and upstream and downstream water level difference and pump guide vane opening of the pumped storage power generation motor respectively; Step 2: obtaining 8 historical health parameters of the pumped storage power generation motor under three historical health states for T time respectively, and dividing the 8 historical health parameters under the three historical health states by using a sliding window method, so as to form M historical sliding windows, each of which contains 8 historical health parameters of a time length for T time under any historical health state; Step 3: decomposing all historical health parameters in each historical sliding window according to the following formula (1) （1）； In formula (1), is an adaptive noise complete ensemble empirical mode decomposition cascaded variational mode decomposition; is the qth historical health parameter in the mth historical sliding window; is the kth historical modal component decomposed from the qth historical health parameter in the mth historical sliding window; Step 4: Calculate the first historical eigenvalue, the second historical eigenvalue and the third historical eigenvalue of each historical modal component according to the following formula (2) respectively （2）； In formula (2), is The total number of sampling points within T time; is The amplitude of the i-th sampling point; is The average value of the amplitude of is The total number of frequencies of is The j-th frequency of is The amplitude on the j-th frequency; Step 4.1: Calculate the fourth historical eigenvalue of each historical modal component according to the following formula (3) （3）； In formula (3), is a fusion vector of the historical modal components decomposed from the historical health parameters in the historical normal state in all the historical sliding windows, T is a transpose; is a mean vector of the fusion vectors of the historical modal components decomposed from the historical health parameters in the historical normal state in all the historical sliding windows; is a covariance matrix of the fusion vectors of the historical modal components decomposed from the historical health parameters in the historical normal state in all the historical sliding windows. Step 4.2: Sample entropy of each historical modality component as its own fifth historical feature value ; The tth historical eigenvalue of the kth historical modal component of the qth historical health parameter decomposed from all historical sliding windows is respectively constructed into a historical feature vector according to the following formula (4) （4）； In formula (4), is the tth historical eigenvalue of the kth historical modal component of the qth historical health parameter in the mth historical sliding window. Step 4.3: Calculate the historical screening coefficient of each historical feature vector according to the following formula (5) （5）； In formula (5); is the mean value of the tth historical eigenvalue of the kth historical modal component of the qth historical health parameter in all historical sliding windows; is the historical output power corresponding to the mth historical sliding window; is the mean value of the historical output power corresponding to all historical sliding windows; Step 4.4: Sort all the historical screening coefficients from small to large, calculate the first historical threshold value according to the following formula (6) （6）； In formula (6), is a quantile function; is a set of all historical screening coefficients; b is a threshold coefficient, 85≤b≤95; First historical threshold The value after +0.01 is used as the second historical threshold, and the value after +0.01 of the second historical threshold is used as the third historical threshold, and so on to obtain the sixth historical threshold. The historical feature vectors corresponding to the historical screening coefficients greater than the first historical threshold value are combined to form a first historical feature vector matrix, the historical feature vectors corresponding to the historical screening coefficients greater than the second historical threshold value are combined to form a second historical feature vector matrix, and the sixth historical feature vector matrix is obtained in the same manner; Step 4.5: establishing 6 historical power prediction models, inputting the first historical feature vector matrix and the historical output power into the first historical power prediction model for training to obtain the first historical prediction mean square error of the trained first historical power prediction model, inputting the second historical feature vector matrix and the historical output power into the second historical power prediction model for training to obtain the second historical prediction mean square error of the trained second historical power prediction model, and obtaining the sixth historical mean square error in the same manner; The historical threshold value corresponding to the minimum historical mean square error in the 6 historical mean square errors is taken as the final historical threshold value, and the historical feature vectors corresponding to the historical screening coefficients greater than the final historical threshold value are combined to form a final historical feature vector matrix; Step 5: inputting the final historical feature vector matrix into a stacked denoising autoencoder for denoising and dimension reduction to obtain a historical fusion feature vector matrix; Step 6: establishing a health parameter prediction model, inputting the historical fusion feature vector matrix and the historical output power into the health parameter prediction model for training to obtain the trained health parameter prediction model; the input of the health parameter prediction model is set as the fusion feature vector matrix, and the output is set as a health parameter prediction vector composed of 8 health parameter prediction values; Step 7: Collecting 8 real-time health parameters of the pumped storage power generation motor in real time The 8 real-time health parameters are divided by using a sliding window method, and a total of real-time sliding windows are formed, each of which contains 8 real-time health parameters in a time period. Step 7.1: Decompose all real-time health parameters in each real-time sliding window using the same method as in Step 3 for decomposing all historical health parameters in each historical sliding window, to obtain the... The first real-time sliding window Decomposition of real-time health parameters One real-time modal component ; Step 7.2: Each real-time modality component is processed in the same way as each historical modality component in Step 4-4.3 to obtain a first eigenvalue , a second eigenvalue , a third eigenvalue , a fourth eigenvalue , and a fifth eigenvalue of each real-time modality component, and a real-time feature vector and a real-time screening coefficient of each real-time feature vector ; Step 7.3: the historical screening coefficients are processed in the same manner as step 4.4 to obtain 6 real-time threshold values and 6 real-time feature vector matrices; Step 7.4: the 6 real-time feature vector matrices are processed in the same manner as step 4.5 to obtain 6 real-time mean square errors and a final real-time feature vector matrix; Step 8: inputting the final real-time feature vector matrix into a stacked denoising autoencoder for denoising and dimension reduction to obtain a real-time fusion feature vector matrix; Step 9: input the real-time fusion feature vector matrix into the trained health parameter prediction model and combine the Monte Carlo method for prediction, and output B health parameter prediction value vectors; The real-time health value L of the pumped storage power generation motor is calculated according to formula (7) （7）； In formula (7), is the bth predicted residual vector; Y1 is a real-time health parameter vector composed of 8 real-time health parameters; is the bth health parameter predicted value vector; is the average vector of B predicted residual vectors; is the first predicted residual vector; If 0 < L ≤ 1, the health status of the pumped storage power generation motor is evaluated as a normal state; If 1 < L ≤ 2, the health status of the pumped storage power generation motor is evaluated as an attention state; If L > 2, the health status of the pumped storage power generation motor is evaluated as an abnormal state.
2. The method of claim 1, wherein: The health parameter prediction model in step 6 is established by a deep learning model transformer based on an attention mechanism.
3. The method of claim 1, wherein: The health parameter prediction model in step 6 is trained by using a particle swarm optimization algorithm.