A Method and System for Constructing Deterioration Index of Hydraulic Turbine Generator Sets Based on Sparse Autoencoder and Extreme Gradient Boosting

Abstract

The present invention discloses a method and system for constructing a deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting, comprising the following steps: S1. According to the vibration data of the hydropower unit, a multi-dimensional feature set of vibration signals is constructed by using time-domain, frequency-domain, and energy feature extraction methods of vibration signals; S2. The sparse autoencoder is trained with multi-dimensional feature samples of the vibration signals of the hydropower unit to obtain a trained sparse autoencoder model as a multi-dimensional feature dimensionality reduction model of the unit; S3. The vibration monitoring data under the normal state of the hydropower unit is collected to form a health data set, and a health model of the vibration characteristics of the unit is established; S4. The deterioration index of the unit is calculated. The present invention takes the vibration signal characteristics of the hydropower unit as the starting point, refines the vibration characteristics of the unit, constructs a real-time deterioration index of the unit, reflects the deterioration trend of the unit, and has important significance for the state maintenance and fault warning of the hydropower unit, and reducing the fault loss of the hydropower unit.

Description

Technical Field

[0001] The present invention relates to the technical field of fault feature extraction of hydro-generator sets, and in particular to a method and system for constructing a degradation index of hydro-generator sets based on a sparse autoencoder and extreme gradient boosting. Background Technique

[0002] Hydraulic generator sets are the core key equipment of hydropower stations. The scale and capacity of hydro-generator sets are constantly increasing, and the structure is more complex. Their health status directly affects the safe and stable operation of the entire hydropower station. As the service life increases, different degrees of fault accidents will occur in hydro-generator sets. To avoid safety accidents, it is extremely important to deeply study the fault diagnosis and condition prediction of hydro-generator sets.

[0003] As an important monitoring index for evaluating the operation status of the unit, the vibration signal can reflect the internal operation and fault information of the unit. By analyzing and processing the vibration signal, equipment abnormalities can be detected in time, potential faults can be predicted, downtime can be reduced, the availability and operation efficiency of the unit can be improved, and the safe and stable operation of the unit can be ensured. By extracting features and constructing a degradation index from the vibration signal, the characteristic information can be highlighted, providing a good data basis for subsequent signal research, helping engineering and technical personnel to more accurately judge the status of the unit, and taking corresponding maintenance and repair measures in time to improve the reliability and operation efficiency of the unit. Therefore, accurately and reliably extracting and identifying fault features in the vibration signal and constructing a degradation index that can reflect the real-time status of the unit are crucial for the normal operation and maintenance of hydro-generator sets. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method and system for constructing a degradation index of hydro-generator sets based on a sparse autoencoder and extreme gradient boosting, using the sparse autoencoder model (SAE) and extreme gradient boosting (XGBoost) in the field of machine learning to construct a degradation index of hydro-generator sets to reflect the degradation trend of the unit, in order to find a more effective new method in the field of fault feature extraction of hydro-generator sets.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: A method for constructing a degradation index of hydro-generator sets based on a sparse autoencoder and extreme gradient boosting, comprising the following steps:

[0006] S1. According to the vibration data of the hydro-generator set, adopt the vibration signal time-domain, frequency-domain, and energy feature extraction methods to construct a multi-dimensional feature set of the vibration signal;

[0007] S2. Train a sparse autoencoder with the multi-dimensional feature samples of the hydro-generator set vibration signal to obtain a trained sparse autoencoder model as the multi-dimensional feature dimensionality reduction model of the unit;

[0008] S3. Collect the vibration monitoring data of the hydropower unit under normal conditions to form a health dataset. Use the normal data to train the Extreme Gradient Boosting algorithm, find the mapping relationship between each vibration dimensionality reduction feature and the working condition data of the hydropower unit, and establish a health model for the vibration characteristics of the unit.

[0009] S4. Calculate the deterioration index of the unit. Use the deviation between the real-time vibration dimensionality reduction feature of the unit obtained in S3 and the corresponding health feature, and fuse the deviations of each vibration dimensionality reduction feature as the deterioration index of the vibration signal of the hydropower unit.

[0010] In S1, the extraction of the time domain, frequency domain, and energy characteristics of the vibration signal specifically includes:

[0011] S11. Extract six characteristics of the time domain mean, standard deviation, peak-to-peak value, root mean square, kurtosis, and skewness of the target signal as follows:

[0012] ; (1)

[0013] ; (2)

[0014] ; (3)

[0015] ; (4)

[0016] ; (5)

[0017] ; (6)

[0018] Among them: represents the signal sequence, N is the number of sampling points, is the time domain mean, is the standard deviation, is the peak-to-peak value, is the root mean square, is the kurtosis, is the skewness;

[0019] S12. Extract seven indicators of the mean frequency, spectral skewness, spectral kurtosis, first-order spectral centroid, second-order spectral centroid, second-order spectral central moment, and standard deviation frequency of the target signal spectrum as the frequency characteristics, as follows:

[0020] ; (7)

[0021] ; (8)

[0022] ; (9)

[0023] ; (10)

[0024] ; (11)

[0025] ; (12)

[0026] ; (13)

[0027] Among them, represents the amplitude of each frequency corresponding to the signal spectrum, represents the frequency of each point on the spectrum, N represents the number of sampling points; represents the mean frequency, represents the spectral skewness, represents the spectral kurtosis, represents the first-order centroid of the spectrum, represents the second-order centroid of the spectrum, represents the second-order central moment of the spectrum, represents the standard deviation frequency;

[0028] S13. Extract the energy entropy of the target signal as its energy feature;

[0029] ; (14)

[0030] Among them, E represents the signal energy, represents the energy entropy of the target signal.

[0031] In S2, a sparse autoencoder is trained using the multi-dimensional feature samples of the unit vibration signal. By adding a sparsity limit to the hidden layer unit z in the autoencoder, the autoencoder can learn the key structures in the data. The steps are as follows:

[0032] S21. Assume that the number of training samples is N. Then the optimization objective of the sparse autoencoder is:

[0033] ; (15)

[0034] Among them, represents the vector composed of all training sample sequences at the input end, is the sparsity metric function, W is the hyperparameter of the sparse autoencoder, represents the optimization objective, represents the nth input sample sequence, represents the output sample sequence decoded by the sparse autoencoder, The weight coefficient representing the sparsity constraint condition represents the weight decay coefficient, whose function is to reduce the weight amplitude and prevent overfitting;

[0035] S22. Given N training samples, the average activity of the j th neuron in the hidden layer is:

[0036] ; (16)

[0037] In the formula, represents the activation probability of the jth neuron in the hidden layer, making close to the given value ; represents the output value of the jth neuron in the hidden layer when the nth group of sample inputs ;

[0038] Using relative entropy as a criterion to measure the probability difference, the calculation method is:

[0039] ; (17)

[0040] S23. When is equal to the given value , , so the sparsity metric function is defined as:

[0041] ; (18)

[0042] In the formula, p represents the number of neurons in the hidden layer, represents KL The divergence is used to limit the difference between and , represents the constant coefficient.

[0043] In S3, the operating condition data of the hydropower unit includes active power P , guide vane opening , blade opening , unit speed N , and water head H .

[0044] The steps to establish the unit vibration characteristic health model are as follows:

[0045] S31. Given a dataset A containing samples and features, the specific calculation method using the extreme gradient boosting algorithm is as follows:

[0046] ; (19)

[0047] ; (20)

[0048] ; (21)

[0049] In the formula, K represents the number of regression trees, represents the predicted value of the th sample, where the predicted value is obtained by accumulating through each tree, B is the solution space of the regression tree, A represents the given data set, represents the i-th sample, represents the i-th feature, represents the overall structure function of the CART tree, and the input sample can be mapped to the leaf node, represents the tree model, represents the output value of the leaf node, represents the structure function of the tree, represents the sample set, represents the tree set, represents the number of trees;

[0050] S32. Given represents the objective function of the extreme gradient boosting algorithm, and the calculation method of the objective function is as follows:

[0051] ; (22)

[0052] ; (23)

[0053] In the formula, represents the loss function, which is used to describe the deviation degree between the true value and the predicted value, represents the regularization term of the function, represents the objective function of the algorithm, represents the division difficulty of the leaf node, represents the regularization coefficient, represents the score of each leaf node;

[0054] S33. Expand the loss function to the second order using Taylor's formula and simplify the objective function. The calculation formula is as follows:

[0055] ; (24)

[0056] ; (25)

[0057] In the formula, represents the sum of the first-order gradients of the regression tree leaf nodes, represents the sum of the second-order gradients of the regression tree leaf nodes;

[0058] To minimize the objective function, the derivative of the objective function is set to zero for solution, and the obtained results are as follows:

[0059] ; (26)

[0060] ; (27)

[0061] In the formula, represents the weight standard value, represents the weight of the jth leaf node, represents the difficulty of leaf node division, represents the penalty coefficient, represents the number of trees, represents the partial derivative of the loss function, represents the partial derivative of the predicted value;

[0062] S34. Input the real-time operating condition data of the hydropower unit into the trained extreme gradient boosting algorithm, and the real-time health value of the vibration characteristics is obtained as follows:

[0063] ; (28)

[0064] ; (29)

[0065] In the formula 、 are the two-dimensional features after multi-dimensional feature dimensionality reduction, represents the active power, represents the guide vane opening, represents the blade opening, represents the unit speed, represents the water head.

[0066] In S4, the steps to calculate the unit deterioration index are as follows:

[0067] S41. Calculate the deviations between the two-dimensional vibration characteristics after dimensionality reduction and the unit health standards respectively as follows:

[0068] ; (30)

[0069] ; (31)

[0070] In the formula, 、 represent the symptom indicators corresponding to the vibration characteristic components, , represents the actual value of the feature, , represents the health value of the two-dimensional feature, , represents the health value of the two-dimensional feature;

[0071] S42. Fuse the corresponding indicators of the vibration feature components according to the weights to obtain the deterioration index of the target signal. The steps are as follows:

[0072] ; (32)

[0073] In the formula, represents the deterioration index of the vibration signal, , represent the weight coefficients, physically represents the degree of deviation of the vibration feature from the health value;

[0074] The vibration characteristics of the unit obtained through multi-dimensional feature dimensionality reduction contain local vibration feature information. Different features have different sensitivities to faults and may even show opposite change directions. Therefore, to avoid mutual cancellation between features, it is stipulated that and are consistent in sign, that is:

[0075] ; (33)

[0076] When calculating the fusion deterioration index of a single component, weights are introduced. The calculation method is as follows:

[0077] ; (34)

[0078] ; (35)

[0079] In the formula, , represent the weight coefficients, represents the health value of feature one, represents the health value of feature two, , are the actual values of the features.

[0080] S43. After completing the construction of the vibration signal deterioration index, further set the fault warning threshold according to the average value of the deterioration index in the normal state of the hydropower unit to define the normal state and the fault warning state of the hydropower unit. The specific calculation method is as follows:

[0081] ; (36)

[0082] In the formula, is the fault warning threshold, is the median of the overall early warning index for the normal state of the hydropower unit, and 𝛿 is the amplification factor;

[0083] Among them, when the deterioration index exceeds the health threshold, it indicates that the unit may have an early fault; otherwise, it indicates that the equipment is in a normal state.

[0084] A system adopting the method for constructing the deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting as described above, the system includes;

[0085] Unit vibration signal data acquisition module: The unit vibration signal data acquisition module is used to obtain the vibration historical data of the unit components under various working conditions, form a unit vibration signal sample set, and each sample contains the vibration waveform data of the key components;

[0086] Multi-dimensional feature extraction module: The multi-dimensional feature extraction module is used to extract the multi-dimensional features of the target signal, and adopts the vibration signal time domain, frequency domain, and energy feature extraction methods to construct a multi-dimensional feature set of the vibration signal;

[0087] Sparse autoencoder feature dimensionality reduction module: The sparse autoencoder feature dimensionality reduction module is used to input the training samples of each measuring point into the sparse autoencoder, train the sparse autoencoder, and use the trained sparse autoencoder model as the feature dimensionality reduction model, input the samples to be measured into the dimensionality reduction model, and obtain the dimensionality reduction features of each sample;

[0088] Extreme gradient boosting health model construction module: The extreme gradient boosting health model construction module is used to select the normal operating condition data of the unit under different working conditions and the vibration characteristics of the corresponding components as the input of the extreme gradient boosting algorithm for training, and input the real-time operating condition of the unit into the trained extreme gradient boosting algorithm to obtain the healthy vibration characteristics of the unit under this condition;

[0089] Deterioration index and its health threshold calculation module: The deterioration index and its health threshold calculation module uses the deviation between the real-time vibration characteristics of each unit and the corresponding health characteristics, fuses the deviation of each vibration characteristic as the deterioration index of the hydropower unit vibration signal, and selects the product of the median of the deterioration index in the healthy state and the magnification factor as the fault threshold;

[0090] Fault early warning module: The fault early warning module is used to calculate the deterioration index of the sample. When the deterioration index of the sample exceeds the early warning threshold, it performs the fault early warning action; otherwise, it indicates that the equipment is in a normal state, and then continues to calculate the deterioration index at the next moment.

[0091] The present invention has the following beneficial effects:

[0092] The present invention proposes a method and system for constructing vibration deterioration indexes of hydro-generator units based on a sparse autoencoder (SAE) and extreme gradient boosting (XGBoost). Multidimensional features are extracted from three aspects of the vibration signal, namely the time domain, frequency domain, and energy. The SAE is used to reduce the dimensionality of the multidimensional features to obtain key features. The mapping relationship between the operating condition data of the hydro-generator unit and the key features is found through XGBoost to construct a health standard model. The deviation between the real-time features and the standard features in the health model is used as the unit deterioration index, and the fault threshold of the unit deterioration index is determined to achieve early fault warning of the hydro-generator unit. Description of the Drawings

[0093] The present invention will be further described below with reference to the drawings and embodiments.

[0094] Figure 1 It is a flowchart of a method for constructing deterioration indexes of hydro-generator units based on a sparse autoencoder and extreme gradient boosting provided for the implementation of the present invention.

[0095] Figure 2 It is a network architecture diagram of the sparse autoencoder provided for the implementation of the present invention.

[0096] Figure 3 It is a flowchart of calculating deterioration indexes provided for the implementation of the present invention.

[0097] Figure 4 is an architecture diagram of a system for constructing deterioration indexes of hydro-generator units based on a sparse autoencoder and extreme gradient boosting provided for the implementation of the present invention. Detailed Embodiments

[0098] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention.

[0099] Embodiment 1:

[0100] A method for constructing deterioration indexes of hydro-generator units based on a sparse autoencoder and extreme gradient boosting proposed in the embodiment of the present invention is as follows with reference to Figures 1-3 , and the specific implementation steps are as follows:

[0101] Step 1: According to the vibration data of the hydro-generator unit, a method for extracting time domain, frequency domain, and energy characteristics of the vibration signal is used to construct a multidimensional feature set of the vibration signal. The construction process is as follows;

[0102] (1) The extracted time domain features include:

[0103] Mean value ;

[0104] Standard deviation ;

[0105] Peak - to - peak value ;

[0106] Root mean square ;

[0107] Kurtosis ;

[0108] Skewness ;

[0109] Among them, represents the signal sequence, N is the number of sampling points.

[0110] (2) Extracting frequency - domain features includes:

[0111] Mean frequency ;

[0112] Standard deviation of frequency ;

[0113] Frequency skewness ;

[0114] Spectral kurtosis ;

[0115] First - order spectral centroid ;

[0116] Second - order spectral centroid ;

[0117] Second - order spectral central moment ;

[0118] Among them represents the amplitude of each frequency corresponding to the signal spectrum, represents the frequency of each point on the spectrum, N is the number of sampling points.

[0119] (3) Extracting energy feature - energy entropy ;

[0120] Among them, E represents the signal energy, represents the energy entropy of the target signal.

[0121] Step 2: Use some samples of the vibration signal multi - dimensional features of the hydropower unit to train the sparse auto - encoder (SAE), and obtain the trained sparse auto - encoder (SAE) model as the multi - dimensional feature dimensionality reduction model of the unit.

[0122] Step 21: Assume that the number of training samples is N, then the optimization objective of the sparse auto - encoder is:

[0123] ;

[0124] wherein, represents the vector composed of all training sample sequences at the input end, is the sparsity metric function, W is the hyperparameter of the sparse autoencoder, represents the optimization objective, represents the nth input sample sequence, represents the output sample sequence obtained by decoding through the sparse autoencoder, represents the weight coefficient of the sparsity constraint condition, represents the weight decay coefficient, and its function is to reduce the weight amplitude and prevent overfitting.

[0125] Step 22: After giving N training samples, the average activity of the j th neuron in the hidden layer is:

[0126] ;

[0127] In the formula, represents the activation probability of the jth neuron in the hidden layer, making close to the given value ; represents the input of the nth group of samples when the output value of the jth neuron in the hidden layer; and through the Kullback-Leibler divergence, it is used as a standard to measure the probability difference, and the calculation method is: .

[0128] Step 23: When is equal to the given value , , so the sparsity metric function is defined as: ;

[0129] In the formula, p is the number of neurons in the hidden layer, represents KL the divergence is used to limit and the difference between them, represents the constant coefficient.

[0130] Step 3: Collect the vibration monitoring data of the hydropower unit under normal conditions to form a health dataset. Use the normal data to train the extreme gradient boosting, and use the trained extreme gradient boosting algorithm to find the vibration characteristics and active power P extracted in each step 2, and the guide vane opening , Blade opening , Unit speed N , Head H Establish a unit vibration characteristic health model for the mapping relationship among them.

[0131] Step 31: Given that a dataset A contains samples and features.

[0132] The specific calculation method of the Extreme Gradient Boosting algorithm (XGBoost) is as follows:

[0133] ,;

[0134] ;

[0135] ;

[0136] In the formula, K represents the number of regression trees, represents the predicted value of the th sample, where the predicted value is obtained by accumulating through each tree, B is the solution space of the regression tree, A represents the given dataset, represents the th sample, represents the overall structure function of the CART tree, and the input sample can be mapped to the leaf node, represents the tree model, represents the output value of the leaf node, represents the structure function of the tree, represents the sample set, represents the tree set, represents the number of trees.

[0137] Step 32: Given represents the objective function of the Extreme Gradient Boosting algorithm (XGBoost), and the calculation method of the objective function is obtained as follows:

[0138] ;

[0139] ;

[0140] Among them, represents the loss function, which is used to describe the deviation degree between the true value and the predicted value. represents the regularization term of the function, represents the objective function of the algorithm, represents the division difficulty of the leaf node, represents the regularization coefficient, represents the score of each leaf node.

[0141] Step 33: Expand the loss function using Taylor's formula to the second order and simplify the objective function. The calculation formula is:

[0142] ;

[0143] ;

[0144] where represents the sum of the first-order gradients of the regression tree leaf nodes, represents the sum of the second-order gradients of the regression tree leaf nodes. To minimize the objective function, set the derivative of the objective function to zero and solve, and the obtained result is:

[0145] and ;

[0146] In the formula, represents the weight standard value, represents the weight of the j-th leaf node, represents the difficulty of leaf node partitioning, represents the penalty coefficient, represents the number of trees, represents the partial derivative of the loss function, represents the partial derivative of the predicted value.

[0147] Step 34: Input the real-time operating condition data of the hydropower unit into the trained Extreme Gradient Boosting (XGBoost) algorithm, and obtain the real-time health value of the vibration characteristics as:

[0148] ;

[0149] ;

[0150] In the formula and are the two-dimensional features after multi-dimensional feature dimensionality reduction, represents the active power, represents the guide vane opening, represents the blade opening, represents the unit speed, represents the water head.

[0151] Step 4: Utilize the deviation between the real-time vibration characteristics of the unit obtained in Step 3 and the corresponding health characteristics, and fuse the deviations of each vibration characteristic as the deterioration index of the vibration signal of the hydropower unit.

[0152] Step 41: Calculate the deviations between the two-dimensional vibration characteristics after dimensionality reduction and the unit health standards respectively as follows:

[0153] 、 ;

[0154] Among them, 、 are the symptom indicators corresponding to the vibration characteristic components, 、 are the actual values of the characteristics, 、 represent the two-dimensional characteristic health values, 、 represent the two-dimensional characteristic health.

[0155] Step 42: Fuse the indicators corresponding to the vibration characteristic components according to the weights to obtain the deterioration index of the target signal:

[0156] ;

[0157] Among them, is the deterioration index of the vibration signal, 、 are the weight coefficients. Physically, it represents the degree of deviation of the vibration characteristics from the health values.

[0158] The unit vibration characteristics obtained through multi-dimensional characteristic dimensionality reduction contain local vibration characteristic information. Different characteristics have different sensitivities to faults, and even opposite change directions may occur. Therefore, to avoid mutual cancellation between characteristics, it is stipulated that and are consistent in sign, that is: 。

[0159] Step 43: To enhance the sensitivity of the deterioration index to abnormal data, weights are introduced when calculating the fusion deterioration index of a single component to more specifically consider the importance of different characteristics in reflecting the changes in the unit state. Such a weight selection can improve the ability to identify potential fault characteristics. The calculation method is as follows:

[0160] ;

[0161] ;

[0162] In the formula, 、 represent the weight coefficients, represents the health value of the first characteristic, represents the health value of the second characteristic, 、 are the actual values of the characteristics.

[0163] Step 5: After constructing the vibration signal deterioration index, further set the fault warning threshold according to the mean value of the deterioration index in the normal state of the hydropower unit to define the normal state and the fault warning state of the hydropower unit. The specific calculation method is as follows: .

[0164] Among them, is the median of the overall warning index in the normal state of the hydropower unit, and 𝛿 is the amplification factor. Take 𝛿 = 1.6. When the deterioration index exceeds the health threshold, it indicates that the unit may have an early fault, and on-site staff need to conduct a preliminary inspection; otherwise, it indicates that the equipment is in a normal state.

[0165] Example 2:

[0166] Refer to Figure 4 , a system for constructing the deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting proposed in an embodiment of the present invention. The system includes:

[0167] Vibration signal data acquisition module for the unit. Relying on the power station real-time monitoring system, obtain the vibration historical data of the unit components under various working conditions to form a vibration signal sample set of the unit (N is the number of samples, M is the number of monitoring variables included in the sample), and each sample should include the vibration waveform data of key components, such as the swing waveform data of each guide bearing, etc.

[0168] Multi-dimensional feature extraction module. Extract multi-dimensional features of the target signal. Adopt vibration signal time-domain, frequency-domain, and energy feature extraction methods to construct a multi-dimensional feature set of vibration signals.

[0169] Sparse autoencoder (SAE) feature reduction module. Input the training samples of each measurement point into the SAE, train the SAE according to the process of Step 2, and use the trained SAE model as the feature reduction model. Input the samples to be measured into the reduction model to calculate the reduced features of each sample.;

[0170] Extreme gradient boosting (XGBoost) health model construction module. Select the normal operating condition data of the unit under different working conditions and the corresponding vibration characteristics of the components as the input for training the extreme gradient boosting algorithm, and input the real-time operating condition of the unit into the trained extreme gradient boosting algorithm to obtain the healthy vibration characteristics of the unit under this condition.

[0171] Deterioration index and its health threshold calculation module. Use the deviation between the real-time vibration characteristics of each unit and the corresponding healthy characteristics, and fuse the deviation of each vibration characteristic as the vibration signal deterioration index of the hydropower unit. Select the product of the median of the deterioration index in the healthy state and the magnification factor as the fault threshold.

[0172] Fault warning module. For the sample to be measured, first calculate the sample deterioration index. When the sample deterioration index exceeds the warning threshold, it indicates that the equipment has early fault signs, and the fault warning action is executed; otherwise, it indicates that the equipment is in a normal state, and then continue to calculate the deterioration index at the next moment.

[0173] Those skilled in the art can easily understand that the above are only the specific implementation steps of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for constructing a deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting, characterized in that It includes the following steps: S1. According to the vibration data of the hydropower unit, adopt the time domain, frequency domain, and energy feature extraction methods of vibration signals to construct a multi-dimensional feature set of vibration signals; S2. Use the multi-dimensional feature samples of the hydropower unit vibration signal to train a sparse autoencoder, and obtain the trained sparse autoencoder model as the multi-dimensional feature dimensionality reduction model of the unit; S3. Collect the vibration monitoring data of the hydropower unit under normal conditions to form a health data set, use the normal data to train the extreme gradient boosting algorithm, find the mapping relationship between each vibration dimensionality reduction feature and the operating condition data of the hydropower unit, and establish a health model of the unit vibration characteristics; S4. Calculate the deterioration index of the unit, use the deviation between the real-time vibration dimensionality reduction feature of the unit obtained in S3 and the corresponding health feature, and fuse the deviation of each vibration dimensionality reduction feature as the deterioration index of the hydropower unit vibration signal; The steps to establish a health model of the unit vibration characteristics are as follows: S31. Given that a dataset A contains samples and features, the specific calculation method using the extreme gradient boosting algorithm is as follows: ；（19） ；（20） ； （21） In the formula, K represents the number of regression trees, represents the predicted value of the -th sample, where the predicted value is obtained by accumulating through each tree, B is the solution space of the regression tree, A represents the given data set, represents the i-th sample, represents the i-th feature, represents the overall structure function of the CART tree. The input sample can be mapped to the leaf node, represents the tree model, represents the output value of the leaf node, represents the structure function of the tree, represents the sample set, represents the tree set, represents the number of trees; S32. Given representing the objective function of the extreme gradient boosting algorithm, the calculation method of the objective function is as follows: ；（22） ； （23） Wherein, represents the loss function, which is used to describe the deviation degree between the true value and the predicted value, represents the regularization term of the function, represents the objective function of the algorithm, represents the partitioning difficulty of the leaf node, represents the regularization coefficient, represents the score of each leaf node; S33. Expand the loss function Use the Taylor formula to expand it to the second order and simplify the objective function. The calculation formula is as follows: ；（24） ；（25） In the formula, represents the sum of the first-order gradients of the regression tree leaf nodes, represents the sum of the second-order gradients of the regression tree leaf nodes; To minimize the objective function, set the derivative of the objective function to zero for solution, and the obtained results are as follows: ；（26） ；（27） In the formula, represents the standard value of the weight, represents the weight of the j leaf nodes, represents the difficulty of leaf node division, represents the penalty coefficient, represents the number of trees, represents the partial derivative of the loss function, represents the partial derivative of the predicted value; S34. Input the real-time operating condition data of the hydropower unit into the trained extreme gradient boosting algorithm, and obtain the real-time health value of the vibration characteristics as follows: ；（28） ；（29） where , are two-dimensional features after multi-dimensional feature dimensionality reduction, represents active power, represents the guide vane opening, represents the blade opening, represents the unit speed, represents the water head.

2. A method for constructing a deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting according to claim 1, characterized in that In S1, the extraction of the time domain, frequency domain, and energy characteristics of the vibration signal specifically includes: S11. Extract the target signal The six features of time-domain mean, standard deviation, peak-to-peak value, root mean square, kurtosis, and skewness are as follows: ；（1） ；（2） ；（3） ；（4） ；（5） ；（6） Wherein: represents the signal sequence, N is the number of sampling points, is the time-domain mean value, is the standard deviation, is the peak-to-peak value, is the root mean square, is the kurtosis, is the skewness; S12. Extract the spectrum of the target signal Take the mean frequency, spectral skewness, spectral kurtosis, first-order spectral centroid, second-order spectral centroid, second-order spectral central moment, and standard deviation frequency of these seven indicators as frequency features, as follows: ；（7） ；（8） ；（9） ；（10） ；（11） ；（12） ；（13） Among them, represents the amplitude of each frequency corresponding to the signal spectrum, represents the frequency of each point on the spectrum, N represents the number of sampling points; represents the mean frequency, represents the spectral skewness, represents the spectral kurtosis, represents the first-order centroid of the spectrum, represents the second-order centroid of the spectrum, represents the second-order central moment of the spectrum, represents the standard deviation frequency; S13. Extract the energy entropy of the target signal as its energy feature; ；（14） Among them, E represents the signal energy, represents the energy entropy of the target signal.

3. A method for constructing a deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting according to claim 1, characterized in that, In S2, use the multi-dimensional feature samples of the unit vibration signal to train a sparse autoencoder. By adding a sparsity limit to the hidden layer unit z in the autoencoder, the autoencoder can learn the key structure in the data. The steps are as follows: S21. Assume that the number of training samples is N, then the optimization objective of the sparse autoencoder is: ； （15） Among them, represents the vector composed of all training sample sequences at the input end, is the sparsity metric function, W is the hyperparameter of the sparse autoencoder, represents the optimization objective, represents the nth input sample sequence, represents the output sample sequence obtained by decoding through the sparse autoencoder, represents the weight coefficient of the sparsity constraint condition, represents the weight decay coefficient, whose function is to reduce the weight amplitude and prevent overfitting; S22. Given N after the training samples, the average activity of the j th neuron in the hidden layer is: ；（16） In the formula, represents the activation probability of the j-th neuron in the hidden layer, such that is close to the given value ; represents the output value of the j-th neuron in the hidden layer when the n-th group of sample inputs are input; Through relative entropy, which is used as a standard to measure the probability difference, the calculation method is: ；（17） S23. When is equal to a given value at this time , so the sparsity metric function is defined as: ；（18） In the formula, p represents the number of neurons in the hidden layer, represents KL The divergence is used to limit the difference between and represents a constant coefficient.

4. A method for constructing a deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting according to claim 1, characterized in that In S3, the operating condition data of the hydro-generating unit includes active power P , guide vane opening , blade opening , unit speed N , and water head H .

5. A method for constructing a deterioration index of a hydropower unit based on a sparse autoencoder and extreme gradient boosting according to claim 1, characterized in that In S4, the steps to calculate the deterioration index of the unit are as follows: S41. Calculate the deviation between the two-dimensional vibration characteristics after dimensionality reduction and the health standard of the unit respectively as follows: ；（30） ； （31） Wherein, and represent the symptom indexes corresponding to the vibration characteristic components, and represent the actual characteristic values, and represent the two-dimensional characteristic health values, and represent the two-dimensional characteristic health values; S42. Fuse the corresponding indicators of the vibration feature components by weight to obtain the deterioration index of the target signal. The steps are as follows: ； （32） In the formula, represents the vibration signal deterioration index, , represents the weight coefficient, physically represents the degree of deviation of the vibration characteristics from the healthy value; The vibration characteristics of the unit obtained through multi-dimensional feature dimensionality reduction contain local vibration feature information. Different features have different sensitivities to faults, and there may even be a phenomenon where the change directions are opposite. Therefore, to avoid mutual cancellation between features, it is stipulated that and be consistent in sign, that is: ；（33） When calculating the fusion deterioration index of a single component, introduce weights. The calculation method is as follows: ；（34） ；（35） In the formula, , represent weight coefficients, represents the health value of feature one, represents the health value of feature two, , are the actual feature values; S43. After completing the construction of the vibration signal deterioration index, further set the fault warning threshold according to the average value of the deterioration index in the normal state of the hydropower unit to define the normal state and the fault warning state of the hydropower unit. The specific calculation method is as follows: ； （36） wherein, is the fault warning threshold value, is the median of the overall warning index of the normal state of the hydropower unit, and 𝛿 is the amplification factor; Among them, when the deterioration index exceeds the health threshold, it indicates that the unit may have early faults; otherwise, it indicates that the equipment is in a normal state.

6. A system for constructing a deterioration index of a hydropower unit using the method according to any one of claims 1-5, characterized in that, The system includes; Unit vibration signal data acquisition module: The unit vibration signal data acquisition module is used to obtain the vibration historical data of the unit components under various operating conditions, form a unit vibration signal sample set, and each sample contains the vibration waveform data of the key components; Multi-dimensional feature extraction module: The multi-dimensional feature extraction module is used to extract the multi-dimensional features of the target signal, and adopt the time domain, frequency domain, and energy feature extraction methods of vibration signals to construct a multi-dimensional feature set of vibration signals; Sparse autoencoder feature dimension reduction module: The sparse autoencoder feature dimension reduction module is used to input the training samples of each measurement point into the sparse autoencoder, train the sparse autoencoder, and use the trained sparse autoencoder model as the feature dimension reduction model. Input the sample to be measured into the dimension reduction model to obtain the dimension-reduced features of each sample; Extreme gradient boosting health model construction module: The extreme gradient boosting health model construction module is used to select the operating condition data of the unit under different normal operating conditions and the vibration characteristics of the corresponding components as the input for training the extreme gradient boosting algorithm. Input the real-time operating condition of the unit into the trained extreme gradient boosting algorithm to obtain the healthy vibration characteristics of the unit under this condition; Deterioration index and its health threshold calculation module: The deterioration index and its health threshold calculation module uses the deviation between the real-time vibration characteristics of each unit and the corresponding health characteristics, fuses the deviation of each vibration characteristic as the deterioration index of the vibration signal of the hydropower unit, and selects the product of the median of the deterioration index and the magnification factor in the healthy state as the fault threshold; Fault warning module: The fault warning module is used to calculate the deterioration index of the sample. When the deterioration index of the sample exceeds the warning threshold, perform the fault warning action; Otherwise, it means that the equipment is in a normal state, and then continue to calculate the deterioration index at the next moment.

Images