Hydroelectric generating set state evaluation method based on HHO-VMD-CNN-BiGRU

Abstract

A hydroelectric generating set state evaluation method based on HHO-VMD-CNN-BiGRU comprises the steps that firstly, an improved variational mode decomposition method for adaptively optimizing decomposition parameters through a new comprehensive evaluation index of cosine similarity-kurtosis-envelope entropy is proposed, and decomposed effective mode components are screened through the index; then, in order to extract more comprehensive fault features, a new composite scaling permutation entropy is introduced to optimize the fault features of each effective modal component; and finally, on the basis of a relay protection mining model of CNN-BiGRU (Convolutional Neural Network Recurrent Unit) deep learning, capturing typical characteristic quantities of state data through a CNN layer, then mining a time sequence dependency relationship by using a BiGRU model, and completing assessment of the health state of the hydroelectric generating set.

Description

Technical Field

[0001] The invention belongs to the technical field of hydropower unit state assessment, and in particular relates to a hydropower unit state assessment method based on HHO-VMD-CNN-BiGRU. Background Art

[0002] Hydropower units are the core equipment in the construction of hydropower stations. Their stable, safe and reliable operation is crucial for power plant management and maintenance personnel. This is not only related to the safety of the power station itself, but also directly affects whether the power grid can continue to provide stable power support for social production. During operation, hydropower units are affected by multiple factors such as electromagnetic, mechanical and hydraulic. As the running time of the unit accumulates, the equipment may experience abnormal vibration, leading to fatigue, damage or even deterioration, thus affecting the safe and reliable operation of the unit. This situation may pose a serious threat to the safety and stability of the power grid and may cause significant social and economic losses. Under the constraints of water-machine-electricity-magnetism multiple coupling, the development of safe operation control theory of hydropower units is the key to ensuring the safety of water conservancy projects. This is of great significance for improving the efficiency of water resource utilization, ensuring the ecological water demand downstream of hydropower stations, and promoting the development of the national economy. In the long-term operation process, hydropower units will go through a process from healthy state, performance degradation to equipment failure and even failure. If the operating status of the unit can be monitored in real time during the unit degradation process, the health status changes can be evaluated, and the future degradation trend can be predicted, it will help to detect abnormalities in time, formulate reasonable maintenance plans, and truly realize the condition maintenance of hydropower units, thereby maximizing the production benefits of power generation companies. As a key link in unit maintenance, the health status assessment and performance degradation trend prediction of hydropower units have important research value and application prospects for improving the stable operation level of hydropower units and reducing the downtime and maintenance losses caused by predictable failures.

[0003] Therefore, in order to ensure the safe and stable operation of the unit equipment, it is urgent to carry out research on the prediction of the performance degradation trend of the hydropower unit. Based on the massive state monitoring data of the hydropower unit, the high-dimensional function mapping relationship between the unit operating parameters and the unit operating status is analyzed, the health status of the hydropower unit equipment is grasped in real time, the health performance of the hydropower unit historical operation process is evaluated, and the future health status of the unit is predicted and evaluated, the abnormal state of the unit is discovered in time, the unit fault is judged scientifically and reasonably, the corresponding maintenance and active maintenance are arranged, and decision-making technical support is provided for the fault detection and active maintenance of the unit, thereby ensuring the safe and stable operation of the hydropower unit.

[0004] The existing degradation trend prediction methods for hydropower units based on vibration signals have the problem that the selected health indicators are too local and single, which makes it difficult for the degradation trend prediction results to reflect the overall operating status of the unit; under complex operating conditions, the degradation trend prediction model is not accurate enough to accurately obtain the degradation trend at future moments; in addition, the degradation prediction model ignores the influence of historical operating parameters themselves on the degradation trend prediction, which further limits the prediction effect and robustness of the model. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a hydropower unit state assessment method based on HHO-VMD-CNN-BiGRU, through which the effective modal components after decomposition are screened; then, in order to extract more comprehensive fault features, a new composite scaled permutation entropy is introduced to quantify each effective modal component.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU, the steps are as follows: Step 1): Establish an improved variational mode decomposition method that adaptively optimizes decomposition parameters using the comprehensive evaluation index of "cosine similarity-kurtosis-envelope entropy"; Step 2): Decompose the vibration signal into several modal components through an improved algorithm; Step 3): Decompose the vibration signal into effective parts and redundant parts, and use comprehensive evaluation indicators to accurately screen out the effective components; Step 4): A new composite scaled permutation entropy is introduced to quantify the fault characteristics of each effective modal component, thereby extracting more comprehensive fault characteristics. Preferably, in step 1), a comprehensive evaluation index that takes into account both periodic characteristics and impulse characteristics is adopted by combining the advantages of envelope entropy, kurtosis and cosine similarity. , as shown in formula (1): (1); In the formula, is the IMF envelope entropy; is the kurtosis of each IMF; is the cosine similarity between each IMF and the initial signal; , are the weight coefficients of envelope entropy and kurtosis, respectively, and + =1, since the envelope entropy and kurtosis index have similar sensitivity to the signal decomposition results, = =0.5.

[0007] Preferably, The value of is shown in formula (2): = (2); In the formula, , They are IMF signal and initial signal respectively; The value of is shown in formula (3): (3); In the formula, is the IMF signal, is the mean of the IMF signal, is the standard deviation of the IMF signal, and E is the mathematical expectation; The value of is shown in formula (4): (4); The value of is shown in formula (5): (5); In the formula, is the envelope signal after Hilbert demodulation; for The normalized form of .

[0008] Preferably, in step 2), the steps of performing modal decomposition of the original signal using the improved algorithm are as follows: Step 2.1: Initialize HHO optimization algorithm parameters; Step 2.2: There are two parameters that the HHO optimization algorithm needs to extract adaptively. Set all populations to a 2-dimensional vector , each population represents a parameter that needs to be optimized, where Indicates the number of modal components , the value range is [3,10], represents the quadratic penalty factor , the value range is [500,6000]; Step 2.3: Use the Soft-max function normalization method to normalize the envelope entropy and kurtosis index so that they are between [0,1]. The definition of Soft-max function normalization is shown in formula (6): (6); In the formula, is the inverse of the envelope entropy or the original value of kurtosis; is the normalized value of the inverse of the envelope entropy or the normalized value of the kurtosis; is the number of modal components; Step 2.4: Use the normalized index obtained in step 3 to calculate the fitness value of the Harris Hawk population and find the initial population position with the minimum fitness value; Step 2.5: Update the population position through the three optimization stages of the population, iterate continuously, recalculate the fitness value, and select the minimum fitness value to determine the optimal population position; Step 2.6: If the current iteration number Greater than the maximum number of iterations , then the iteration ends and the optimal parameters of VMD are obtained and ; Otherwise, return to step 5. , until the judgment condition is met.

[0009] Preferably, in step 3), the original signal is decomposed into IMFs, calculate the comprehensive evaluation index of each IMF and The average value of the comprehensive evaluation indicators of the IMF ; like , then the IMF is a valid component; otherwise it is considered a redundant component; This criterion is used to screen the effective IMF after decomposition of each sample. Analysis shows that the number of effective IMFs obtained by screening each sample is not exactly the same. The size of the number of effective IMFs of each sample is compared, and the maximum value is set to , the number of remaining valid IMFs is less than The sample of the value is expanded by overlapping the effective IMF method, so that the number of effective IMFs is equal to .

[0010] Preferably, in step 4), the newly proposed composite scaling permutation entropy uses multiple wavelets to synchronously decompose the original signal into a spectrum scaling time series, which overcomes the shortcomings of multi-scale permutation entropy and hierarchical permutation entropy, enriches fault features with multiple resolutions over the entire frequency band, and improves the stability of feature extraction; Extract each sample in step 3) features, and the feature vector is ,in, is the total number of samples); select samples are training samples, and their feature vectors are ; test samples, whose feature vector is ; Normalize the feature vector and use it as the input of the CNN-BiGRU network.

[0011] A method for mining the status assessment of hydropower units based on CNN-BiGRU deep learning is proposed. The convolutional network layer CNN is used to capture typical feature quantities of status data, and then the BiGRU model is used to mine the temporal relationship to complete the evaluation of the health status of the hydropower unit. The twin assessment structure of the relay protection health status based on CNN-BiGRU deep learning includes an input layer, a CNN layer, a BiGRU layer, a pooling layer, and a fully connected layer. The process of implementing the health status assessment of the hydropower unit through this model is as follows: Step 1: The input layer is the original state indicator matrix to be evaluated in a time-series continuous manner, and the CNN layer completes the input vector convolution operation through the convolution kernel; Step 2: After analyzing the indicator features, CNN outputs them to BiGRU to complete the storage and operation of the time series input vector; Step 3: The pooling layer uses the k-max pooling method to reduce the dimension of the input vector to obtain a more representative vector; Step 4: The fully connected layer performs a weighted operation on the input vector after step 3, as shown in formula (8) After the function is nonlinearly calculated, the output is a vector consisting of the probabilities of the four categories of "normal", "caution", "abnormal" and "severe abnormality", and the category with the highest probability is determined as the evaluation result.

[0012] Preferably, in step 1, the calculation process is as follows: Assume that the size of a single convolution kernel is , the single indicator value to be processed is , the weight, bias and output assigned to the convolution kernel index value are set as , , Then we have formula (7) and formula (8): (7); (8).

[0013] Preferably, in step 2, the BiGRU layer is designed by stacking multiple GRU neurons in sequence, and the neurons use a reset gate With update gate , combined with formula (9) to complete the storage and operation of the time series input vector: (9); Where: is the sigmoid function, and They are the hidden state vectors of this GRU output and the previous GRU transmission, is the input vector; , , They are the weights assigned by the reset gate, update gate and the hidden state of this GRU to the input vector. , , is the corresponding bias; , They are the weights assigned by the reset gate and the update gate to the hidden state vector passed by the previous GRU; Output hidden state vector of this GRU to hidden state vector of this GRU The weight assigned by the Hadamard product of the hidden state vector passed by the previous GRU.

[0014] Preferably, in step 3), the pooling layer uses the k-max pooling method to reduce the dimension of the input vector to obtain a more representative vector. The pooling process is shown in formula (10): (10); In the formula = is the input of the pooling layer, is the output; In step 4), the fully connected layer performs a weighted operation on the input vector, as shown in formula (8) After the function is nonlinearly calculated, the output is a vector consisting of the probabilities of the four categories of "normal", "caution", "abnormal" and "severe abnormality", and the category with the highest probability is determined as the evaluation result.

[0015] The present invention can achieve the following beneficial effects: 1. The present invention uses the index to screen the effective modal components after decomposition; then, in order to extract more comprehensive fault features, a new composite scaling permutation entropy is introduced to quantify each effective modal component.

[0016] 2. The newly proposed composite scaled permutation entropy (CZPE) of the present invention uses multiple wavelets to synchronously decompose the original signal into a spectrum-scaled time series, which overcomes the shortcomings of multi-scale permutation entropy and hierarchical permutation entropy, can enrich fault features with multiple resolutions over the entire frequency band, and improve the stability of feature extraction. The advantages of CZPE are twofold: (1) Compared with MPE and HPE, CZPE uses multiple wavelets to synchronously decompose the original signal into a spectrum-scaled time series (SZTS). This scaling process can enrich fault features with multiple resolutions over the entire frequency band. (2) Compared with CMPE, CZPE provides a physical explanation for composite analysis, improves the stability of extracted features, and makes the performance of CZPE better than other methods.

[0017] 3. The invention proposes a hydropower unit status assessment mining model based on CNN-BiGRU deep learning, which captures typical feature quantities of status data through the convolutional network layer CNN, and then uses the BiGRU model to mine the temporal relationship to complete the evaluation of the health status of the hydropower unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is the flow chart of CNN-BiGRU of the present invention; Figure 2 The internal structure diagram of BiGRU of the present invention; Figure 3 The data source of the present invention and the distribution diagram of vibration measurement points of the hydropower unit; Figure 4 This is a graph showing the accuracy of health status assessment of the training set and the validation set of the simulation results of the present invention; Figure 5 Health status assessment of the test set for the present invention and Indicator value display chart; Figure 6 The present invention and other different model test sets and Comparison chart. DETAILED DESCRIPTION

[0019] The preferred solution is Figures 1 to 5 As shown in FIG. 1 , a method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU is shown in FIG. 1 , and the steps are as follows: Step 1): First, an improved variational mode decomposition method is proposed which adaptively optimizes the decomposition parameters by using the new comprehensive evaluation index of “cosine similarity-kurtosis-envelope entropy”.

[0020] Step 2): Decompose the vibration signal into several modal components through an improved algorithm.

[0021] Step 3): Decompose the vibration signal into effective parts and redundant parts, and use comprehensive evaluation indicators to accurately screen out the effective components; Step 4): In order to extract more comprehensive fault features, a new composite scaled permutation entropy is introduced to quantify the fault features of each effective modal component.

[0022] Furthermore, in step 1), the main thing is to determine the fitness function of the algorithm. The envelope entropy can reflect the sparsity of the signal. If the signal sparsity is very weak, the envelope entropy value is very large; conversely, if the signal contains a lot of fault information, the signal sparsity is very strong and the envelope entropy value is very small. However, the envelope entropy can only reflect the periodic characteristics of the signal. The actual vibration signal of the rotating mechanical bearing has not only periodic characteristics but also impact characteristics. Considering that the kurtosis is very sensitive to fault impact, the larger its value is, the more obvious the signal impact phenomenon and fault information are; cosine similarity can characterize the similarity between each component and the original signal. The closer the cosine similarity value is to 1, the closer the component is to the original signal. This paper combines the advantages of envelope entropy, kurtosis and cosine similarity, and proposes a comprehensive evaluation index that takes into account both periodic characteristics and impact characteristics. , as shown in formula (1): (1) ; In the formula, is the IMF envelope entropy; is the kurtosis of each IMF; is the cosine similarity between each IMF and the initial signal; , are the weight coefficients of envelope entropy and kurtosis, respectively, and + =1, since the envelope entropy and kurtosis index have similar sensitivity to the signal decomposition results, = =0.5. Among them, The value of is shown in formula (2): = (2); In the formula, , are the IMF signal and the initial signal respectively.

[0023] The value of is shown in formula (3): (3); In the formula, is the IMF signal, is the mean of the IMF signal, is the standard deviation of the IMF signal, and E is the mathematical expectation.

[0024] The value of is shown in formula (4): (4); The value of is shown in formula (5): (5); In the formula, is the envelope signal after Hilbert demodulation; for The normalized form of .

[0025] Furthermore, in step 2), the steps of performing modal decomposition of the original signal using the improved algorithm are as follows: Step 2.1: Initialize HHO optimization algorithm parameters; Step 2.2: There are two parameters that the HHO optimization algorithm needs to extract adaptively. Set all populations to a 2-dimensional vector , each population represents a parameter that needs to be optimized, where Indicates the number of modal components , the value range is [3,10], represents the quadratic penalty factor , the value range is [500,6000]; Step 2.3: Use the Soft-max function normalization method to normalize the envelope entropy and kurtosis index so that they are between [0,1]. The definition of Soft-max function normalization is shown in formula (6): (6) In the formula, is the inverse of the envelope entropy or the original value of kurtosis; is the normalized value of the inverse of the envelope entropy or the normalized value of the kurtosis; is the number of modal components.

[0026] Step 2.4: Use the normalized index obtained in step 3 to calculate the fitness value of the Harris Hawk population and find the initial population position with the minimum fitness value; Step 2.5: Update the population position through the three optimization stages of the population, iterate continuously, recalculate the fitness value, and select the minimum fitness value to determine the optimal population position; Step 2.6: If the current iteration number Greater than the maximum number of iterations , then the iteration ends and the optimal parameters of VMD are obtained and ; Otherwise, return to step 5. , until the judgment condition is met.

[0027] Further, in step 3), the original signal is decomposed into IMFs, calculate the comprehensive evaluation index of each IMF and The average value of the comprehensive evaluation indicators of the IMF .like , then the IMF is a valid component; otherwise it is considered a redundant component. This criterion is used to screen the valid IMF after each sample is decomposed. Analysis shows that the number of valid IMFs screened from each sample is not exactly the same. By comparing the number of valid IMFs for each sample, the maximum value is set to , the number of remaining valid IMFs is less than The sample of the value is expanded by overlapping the effective IMF method, so that the number of effective IMFs is equal to .

[0028] Furthermore, in step 4), the newly proposed composite scaled permutation entropy (CZPE) uses multiple wavelets to synchronously decompose the original signal into a spectral scaled time series, which overcomes the shortcomings of multi-scale permutation entropy and hierarchical permutation entropy, can enrich fault features with multiple resolutions across the entire frequency band, and improve the stability of feature extraction. The advantages of CZPE are twofold: (1) Compared with MPE and HPE, CZPE uses multiple wavelets to synchronously decompose the original signal into a spectral scaled time series (SZTS). This scaling process can enrich fault features with multiple resolutions across the entire frequency band. (2) Compared with CMPE, CZPE provides a physical explanation for composite analysis, improves the stability of extracted features, and makes the performance of CZPE better than other methods. Extract each sample in step 3) features, and the feature vector is (in, is the total number of samples). Select samples are training samples, and their feature vectors are ; test samples, whose feature vector is . Normalize the feature vector and use it as the input of the CNN-BiGRU network.

[0029] On the other hand, the present invention proposes a hydropower unit status assessment mining model based on CNN-BiGRU deep learning, which captures typical feature quantities of status data through the convolutional network layer CNN, and then uses the BiGRU model to mine the temporal relationship to complete the evaluation of the health status of the hydropower unit. The twin assessment structure of the health status of the relay protection based on CNN-BiGRU deep learning is as follows: Figure 1 As shown in the figure, it includes input layer, CNN layer, BiGRU layer, pooling layer and fully connected layer. The health status assessment of hydropower units through this model mainly consists of the following steps: Step 1) The input layer is a time-series continuous matrix of original state indicators to be evaluated, and the CNN layer completes the input vector convolution operation through the convolution kernel.

[0030] Step 2) After CNN analyzes the indicator features, it outputs them to BiGRU to complete the storage and operation of the time series input vector Step 3) The pooling layer uses the k-max pooling method to reduce the dimension of the input vector to obtain a more representative vector.

[0031] Step 4) The fully connected layer performs a weighted operation on the input vector after step 3, as shown in formula (8) After the function is nonlinearly calculated, the output is a vector consisting of the probabilities of the four categories of "normal", "caution", "abnormal" and "severe abnormality", and the category with the highest probability is determined as the evaluation result.

[0032] In step 1), the calculation process is as follows: Assume that the size of a single convolution kernel is , the single indicator value to be processed is , the weight, bias and output assigned to the convolution kernel index value are set as , , Then we have formula (7) and formula (8): (7); (8); In step 2), the BiGRU layer is designed by stacking multiple GRU neurons in sequence, such as Figure 1 As shown, a single GRU structure is as follows Figure 2 As shown. Neurons use reset gates With update gate , combined with formula (9), the storage and operation of the time series input vector are completed.

[0033] (9); Where: is the sigmoid function, and They are the hidden state vectors of this GRU output and the previous GRU transmission, is the input vector; , , They are the weights assigned by the reset gate, update gate and the hidden state of this GRU to the input vector. , , is the corresponding bias; , They are the weights assigned by the reset gate and the update gate to the hidden state vector passed by the previous GRU; Output hidden state vector of this GRU to hidden state vector of this GRU The weight assigned by the Hadamard product of the hidden state vector passed by the previous GRU.

[0034] In step 3), the pooling layer uses the k-max pooling method to reduce the dimension of the input vector to obtain a more representative vector. The pooling process is shown in formula (10): (10); In the formula = is the input of the pooling layer, is the output.

[0035] In step 4), the fully connected layer performs a weighted operation on the input vector, as shown in formula (8) After the function is nonlinearly calculated, the output is a vector consisting of the probabilities of the four categories of "normal", "caution", "abnormal" and "severe abnormality", and the category with the highest probability is determined as the evaluation result.

[0036] In order to test the performance of the above model, a two-class evaluation method was used to verify the feasibility of the model. In the two-class evaluation problem, the accuracy

[0037] is the number of positive examples in the sample The ratio of all positive examples to the recall rate Refers to the number of positive examples in the sample that are judged as positive examples The ratio of all positive examples, and The upper and lower limits of the model reliability are measured respectively, and their values ​​are positively correlated with the reliability. In order to achieve a comprehensive consideration of the two, the harmonic mean of accuracy and recall is often used for binary classification problems. As a model performance evaluation indicator, as shown in formula (11): (11); Since the health assessment task in this paper is a four-classification judgment task, this paper selects the multi-classification task model measurement index and As shown in formula (12): , (12) The correlation between the evaluation indicators is shown in Table 3-1. Ideally, the accuracy and recall rates are both 100%, and The closer it is to 100%, the better the overall performance of the model.

[0038]

[0039] The experimental data comes from the No. 4 axial-flow propeller turbine generator set of a power plant. Since most of the hydroelectric generator sets in the hydropower plant are of the same type and rated capacity, and the impact of other very small parts of different types of hydroelectric generator sets can be ignored, this paper selects the monitoring data of a hydroelectric generator set to evaluate the operating risk status of the hydroelectric generator set. The unit is equipped with several pressure point sensors, three pendulum sensors, five vibration sensors, and four pressure point sensors. The vibration monitoring points are distributed as follows: Figure 3 As shown. In this experiment, the X and Y horizontal vibration data measured at four vibration points, namely the upper frame, lower frame, top cover and oil tank, were selected. The data measured at these four vibration points were feature extracted and aggregated. They were input into the evaluation model, and the first 80% of the data was used for training, and the last 20% was used for testing. The model hyperparameter settings are shown in Table 3

[0040] Figure 4 and Figure 5 The evaluation accuracy of the twin assessment model for the health status of hydropower units built based on the above parameters on the training set and the validation set is shown. As the number of training times increases, the classification accuracy of the training set and the validation set gradually converges. The final evaluation accuracy of the training set is 99.68%, and the final evaluation accuracy of the validation set is 94.76%. and The values ​​were 93.98% and 94.19% respectively. The fluctuation range was within 2%. On this basis, this paper further plotted the subject curve of the model calculation results. The larger the area under the curve, the better the evaluation performance, and the maximum value is 100%. States 1 to 4 in the figure represent "normal", "caution", "abnormal" and "severe abnormality" respectively. Figure 4 It can be seen that the model and The AUCs of the two ROC curves are 94% and 99% respectively, which are very close to 100%, proving that the model has excellent performance in assessing the health status of hydropower units.

[0041] In order to verify the superiority of the CNN-BiGRU hydropower unit health status assessment model, this paper selected five typical machine learning and shallow neural network models that have been applied to equipment health status assessment, namely CNN, BiLSTM, MLP, BiGRU, SVM, Naive Bayes, Logistic Regression and Random Forest, for comparison. Among them, the number of neurons in LSTM and BiLSTM is consistent with the number of neurons in the BiGRU layer in the model proposed in this paper, and the convolution parameters of CNN are consistent with the CNN layer parameters of the model in this paper. The models are updated based on the training set and the validation set, and the test set is used. and The experimental results are as follows: Figure 6 As shown in Figure 2, it can be seen that the proposed model is superior to other models in terms of convergence speed and verification accuracy in the state assessment process, and has excellent assessment capabilities.

[0042] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall 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. A hydropower unit state assessment method based on HHO-VMD-CNN-BiGRU, characterized in that The following steps are involved: Step 1): Establish an improved variational mode decomposition method for adaptively optimizing decomposition parameters using the comprehensive evaluation index of "cosine similarity-kurtosis-envelope entropy"; Step 2): Decompose the vibration signal into several modal components through an improved algorithm; Step 3): Decompose the vibration signal into effective parts and redundant parts, and use comprehensive evaluation indicators to accurately screen out the effective components; Step 4): A new composite scaled permutation entropy is introduced to quantify the fault characteristics of each effective modal component, thereby extracting more comprehensive fault characteristics.

2. The method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU according to claim 1 is characterized in that: In step 1), a comprehensive evaluation index that takes into account both periodic characteristics and impulse characteristics is adopted by combining the advantages of envelope entropy, kurtosis and cosine similarity. , as shown in formula (1): （1）； In the formula, is the IMF envelope entropy; is the kurtosis of each IMF; is the cosine similarity between each IMF and the initial signal; , are the weight coefficients of envelope entropy and kurtosis, respectively, and + =1, since the envelope entropy and kurtosis index have similar sensitivity to the signal decomposition results, = =0.

5.

3. The method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU according to claim 1 is characterized in that: The value of is shown in formula (2): = （2）； In the formula, , They are IMF signal and initial signal respectively; The value of is shown in formula (3): （3）； In the formula, is the IMF signal, is the mean of the IMF signal, is the standard deviation of the IMF signal, and E is the mathematical expectation; The value of is shown in formula (4): （4）； The value of is shown in formula (5): （5）； In the formula, is the envelope signal after Hilbert demodulation; for The normalized form of .

4. The method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU according to claim 1 is characterized in that: In step 2), the steps of modal decomposition of the original signal using the improved algorithm are as follows: Step 2.1: Initialize HHO optimization algorithm parameters; Step 2.2: There are two parameters that the HHO optimization algorithm needs to extract adaptively. Set all populations to a 2-dimensional vector , each population represents a parameter that needs to be optimized, where Indicates the number of modal components , the value range is [3,10], represents the quadratic penalty factor , the value range is [500,6000]; Step 2.3: Use the Soft-max function normalization method to normalize the envelope entropy and kurtosis index so that they are between [0,1]. The definition of Soft-max function normalization is shown in formula (6): （6）； In the formula, is the inverse of the envelope entropy or the original value of kurtosis; is the normalized value of the inverse of the envelope entropy or the normalized value of the kurtosis; is the number of modal components; Step 2.4: Use the normalized index obtained in step 3 to calculate the fitness value of the Harris Hawk population and find the initial population position with the minimum fitness value; Step 2.5: Update the population position through the three optimization stages of the population, iterate continuously, recalculate the fitness value, and select the minimum fitness value to determine the optimal population position; Step 2.6: If the current iteration number Greater than the maximum number of iterations , then the iteration ends and the optimal parameters of VMD are obtained and ; Otherwise, return to step 5. , until the judgment condition is met.

5. The method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU according to claim 3 is characterized in that: In step 3), the original signal is decomposed into IMFs, calculate the comprehensive evaluation index of each IMF and The average value of the comprehensive evaluation indicators of the IMF ; like , then the IMF is a valid component; otherwise it is considered a redundant component; This criterion is used to screen the effective IMF after decomposition of each sample. Analysis shows that the number of effective IMFs obtained by screening each sample is not exactly the same. The size of the number of effective IMFs of each sample is compared, and the maximum value is set to , the number of remaining valid IMFs is less than The sample of the value is expanded by overlapping the effective IMF method, so that the number of effective IMFs is equal to .

6. The method for evaluating the state of a hydropower unit based on HHO-VMD-CNN-BiGRU according to claim 1 is characterized in that: In step 4), the newly proposed composite scaling permutation entropy uses multiple wavelets to synchronously decompose the original signal into a spectrum scaling time series, which overcomes the shortcomings of multi-scale permutation entropy and hierarchical permutation entropy, enriches fault features with multiple resolutions over the entire frequency band, and improves the stability of feature extraction; Extract each sample in step 3) features, and the feature vector is ,in, is the total number of samples); select samples are training samples, and their feature vectors are ; test samples, whose feature vector is ; Normalize the feature vector and use it as the input of the CNN-BiGRU network.

7. A hydropower unit state assessment mining method based on CNN-BiGRU deep learning, characterized by: The convolutional network layer CNN is used to capture the typical feature quantities of the status data, and then the BiGRU model is used to mine the temporal relationship to complete the evaluation of the health status of the hydropower unit; the relay protection health status twin evaluation structure based on CNN-BiGRU deep learning includes input layer, CNN layer, BiGRU layer, pooling layer and fully connected layer; the process of realizing the health status evaluation of the hydropower unit through this model is as follows: Step 1: The input layer is the original state indicator matrix to be evaluated in a time-series continuous manner, and the CNN layer completes the input vector convolution operation through the convolution kernel; Step 2: After analyzing the indicator features, CNN outputs them to BiGRU to complete the storage and operation of the time series input vector; Step 3: The pooling layer uses the k-max pooling method to reduce the dimension of the input vector to obtain a more representative vector; Step 4: The fully connected layer performs a weighted operation on the input vector after step 3, as shown in formula (8) After the function is nonlinearly calculated, the output is a vector consisting of the probabilities of the four categories of "normal", "caution", "abnormal" and "severe abnormality", and the category with the highest probability is determined as the evaluation result.

8. A method for hydropower unit state assessment and mining based on CNN-BiGRU deep learning according to claim 7, characterized in that: In step 1, the calculation process is as follows: Assume that the size of a single convolution kernel is , the single indicator value to be processed is , the weight, bias and output assigned to the convolution kernel index value are set as , , Then we have formula (7) and formula (8): （7）； （8）。 9. The method for hydropower unit state assessment and mining based on CNN-BiGRU deep learning according to claim 7 is characterized in that: In step 2, the BiGRU layer is designed by stacking multiple GRU neurons in sequence, and the neurons use the reset gate With update gate , combined with formula (9) to complete the storage and operation of the time series input vector: （9）； Where: is the sigmoid function, and They are the hidden state vectors of this GRU output and the previous GRU transmission, is the input vector; , , They are the weights assigned by the reset gate, update gate and the hidden state of this GRU to the input vector. , , is the corresponding bias; , They are the weights assigned by the reset gate and the update gate to the hidden state vector passed by the previous GRU; Output hidden state vector of this GRU to hidden state vector of this GRU The weight assigned by the Hadamard product of the hidden state vector passed by the previous GRU.

10. A method for hydropower unit state assessment and mining based on CNN-BiGRU deep learning according to claim 7, characterized in that: In step 3), the pooling layer uses the k-max pooling method to reduce the dimension of the input vector to obtain a more representative vector. The pooling process is shown in formula (10): （10）； In the formula = is the input of the pooling layer, is the output; In step 4), the fully connected layer performs a weighted operation on the input vector, as shown in formula (8) After the function is nonlinearly calculated, the output is a vector consisting of the probabilities of the four categories of "normal", "caution", "abnormal" and "severe abnormality", and the category with the highest probability is determined as the evaluation result.

Images