A method and system for evaluating the degradation degree of a hydroelectric generating unit shafting

Abstract

The application relates to a water turbine unit shaft system degradation degree evaluation method, which comprises the following steps: analyzing the correlation between the health waveform sample mean of the water turbine unit in the health running stage and the corresponding initial working condition parameters, screening high-correlation working condition parameters, and constructing corresponding high-correlation working condition parameter characteristic vectors; performing K-means clustering analysis on the high-correlation working condition parameter characteristic vectors corresponding to the vibration waveform samples, combining a one-hot encoding method, and obtaining the classification working condition labels of the vibration waveform samples; inputting the classification working condition labels of the degradation waveform samples in the degradation running stage into a health model in sequence to generate corresponding health waveforms, comparing the actual degradation waveforms with the generated health waveforms to obtain degradation degrees, and constructing a sub-degradation sequence; and according to the sub-degradation sequence and the weight of the sub-degradation sequence, a comprehensive degradation sequence is calculated to evaluate the degradation degree of the water turbine unit shaft system; the application can effectively avoid the situation that only the vibration difference is considered and the unit state misjudgment caused by different working conditions is ignored.

Description

Technical Field

[0001] This application relates to the field of hydropower unit condition assessment technology, and in particular to a method and system for assessing the deterioration of the shaft system of a hydropower unit. Background Technology

[0002] As the core equipment of a hydropower station, the safe and stable operation of hydropower units directly affects the safe production of the station and the stability of the power system, and has enormous economic value. The shaft system of a hydropower unit, containing several guide bearings to withstand complex coupled unbalanced forces from electromagnetic, mechanical, and hydraulic forces, is a key component affecting the vibration characteristics and stability of the hydropower unit. Researching deterioration assessment methods for the vibration of the unit's shaft system is beneficial for timely understanding of the station's unit operating status, improving unit stability, reducing operation and maintenance costs, and thus improving economic efficiency.

[0003] Existing methods for assessing the deterioration of hydropower unit shaft systems based on vibration signals often select only one or more specific vibration indices to construct a health model, ignoring the impact of differences in unit vibration under different operating conditions on the deterioration process, leading to inaccurate assessment results. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and system for assessing the deterioration of the shaft system of hydropower units in response to the above-mentioned technical problems. This aims to solve the problem that the existing methods for assessing the deterioration of the shaft system of hydropower units based on vibration signals only model and analyze the vibration signals of the unit, ignoring the impact of the differences in unit vibration under different operating conditions on the deterioration process, which leads to inaccurate assessment results.

[0005] This invention provides a method for assessing the deterioration of a hydroelectric generator shaft system, the method comprising the following steps:

[0006] K-means clustering analysis was performed on the feature vectors of highly correlated operating condition parameters corresponding to the vibration waveform samples. Combined with the one-hot encoding method, the classification operating condition labels of the vibration waveform samples were obtained. Among them, the vibration waveform samples include healthy waveform samples in the healthy operation stage and deteriorated waveform samples in the deteriorated operation stage.

[0007] The classification condition labels of the degraded waveform samples are sequentially input into the trained health model to generate the corresponding health waveforms. The degree of degradation is obtained by comparing the actual degraded waveforms with the generated health waveforms, and the degree of degradation is sorted by time to construct a sub-degrade sequence.

[0008] Based on the sub-deterioration sequences and their weights, a comprehensive deterioration sequence is calculated, and the degree of deterioration of the hydropower unit's shaft system is assessed.

[0009] Furthermore, several highly correlated feature vectors of operating conditions are randomly selected as initial cluster centers;

[0010] K-means clustering analysis was used to calculate the distance from the feature vectors of the remaining highly correlated operating conditions to each cluster center, classify the feature vectors of the remaining highly correlated operating conditions, and calculate and update the cluster centers of each category.

[0011] Based on the correspondence between vibration waveform samples and feature vectors of high-performance operating condition parameters, the classification of vibration waveform samples is obtained;

[0012] Based on the number of classifications of the vibration waveform samples, the classifications of the vibration waveform samples are converted into one-hot codes of the corresponding dimensions to obtain the classification condition labels of the vibration waveform samples.

[0013] Furthermore, the health model includes a generator and a discriminator, and the training process of the health model specifically includes:

[0014] A generator is constructed using a one-dimensional deconvolutional network. The generator takes random noise and the classification labels corresponding to the actual healthy waveform samples during the healthy operation phase as input to generate the distribution of healthy waveform samples under different operating conditions during the healthy operation phase.

[0015] A discriminator is constructed using a one-dimensional convolutional network. The discriminator takes the generated health waveform samples of the healthy operation phase and the corresponding actual health waveform samples and their operating condition labels as input to determine the authenticity of the generated health waveforms of the healthy operation phase.

[0016] The objective function is to minimize the difference between the generated healthy waveform samples and the corresponding actual healthy waveform samples during the healthy operation phase, and the healthy model is trained using binary cross-entropy loss as the loss function.

[0017] Furthermore, the classification condition labels of the degraded waveform samples are sequentially input into the trained health model to generate corresponding healthy waveforms. The degree of degradation is obtained by comparing the actual degraded waveforms with the generated healthy waveforms, and the degradation degrees are sorted by time to construct a sub-degradation sequence, specifically including:

[0018] Input the classification condition labels of the degraded waveform samples in the degraded operation phase into the generator, and obtain the healthy waveforms generated under each condition in the degraded operation phase in sequence.

[0019] Calculate the degradation index of the actual degradation waveform and the generated healthy waveform during the degradation operation phase, and take the difference between the two as the degree of unit degradation.

[0020] The deterioration degree of the units in the deterioration operation stage is sorted by time to construct a sub-deterioration sequence.

[0021] Furthermore, the weights of the sub-degraded sequences are determined, specifically including:

[0022] Substitute the weights of the N initial sub-deterioration sequences into the calculation formula of the overall deterioration sequence to generate an initial population of size N as the parent population; where individuals in the population represent the overall deterioration sequence.

[0023] The coefficient of variation of the overall deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase were calculated using the parent population.

[0024] The coefficient of variation of the comprehensive deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase are sorted non-dominatedly, and the parent population is selected, crossed over and mutated using a genetic algorithm to generate the s-th generation sub-population; where the initial value of s is 1.

[0025] The parent and offspring populations were merged, and the coefficient of variation of the comprehensive deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration operation phase were non-dominated and sorted, and the individual crowding and non-matching order were calculated respectively.

[0026] Select several individuals based on the non-dominated order and individual crowding to update the parent population, increment the value of s by 1; and return the steps to calculate the coefficient of variation of the overall deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration phase.

[0027] Once the preset number of iterations is reached, the primary objective is to maximize the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase, while the secondary objective is to minimize the coefficient of variation of the comprehensive deterioration sequence under the healthy operation phase. The weights of the sub-deterioration sequences are then selected from the optimal solution of the multi-objective optimization problem.

[0028] Furthermore, before performing K-means clustering analysis on the feature vectors of highly correlated operating condition parameters corresponding to the vibration waveform samples, the method further includes: using the maximum information coefficient to analyze the correlation between the mean of healthy waveform samples of the hydropower unit during the healthy operation phase and the corresponding initial operating condition parameters, screening out highly correlated operating condition parameters with MIC indices greater than a threshold, and constructing the corresponding feature vectors of highly correlated operating condition parameters; wherein,

[0029] Initial operating parameters include: head, active power, reactive power, excitation voltage, excitation current, opening degree, and rotational speed;

[0030] Degradation metrics include: time-domain degradation metrics, frequency-domain degradation metrics, and data distribution metrics;

[0031] The time-domain degradation index is the mean and standard deviation of the actual degraded waveform or the generated healthy waveform during the degradation operation phase; the frequency-domain degradation index is the frequency standard deviation of the actual degraded waveform or the generated healthy waveform; the data distribution index is kurtosis and JS divergence.

[0032] Another embodiment of the present invention provides a system for assessing the deterioration of a hydroelectric generator shaft system, the system comprising:

[0033] The operating condition parameter screening module is used to analyze the correlation between the mean of the health waveform samples of the hydropower unit during the healthy operation phase and the corresponding initial operating condition parameters using the maximum information coefficient, and to screen out the highly correlated operating condition parameters whose MIC index is greater than the threshold, thus forming a feature vector of highly correlated operating condition parameters.

[0034] The classification condition label acquisition module performs K-means clustering analysis on the feature vectors of highly correlated operating condition parameters corresponding to vibration waveform samples, and combines the one-hot encoding method to obtain the classification condition labels of vibration waveform samples; among them, vibration waveform samples include healthy waveform samples in the healthy operation stage and deteriorated waveform samples in the deteriorated operation stage.

[0035] The degradation sequence construction module sequentially inputs the classification condition labels of the degradation waveform samples into the trained health model to generate corresponding health waveforms. It obtains the degradation degree by comparing the actual degradation waveform with the generated health waveform and constructs sub-degradation sequences by sorting the degradation degree by time.

[0036] The hydropower unit shaft system assessment module is used to calculate the comprehensive deterioration sequence based on the sub-deterioration sequence and its weight, and to assess the degree of deterioration of the hydropower unit shaft system.

[0037] The health model construction module uses health waveform samples from the health operation phase and corresponding classification condition labels to train a health model based on a conditional convolutional adversarial generative network.

[0038] Furthermore, the classification condition label acquisition module includes:

[0039] The initial cluster center selection unit is used to randomly select several highly correlated working condition parameter feature vectors as initial cluster centers;

[0040] The cluster center update unit is used to calculate the distance from the feature vectors of the remaining highly correlated operating conditions to each cluster center using K-means clustering analysis, classify the feature vectors of the remaining highly correlated operating conditions, and calculate and update the cluster centers of each category.

[0041] The classification unit for vibration waveform samples is used to obtain the classification of vibration waveform samples based on the correspondence between vibration waveform samples and feature vectors of high-performance operating condition parameters.

[0042] The classification condition label acquisition unit is used to convert the classification of vibration waveform samples into one-hot encoding of the corresponding dimension based on the number of classifications of the vibration waveform samples, and to obtain the classification condition label of the vibration waveform samples.

[0043] Furthermore, the health model construction module includes:

[0044] The generator, constructed using a one-dimensional deconvolutional network, is used to generate the distribution of health waveform samples under different operating conditions during the health operation phase, taking random noise and the classification labels corresponding to the actual health waveform samples during the health operation phase as input.

[0045] The discriminator, constructed using a one-dimensional convolutional network, is used to determine the authenticity of the generated healthy waveforms during the healthy operation phase, taking the generated healthy waveform samples and the corresponding actual healthy waveform samples and their operating condition labels as inputs; the generator and discriminator together form a health model.

[0046] The building unit is used to train the health model with the objective function of minimizing the difference between the generated healthy waveform samples and the corresponding actual healthy waveform samples, and with binary cross-entropy loss as the loss function.

[0047] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0048] (1) The present invention uses healthy waveform samples and corresponding classification operating condition labels during the healthy operation phase to train a health model based on conditional convolutional adversarial generative network. It can be seen that the present invention links highly correlated operating condition parameter information with vibration waveform samples, so that the health model can learn the distribution of vibration waveforms during the healthy operation phase and the differences in unit vibration waveforms under different operating conditions during the training process. This can effectively avoid simply considering vibration differences and ignoring the misjudgment of unit status caused by different operating conditions.

[0049] (2) In the process of constructing the health model, the generator takes random noise and the classification condition labels corresponding to the actual health waveform samples in the health operation stage as input, learns the distribution of health waveform samples under different conditions in the health operation stage, and generates the corresponding health waveforms; the discriminator takes the generated health waveform samples in the health operation stage and the corresponding actual health waveform samples and their condition labels as input, and judges the authenticity of the generated health waveforms in the health operation stage; the generator and the discriminator compete against each other to complete the training of the health model; the waveform output of the generator is the final output of the health model; wherein, the initial condition parameters include head, active power, reactive power, excitation voltage, excitation current, opening degree and speed; it can be seen that by using adversarial generative networks to establish a health model with the original vibration waveform as the output, the problem of loss of deterioration information when only one or more specific vibration indicators are selected to construct the health model is effectively avoided.

[0050] (3) This invention uses indicators from multiple perspectives, including time domain, frequency domain, and data distribution, to construct sub-deterioration sequences; it obtains the weight of each sub-deterioration sequence through multi-objective optimization; and the comprehensive deterioration sequence is composed of each sub-deterioration sequence and its corresponding weight. The algorithm's optimization objective is to consider the differences in sensitivity of different analytical indicators to the deterioration process and the reliability of the indicators. The multi-objective optimization method determines the weight of each indicator, thereby improving the comprehensive indicator's ability to represent the deterioration process and its reliability, thus solving the problem that manually selected deterioration indicators can only reflect a single characteristic. Attached Figure Description

[0051] Figure 1 This is a flowchart of the hydroelectric generator shaft system deterioration assessment provided in an embodiment of the present invention;

[0052] Figure 2 This is a structural diagram of the health model provided in an embodiment of the present invention;

[0053] Figure 3 This is a graph showing the clustering analysis results of the K-means of the unit operating parameters provided in this embodiment of the invention;

[0054] Figure 4 These are health waveforms generated by the health model provided in this embodiment of the invention and real sample images;

[0055] Figure 5 This is a graph showing the results of multiple indicator degradation sequences generated according to an embodiment of the present invention;

[0056] Figure 6 This is the result of multi-objective optimization for the weights of each degraded sequence provided in the embodiments of the present invention;

[0057] Figure 7 This is a diagram showing the overall degradation sequence results provided in an embodiment of the present invention;

[0058] Figure 8 This is a diagram of the hydroelectric generator shaft system deterioration assessment system provided in an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] This invention provides a method for assessing the deterioration of a hydroelectric generator shaft system. The overall scheme is as follows: First, the operating parameters are screened and analyzed using the maximum information coefficient and K-means clustering to obtain operating condition classification labels; a health model based on a conditional convolutional adversarial generative network is established to obtain healthy vibration waveform samples under each operating condition; five deterioration sequences are constructed by comparing the time domain, frequency domain, and data distribution differences between the healthy waveforms and the actual deteriorated waveforms; the optimal weights for each deterioration sequence are determined using a multi-objective optimization method to generate a comprehensive deterioration sequence, with the optimization objective being to reduce the volatility of the comprehensive deterioration sequence while providing sensitivity to the deterioration process.

[0061] like Figure 1 As shown, the method for assessing the deterioration degree of the shaft system of a hydroelectric generator provided in this embodiment of the invention includes the following steps:

[0062] Step (1): Obtain vibration waveform samples V during the healthy operation phase of the hydropower unit (hereinafter referred to as: healthy waveform). k =D k (i) and the initial operating condition parameter vector W of the hydropower unit k =[H k ,P k Q k U k ,I k O k N k The correlation (MIC) between various operating condition parameters and the mean of healthy waveform samples was analyzed using the maximum information coefficient (MIC). Several operating condition parameters with high cross-correlation were selected to form corresponding feature vectors of highly correlated operating condition parameters. K-means clustering analysis was performed on the feature vectors of highly correlated operating condition parameters to obtain operating condition classification labels, which were then one-hot encoded to obtain the label l. c .

[0063] Among them, the healthy waveform sample can be the waveform sample of the upper guide swing X-direction measuring point; H k For water head; P k Q represents active power. k Reactive power; U k I is the excitation voltage; k For excitation current; O k For opening; N k V is the rotational speed; k For the k-th healthy waveform sample; D k (i) represents the i-th point in the k-th health waveform sample; i∈[1,n], k∈[1,m], n represents the number of sampling points in each health waveform sample; m represents the number of health waveform samples.

[0064] Step (1) specifically includes the following:

[0065] (1-1) Obtaining health waveform samples V during the healthy operation phase k =D k (i), i∈[1,n], k∈[1,m], n represents the number of sampling points for each healthy waveform sample; m represents the number of healthy waveform samples; the operating parameters corresponding to each healthy waveform sample are obtained, including head, active power, reactive power, excitation voltage, excitation current, opening degree and speed, denoted as a vector: W k =[H k ,P k Q k U k ,I k O k N k The correlation between each operating condition parameter and the healthy waveform sample is calculated sequentially using maximum information coefficient analysis. The correlation is represented by the MIC index, and its calculation method is as follows:

[0066]

[0067] Where X represents a set of parameter variables for a certain operating condition. P(x,y) represents the mean sample of healthy waveforms; P(x,y) represents the joint density function of x and y; x is a variable in the set X; y is an element in the sample Y; where healthy waveform samples and degraded waveform samples are used to construct vibration waveform samples.

[0068] (1-2) Based on the MIC index, low-correlation operating parameters in the vibration waveform samples are removed. K-means clustering analysis is then performed on the feature vector W′ of the highly correlated operating parameters. The process is as follows:

[0069] (1-2-1) Randomly select c highly correlated feature vectors of operating condition parameters as initial cluster centers, denoted as a = a1, a2, ..., a c ;

[0070] (1-2-2) For any element W in the feature vector of the remaining highly correlated operating condition parameters k Calculate the distance from each cluster center and assign the remaining highly correlated operating condition parameter feature vectors to the corresponding categories;

[0071] (1-2-3) After completing the category division in (1-2-2), recalculate the cluster centers for each category. cj is the set of remaining working condition parameter vectors in the j-th category; j = 1, 2, ..., c;

[0072] (1-2-4) Repeat (1-2-2) and (1-2-3) until the specified number of iterations is reached;

[0073] (1-2-5) After the iteration is completed, based on the correspondence between the vibration waveform samples and the feature vectors of the high-performance operating condition parameters, the classification operating condition labels of healthy waveform samples and degraded waveform samples are obtained. c Based on the number of categories c, it is converted into a one-hot encoding of the corresponding c-dimensional dimension.

[0074] Step (2): As Figure 2 As shown, a health model based on a conditional convolutional adversarial generative network is constructed. The generator learns the distribution of health waveform samples under different operating conditions; the discriminator judges the realism of the generated samples. The goal of model training is to increase the discriminative loss of the generated samples, while the discriminator training aims to reduce the discriminative loss. The two interact, thus reducing the gap between the generated samples and the actual samples. This invention uses health waveform samples and their operating condition labels. c The model training is completed using binary cross-entropy loss as the loss function; the specific steps include the following:

[0075] (2-1) Construct a generator using a one-dimensional deconvolutional network; it contains two input layers, which take random noise z and the working condition label l corresponding to the healthy waveform sample as inputs, respectively. c The output of the input layer is as follows:

[0076]

[0077] Where deconv represents deconvolution; bn represents regularization; relu represents the ReLU activation function; ii is the ii-th convolutional kernel; nn is the number of convolutional kernels in this layer; w and b represent the convolutional kernel and bias, respectively; the outputs of the hidden and output layers in the generator are as follows:

[0078]

[0079] Where l is the number of hidden or output layers.

[0080] (2-2) Constructing the discriminator for the health model. The discriminator uses a one-dimensional convolutional network, which also contains two input layers. The inputs are the health waveform sample v and the corresponding working condition label l, respectively. c The output of the input layer is as follows:

[0081]

[0082] Where conv represents convolution; the outputs of the hidden and output layers in the discriminator are as follows:

[0083]

[0084] Where, leakyrelu and sigmoid represent the leakyrelu ReLU activation function and the sigmoid activation function, respectively; maxpool and avgpool represent max pooling and average pooling, l = L represents the output layer; ip is the ip-th convolutional kernel; and np is the number of convolutional kernels in this layer.

[0085] (2-3) The health model is trained, and the objective function of the entire health model is expressed as follows:

[0086]

[0087] Where V(D,G) is the objective function; D represents the discriminator, G represents the generator; v is a healthy waveform sample; pdata is the true distribution of the healthy waveform samples; pz(z) is the distribution of the random number z; D(v|l c G(z|l) represents the conditional probability of the discriminator; c The conditional probability of the generator is denoted as ; in this invention, binary cross-entropy loss is used as the loss function, expressed as follows:

[0088] L BCE =y B ·logo L +(1-y B log(1-O) L )

[0089] Among them, O L This represents the output of the discriminator output layer in (2-2); y B Indicates the type of sample, taking the value 0 or 1; 0 indicates that the sample is a generated health waveform sample, and 1 indicates that the sample is an actual health waveform sample.

[0090] Step (3): Input the classification condition parameter labels corresponding to the deteriorated waveform samples in the deterioration operation stage into the health model in sequence to generate the corresponding health waveform. By comparing the actual deteriorated waveform with the generated health waveform, the sub-deterioration sequence is obtained. The difference values ​​of mean, standard deviation, kurtosis, frequency standard deviation and JS divergence index are calculated respectively to reflect the deterioration process from the three aspects of time domain, frequency domain and data distribution.

[0091] The calculation method for the five deterioration sequences includes the following steps:

[0092] (3-1) Input the classification condition labels of the deteriorated waveform samples in the deterioration operation stage into the generator completed in step (2) in sequence to obtain the healthy waveforms V generated under each condition in the deterioration operation stage. lk =D lk (i), i∈[1,n], lk∈[1,t], t represents the number of degraded samples, by comparing with the actual degraded waveform V lk ′=Dlk (i) Establish a degradation index based on the difference between the generated healthy waveform and the original healthy waveform.

[0093] (3-2) Calculate the degradation index of the actual degradation waveform and the generated healthy waveform during the degradation operation stage respectively. The degradation index includes time domain, frequency domain and data distribution index. Calculate the difference between the two as the degree of unit degradation.

[0094] The time-domain degradation index is the average of the actual degradation waveform or the generated healthy waveform during the degradation operation phase. The formula for the sum and standard deviation σ is as follows:

[0095]

[0096] The frequency domain degradation index is the frequency standard deviation (RVF) of the actual degraded waveform or the generated healthy waveform, and the calculation formula is as follows:

[0097]

[0098] Where f is the frequency of the actual degraded waveform or the generated healthy waveform; P(f) is the power spectrum of the healthy waveform signal; FC represents the signal centroid frequency;

[0099] The data distribution metrics are kurtosis (K) and JS divergence (DJS), calculated using the following formulas:

[0100]

[0101]

[0102]

[0103] Where p(x) and q(x) represent the actual degraded waveform distribution and the generated healthy waveform distribution, respectively.

[0104] (3-3) Based on the indicators in (3-2), the degree of degradation can be calculated, and the degree of degradation in the degradation operation stage can be sorted by time to construct five sub-degradation sequences as follows:

[0105]

[0106] R3(k)=RVF′ k -RVF k R4(k)=K k '-K k

[0107] R5(k)=D JS (p||q)

[0108] Where k∈[1,t].

[0109] Step (4): Use NSGA-II multi-objective optimization to determine the weights of the five sub-deterioration sequences to obtain the overall deterioration sequence of the hydropower unit; the first optimization objective is to reduce the volatility of the overall deterioration sequence of the hydropower unit in the healthy phase, that is, to reduce the coefficient of variation of the overall sequence in the healthy interval; the second objective is to improve the sensitivity of the overall deterioration sequence of the hydropower unit to the degree of deterioration, that is, to increase the difference between the average values ​​of the healthy interval and the deterioration interval; specifically, the following steps are included:

[0110] (4-1) The comprehensive degradation sequence of the vibration waveform is as follows: k∈[1,m+t], where m and t are the number of healthy waveforms and degraded waveforms, respectively;

[0111] The first optimization objective is to minimize the coefficient of variation (cv) of the overall degradation sequence during the healthy operation phase, as shown in the following formula:

[0112]

[0113] The second optimization objective is to maximize the difference Δu between the means of the sequences in the healthy and deteriorated states, as shown in the following formula:

[0114]

[0115] Therefore, the multi-objective optimization problem can be summarized as follows:

[0116]

[0117] (4-2) The NSGA-II algorithm is used to perform multi-objective optimization. The algorithm flow is as follows:

[0118] (4-2-1) Substitute the weights of the N sub-degraded sequences into the calculation formula of the comprehensive degraded sequence, randomly generate an initial population of size N, and substitute them into the formula for calculating the coefficient of variation cv and the mean difference to obtain the initial coefficient of variation cv and the mean difference Δu; perform non-dominated sorting on the coefficient of variation cv and the mean difference Δu, and use a genetic algorithm to select, crossover and mutate the weights of the degraded sequences to generate the first generation of sub-populations;

[0119] (4-2-2) Merge the parent and offspring populations and perform fast non-dominated sorting, then calculate the crowding degree and non-matching order of individuals respectively;

[0120] (4-2-3) After the fast non-dominated sorting is completed, each individual i in the population contains a non-dominated order irank and a crowding degree id. Based on the non-dominated order and the individual crowding degree, select several individuals to form a new parent population. Then, through selection, crossover and mutation, obtain new offspring. Go to (4-2-2) until the preset number of iterations is reached.

[0121] (4-3) Of the two objectives, maxΔμ is the primary objective that should be satisfied first, and min cv As a secondary objective, and satisfying Δμ>0.5,c v If the value is less than 0.1, appropriate sequence weights can be selected from the multi-objective optimal solution.

[0122] Example

[0123] To more clearly illustrate the present invention and highlight its advantages, monitoring data from Unit 2 of a hydropower station are used to further explain the present invention.

[0124] Step 1: According to the power plant unit maintenance report, Unit 2 operated normally from January to July 14, 2018. On July 14, the unit began to experience abnormal vibration, which continued until the regular maintenance work was carried out. 327 vibration waveform samples of the X-direction swing of the guide bearing on the unit's shaft system from January to July were selected as healthy waveform samples to construct a health model. Deterioration analysis was performed on 490 vibration waveform samples after July. The number of sampling points for each vibration waveform sample was 256.

[0125] The operating parameters corresponding to 817 vibration waveform samples were analyzed, including head, guide vane opening, active power, reactive power, excitation current, excitation voltage, and rotational speed. The mean of each vibration waveform sample was calculated, and the correlation between the waveform mean and each operating parameter was calculated using the maximum information coefficient. The values ​​were 0.3294, 0.8484, 0.6708, 0.0120, 0.4841, 0.0993, and 0.0794, respectively. The correlation coefficients showed that the correlations of reactive power, excitation voltage, and rotational speed were all below 0.1, so these three operating parameters were discarded.

[0126] Feature vectors were constructed using four highly correlated operating parameters: active power, guide vane opening, excitation current, and head. K-means clustering analysis was then performed on the unit's operating conditions. An appropriate number of clusters was selected; in this invention, 5 clusters were chosen. The visualization of the clustering results is shown below. Figure 3 As shown; according to the working condition classification, a working condition label is added to each vibration waveform sample. One-hot encoding is used. According to the number of categories, the label vector dimension is 5, that is, there are five category labels: [1,0,0,0,0], [0,1,0,0,0], ... [0,0,0,0,1].

[0127] Step 2: Use 327 healthy vibration waveforms and working condition labels to train and test the conditional convolutional adversarial generative network to generate a healthy model, thereby obtaining the distribution of healthy vibration waveforms under various working conditions; the input of the generator is random noise with a vector dimension of 50 and the one-hot encoding of the working condition label with a dimension of 5 in Step 1, and the input of the discriminator is a vibration waveform of length 256 (including the actual waveform and the generated waveform) and the corresponding working condition label;

[0128] The training and test sets were split in a 9:1 ratio. The network was trained 100 times, with a generator learning rate of 0.0001 and a discriminator learning rate of 0.00003. The optimizer used was Adam, and the batch size for each training run was 16. After training, the healthy vibration waveforms generated by the network on the test set matched the actual vibration waveforms as follows: Figure 4 As shown, the two are highly similar;

[0129] Step 3: Following the steps in (3), input the degradation process condition labels into the health model sequentially to obtain the corresponding healthy vibration waveforms. Calculate the difference between the generated healthy vibration waveforms and the actual vibration waveforms using the waveform mean, waveform standard deviation, waveform frequency standard deviation, waveform kurtosis, and waveform JS divergence. Obtain five degradation sequences. Some degradation sequence results are shown below. Figure 5 As shown; based on the number of healthy waveform samples and degraded waveform samples, there are 327 healthy operating intervals in the degraded sequence, and a total of 490 points in the degraded process. Figure 5 It can be seen that the sensitivity of the degradation sequence to the degradation process varies under different indicators. In order to improve the degree of response of the degradation sequence to the degradation process, a multi-objective optimization method is used to determine the weight of each degradation sequence and obtain the comprehensive degradation sequence.

[0130] According to the NSGA-II multi-objective optimization method in step (5), the primary objective is to maximize the mean difference maxΔμ, that is, to maximize the difference between the mean of the healthy interval and the mean of the deterioration interval of the overall deterioration sequence, thereby improving the sensitivity of the deterioration sequence to the unit deterioration process; the secondary objective is to minimize the coefficient of variation min cv of the healthy interval of the overall sequence, to reduce the volatility of the healthy operating state of the overall deterioration sequence, thereby improving the reliability of the fluctuation of the deterioration process; either of the above two objectives being too large or too small will cause the deterioration sequence to fail, so Δμ>0.5 is set. v <0.1; Multi-objective optimization results are as follows Figure 6 As shown, in the optimal solution set, the solutions within the dashed box represent the set of solutions that satisfy the given requirements. Based on the principle of optimally satisfying the primary objective, we can deduce that the point that maximizes the mean difference should be selected. Figure 6 The optimal value point is marked in the middle; at this point, the maximum mean difference and coefficient of variation of the two objective functions are 0.7369 and 0.0967, respectively; the weights of the five degraded sequences (JS divergence, mean, standard deviation, kurtosis, and frequency standard deviation) are 0.0060, 0.0683, 0.0089, 0.1671, and 0.9165, respectively; multiplying the five sequences by their weights yields the comprehensive degraded sequence as follows. Figure 7 As shown, the degradation sequence is more stable within the healthy operating range, and the indicators respond more significantly to the degradation process, indicating that the proposed comprehensive degradation index has improved sensitivity to the degradation process.

[0131] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0132] (1) The present invention uses healthy waveform samples and corresponding classification operating condition labels during the healthy operation phase to train a health model based on conditional convolutional adversarial generative network. It can be seen that the present invention links highly correlated operating condition parameter information with vibration waveform samples, so that the health model can learn the distribution of vibration waveforms during the healthy operation phase and the differences in unit vibration waveforms under different operating conditions during the training process. This can effectively avoid simply considering vibration differences and ignoring the misjudgment of unit status caused by different operating conditions.

[0133] (2) In the process of constructing the health model, the generator takes random noise and the classification condition labels corresponding to the actual health waveform samples in the health operation stage as input, learns the distribution of health waveform samples under different conditions in the health operation stage, and generates the corresponding health waveforms; the discriminator takes the generated health waveform samples in the health operation stage and the corresponding actual health waveform samples and their condition labels as input, and judges the authenticity of the generated health waveforms in the health operation stage; the generator and the discriminator compete against each other to complete the training of the health model; the waveform output of the generator is the final output of the health model; wherein, the initial condition parameters include head, active power, reactive power, excitation voltage, excitation current, opening degree and speed; it can be seen that by using adversarial generative networks to establish a health model with the original vibration waveform as the output, the problem of loss of deterioration information when only one or more specific vibration indicators are selected to construct the health model is effectively avoided.

[0134] (3) This invention uses indicators from multiple perspectives, including time domain, frequency domain, and data distribution, to construct sub-deterioration sequences; it obtains the weight of each sub-deterioration sequence through multi-objective optimization; and the comprehensive deterioration sequence is composed of each sub-deterioration sequence and its corresponding weight. The algorithm's optimization objective is to consider the differences in sensitivity of different analytical indicators to the deterioration process and the reliability of the indicators. The multi-objective optimization method determines the weight of each indicator, thereby improving the comprehensive indicator's ability to represent the deterioration process and its reliability, thus solving the problem that manually selected deterioration indicators can only reflect a single characteristic.

[0135] Please see Figure 8 The present invention also provides a device for assessing the deterioration of a hydroelectric generator shaft system, the system comprising:

[0136] The operating condition parameter screening module 21 is used to analyze the correlation between the mean of the health waveform samples of the hydropower unit during the healthy operation phase and the corresponding initial operating condition parameters using the maximum information coefficient, and to screen out the highly correlated operating condition parameters whose MIC index is greater than the threshold, thus forming a feature vector of highly correlated operating condition parameters.

[0137] The classification condition label acquisition module 22 performs K-means clustering analysis on the feature vectors of highly correlated condition parameters corresponding to the vibration waveform samples, and obtains the classification condition labels of the vibration waveform samples by combining the one-hot encoding method; wherein, the vibration waveform samples include healthy waveform samples in the healthy operation stage and deteriorated waveform samples in the deteriorated operation stage.

[0138] Preferably, the classification condition label acquisition module includes:

[0139] The initial cluster center selection unit is used to randomly select several highly correlated working condition parameter feature vectors as initial cluster centers;

[0140] The cluster center update unit is used to calculate the distance from the feature vectors of the remaining highly correlated operating conditions to each cluster center using K-means clustering analysis, classify the feature vectors of the remaining highly correlated operating conditions, and calculate and update the cluster centers of each category.

[0141] The classification unit for vibration waveform samples is used to obtain the classification of vibration waveform samples based on the correspondence between vibration waveform samples and feature vectors of high-performance operating condition parameters.

[0142] The classification condition label acquisition unit is used to convert the classification of vibration waveform samples into one-hot encoding of the corresponding dimension based on the number of classifications of the vibration waveform samples, and to obtain the classification condition label of the vibration waveform samples.

[0143] The health model construction module 23 uses health waveform samples from the health operation phase and corresponding classification condition labels to train a health model based on a conditional convolutional adversarial generative network.

[0144] Preferably, the health model construction module 23 includes:

[0145] The generator, constructed using a one-dimensional deconvolutional network, is used to generate the distribution of health waveform samples under different operating conditions during the health operation phase, taking random noise and the classification labels corresponding to the actual health waveform samples during the health operation phase as input.

[0146] The discriminator, constructed using a one-dimensional convolutional network, is used to determine the authenticity of the generated healthy waveforms during the healthy operation phase, taking the generated healthy waveform samples and the corresponding actual healthy waveform samples and their operating condition labels as inputs; the generator and discriminator together form a health model.

[0147] The building unit is used to train the health model with the objective function of minimizing the difference between the generated healthy waveform samples and the corresponding actual healthy waveform samples, and with binary cross-entropy loss as the loss function.

[0148] The degradation sequence construction module 24 inputs the classification condition labels of the degradation waveform samples into the trained health model in sequence to generate the corresponding health waveforms. It obtains the degradation degree by comparing the actual degradation waveforms with the generated health waveforms and constructs sub-degradation sequences by sorting the degradation degree by time.

[0149] Preferably, the degraded sequence construction module 24 includes:

[0150] The unit deterioration degree calculation unit is used to calculate the deterioration index of the actual deteriorated waveform and the generated healthy waveform during the deterioration operation stage, and the difference between the two is the unit deterioration degree.

[0151] The method for obtaining the healthy waveform generated during the deterioration operation phase is as follows: input the classification condition label of the deterioration waveform sample during the deterioration operation phase into the generator, and obtain the healthy waveform generated under each condition during the deterioration operation phase in sequence.

[0152] The sub-deterioration sequence construction unit is used to construct a sub-deterioration sequence by sorting the degree of unit deterioration in the deterioration operation phase according to time.

[0153] The hydropower unit shaft system evaluation module 25 is used to calculate the comprehensive deterioration sequence based on the sub-deterioration sequence and its weight, and to evaluate the degree of deterioration of the hydropower unit shaft system.

[0154] Preferably, the determination of the weights of the sub-degraded sequences specifically includes:

[0155] Substitute the weights of the N initial sub-deterioration sequences into the calculation formula of the overall deterioration sequence to generate an initial population of size N as the parent population; where individuals in the population represent the overall deterioration sequence.

[0156] The coefficient of variation of the overall deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase were calculated using the parent population.

[0157] The coefficient of variation of the comprehensive deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase are sorted non-dominatedly, and the parent population is selected, crossed over and mutated using a genetic algorithm to generate the s-th generation sub-population; where the initial value of s is 1.

[0158] The parent and offspring populations were merged, and the coefficient of variation of the comprehensive deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration operation phase were non-dominated and sorted, and the individual crowding and non-matching order were calculated respectively.

[0159] Select several individuals based on the non-dominated order and individual crowding to update the parent population, increment the value of s by 1; and return the steps to calculate the coefficient of variation of the overall deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration phase.

[0160] Once the preset number of iterations is reached, the primary objective is to maximize the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase, while the secondary objective is to minimize the coefficient of variation of the comprehensive deterioration sequence under the healthy operation phase. The weights of the sub-deterioration sequences are then selected from the optimal solution of the multi-objective optimization problem.

[0161] Furthermore, the initial operating parameters include: head, active power, reactive power, excitation voltage, excitation current, opening degree, and rotational speed;

[0162] Degradation metrics include: time-domain degradation metrics, frequency-domain degradation metrics, and data distribution metrics;

[0163] The time-domain degradation index is the mean and standard deviation of the actual degraded waveform or the generated healthy waveform during the degradation operation phase; the frequency-domain degradation index is the frequency standard deviation of the actual degraded waveform or the generated healthy waveform; the data distribution index is kurtosis and JS divergence.

[0164] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0165] (1) The present invention uses healthy waveform samples and corresponding classification operating condition labels during the healthy operation phase to train a health model based on conditional convolutional adversarial generative network. It can be seen that the present invention links highly correlated operating condition parameter information with vibration waveform samples, so that the health model can learn the distribution of vibration waveforms during the healthy operation phase and the differences in unit vibration waveforms under different operating conditions during the training process. This can effectively avoid simply considering vibration differences and ignoring the misjudgment of unit status caused by different operating conditions.

[0166] (2) In the process of constructing the health model, the generator takes random noise and the classification condition labels corresponding to the actual health waveform samples in the health operation stage as input, learns the distribution of health waveform samples under different conditions in the health operation stage, and generates the corresponding health waveforms; the discriminator takes the generated health waveform samples in the health operation stage and the corresponding actual health waveform samples and their condition labels as input, and judges the authenticity of the generated health waveforms in the health operation stage; the generator and the discriminator compete against each other to complete the training of the health model; the waveform output of the generator is the final output of the health model; wherein, the initial condition parameters include head, active power, reactive power, excitation voltage, excitation current, opening degree and speed; it can be seen that by using adversarial generative networks to establish a health model with the original vibration waveform as the output, the problem of loss of deterioration information when only one or more specific vibration indicators are selected to construct the health model is effectively avoided.

[0167] (3) This invention uses indicators from multiple perspectives, including time domain, frequency domain, and data distribution, to construct sub-deterioration sequences; it obtains the weight of each sub-deterioration sequence through multi-objective optimization; and the comprehensive deterioration sequence is composed of each sub-deterioration sequence and its corresponding weight. The algorithm's optimization objective is to consider the differences in sensitivity of different analytical indicators to the deterioration process and the reliability of the indicators. The multi-objective optimization method determines the weight of each indicator, thereby improving the comprehensive indicator's ability to represent the deterioration process and its reliability, thus solving the problem that manually selected deterioration indicators can only reflect a single characteristic.

[0168] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assessing the deterioration degree of a hydroelectric generator shaft system, characterized in that, The method includes the following steps: The correlation between the mean of the health waveform samples and the corresponding initial operating parameters of the hydropower unit during the healthy operation phase was analyzed using the maximum information coefficient (MIC). Highly correlated operating parameters with MIC values ​​greater than a threshold were selected, forming the corresponding feature vectors of these highly correlated operating parameters. Initial operating parameters include: head, active power, reactive power, excitation voltage, excitation current, opening degree, and rotational speed; K-means clustering analysis was performed on the feature vectors of highly correlated operating condition parameters corresponding to the vibration waveform samples. Combined with the one-hot encoding method, the classification operating condition labels of the vibration waveform samples were obtained. Among them, the vibration waveform samples include healthy waveform samples in the healthy operation stage and deteriorated waveform samples in the deteriorated operation stage. The classification condition labels of the degraded waveform samples are sequentially input into the trained health model to generate the corresponding health waveforms. The degree of degradation is obtained by comparing the actual degraded waveforms with the generated health waveforms, and the degree of degradation is sorted by time to construct a sub-degrade sequence. Based on the sub-deterioration sequences and their weights, a comprehensive deterioration sequence is calculated, and the degree of deterioration of the hydropower unit's shaft system is assessed. The determination of the weights of the sub-degraded sequences specifically includes: Substitute the weights of the N initial sub-deterioration sequences into the calculation formula of the overall deterioration sequence to generate an initial population of size N as the parent population; where individuals in the population represent the overall deterioration sequence. The coefficient of variation of the overall deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase were calculated using the parent population. The coefficient of variation of the comprehensive deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase are sorted non-dominatedly, and the parent population is selected, crossed over and mutated using a genetic algorithm to generate the s-th generation sub-population; where the initial value of s is 1. The parent and offspring populations were merged, and the coefficient of variation of the comprehensive deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration operation phase were non-dominated and sorted, and the individual crowding and non-matching order were calculated respectively. Select several individuals based on the non-dominated order and individual crowding to update the parent population, increment the value of s by 1; and return the steps to calculate the coefficient of variation of the overall deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration phase. Once the preset number of iterations is reached, the primary objective is to maximize the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase, while the secondary objective is to minimize the coefficient of variation of the comprehensive deterioration sequence under the healthy operation phase. The weights of the sub-deterioration sequences are then selected from the optimal solution of the multi-objective optimization problem.

2. The method for assessing the deterioration degree of the shaft system of a hydroelectric generator unit according to claim 1, characterized in that, Obtain the classification condition labels for vibration waveform samples, specifically including: Several highly correlated feature vectors of operating conditions are randomly selected as initial cluster centers; K-means clustering analysis was used to calculate the distance from the feature vectors of the remaining highly correlated operating conditions to each cluster center, classify the feature vectors of the remaining highly correlated operating conditions, and calculate and update the cluster centers of each category. Based on the correspondence between vibration waveform samples and feature vectors of high-performance operating condition parameters, the classification of vibration waveform samples is obtained; Based on the number of classifications of the vibration waveform samples, the classifications of the vibration waveform samples are converted into one-hot codes of the corresponding dimensions to obtain the classification condition labels of the vibration waveform samples.

3. The method for assessing the deterioration degree of the shaft system of a hydropower unit according to claim 1, characterized in that, The health model includes a generator and a discriminator. The training process of the health model specifically includes: A generator is constructed using a one-dimensional deconvolutional network. The generator takes random noise and the classification labels corresponding to the actual healthy waveform samples during the healthy operation phase as input to generate the distribution of healthy waveform samples under different operating conditions during the healthy operation phase. A discriminator is constructed using a one-dimensional convolutional network. The discriminator takes the generated health waveform samples of the healthy operation phase and the corresponding actual health waveform samples and their operating condition labels as input to determine the authenticity of the generated health waveforms of the healthy operation phase. The objective function is to minimize the difference between the generated healthy waveform samples and the corresponding actual healthy waveform samples during the healthy operation phase, and the healthy model is trained using binary cross-entropy loss as the loss function.

4. The method for assessing the deterioration degree of the shaft system of a hydroelectric generator unit according to claim 3, characterized in that, The classification labels of the degraded waveform samples are sequentially input into the trained health model to generate corresponding healthy waveforms. The degradation degree is obtained by comparing the actual degraded waveforms with the generated healthy waveforms, and the degradation degree is sorted by time to construct a sub-degradation sequence, specifically including: Input the classification condition labels of the degraded waveform samples in the degraded operation phase into the generator, and obtain the healthy waveforms generated under each condition in the degraded operation phase in sequence. Calculate the degradation index of the actual degradation waveform and the generated healthy waveform during the degradation operation phase, and take the difference between the two as the degree of unit degradation. The deterioration degree of the units in the deterioration operation stage is sorted by time to construct a sub-deterioration sequence.

5. The method for assessing the deterioration degree of the shaft system of a hydropower unit according to claim 4, characterized in that, The degradation indicators include: time-domain degradation indicators, frequency-domain degradation indicators, and data distribution indicators; The time-domain degradation index is the mean and standard deviation of the actual degraded waveform or the generated healthy waveform during the degradation operation phase; the frequency-domain degradation index is the frequency standard deviation of the actual degraded waveform or the generated healthy waveform; the data distribution index is kurtosis and JS divergence.

6. A system for assessing the deterioration of a hydroelectric generator shaft system, characterized in that, The system includes: The operating condition parameter screening module is used to analyze the correlation between the mean of the healthy waveform samples of the hydropower unit during the healthy operation phase and the corresponding initial operating condition parameters using the maximum information coefficient (MIC). It then filters out highly correlated operating condition parameters whose MIC index is greater than a threshold, forming a feature vector of highly correlated operating condition parameters. Initial operating parameters include: head, active power, reactive power, excitation voltage, excitation current, opening degree, and rotational speed; The classification condition label acquisition module performs K-means clustering analysis on the feature vectors of highly correlated operating condition parameters corresponding to vibration waveform samples, and combines the one-hot encoding method to obtain the classification condition labels of vibration waveform samples; among them, vibration waveform samples include healthy waveform samples in the healthy operation stage and deteriorated waveform samples in the deteriorated operation stage. The health model construction module uses health waveform samples from the health operation phase and corresponding classification condition labels to train a health model based on a conditional convolutional adversarial generative network. The degradation sequence construction module sequentially inputs the classification condition labels of the degradation waveform samples into the trained health model to generate corresponding health waveforms. It obtains the degradation degree by comparing the actual degradation waveform with the generated health waveform and constructs sub-degradation sequences by sorting the degradation degree by time. The hydropower unit shaft system assessment module is used to calculate the comprehensive deterioration sequence based on the sub-deterioration sequence and its weight, and to assess the degree of deterioration of the hydropower unit shaft system. The determination of the weights of the sub-degraded sequences specifically includes: Substitute the weights of the N initial sub-deterioration sequences into the calculation formula of the overall deterioration sequence to generate an initial population of size N as the parent population; where individuals in the population represent the overall deterioration sequence. The coefficient of variation of the overall deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase were calculated using the parent population. The coefficient of variation of the comprehensive deterioration sequence during the healthy operation phase and the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase are sorted non-dominatedly, and the parent population is selected, crossed over and mutated using a genetic algorithm to generate the s-th generation sub-population; where the initial value of s is 1. The parent and offspring populations were merged, and the coefficient of variation of the comprehensive deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration operation phase were non-dominated and sorted, and the individual crowding and non-matching order were calculated respectively. Select several individuals based on the non-dominated order and individual crowding to update the parent population, increment the value of s by 1; and return the steps to calculate the coefficient of variation of the overall deterioration sequence in the healthy operation phase and the mean difference between the sub-deterioration sequences in the healthy operation phase and the deterioration phase. Once the preset number of iterations is reached, the primary objective is to maximize the mean difference between the sub-deterioration sequences under the healthy operation phase and the deterioration operation phase, while the secondary objective is to minimize the coefficient of variation of the comprehensive deterioration sequence under the healthy operation phase. The weights of the sub-deterioration sequences are then selected from the optimal solution of the multi-objective optimization problem.

7. The hydropower unit shaft system deterioration assessment system according to claim 6, characterized in that, The classification condition label acquisition module includes: The initial cluster center selection unit is used to randomly select several highly correlated working condition parameter feature vectors as initial cluster centers; The cluster center update unit is used to calculate the distance from the feature vectors of the remaining highly correlated operating conditions to each cluster center using K-means clustering analysis, classify the feature vectors of the remaining highly correlated operating conditions, and calculate and update the cluster centers of each category. The classification unit for vibration waveform samples is used to obtain the classification of vibration waveform samples based on the correspondence between vibration waveform samples and feature vectors of high-performance operating condition parameters. The classification condition label acquisition unit is used to convert the classification of vibration waveform samples into one-hot encoding of the corresponding dimension based on the number of classifications of the vibration waveform samples, and to obtain the classification condition label of the vibration waveform samples.

8. The hydropower unit shaft system deterioration assessment system according to claim 6, characterized in that, The health model construction module includes: The generator, constructed using a one-dimensional deconvolutional network, is used to generate the distribution of health waveform samples under different operating conditions during the health operation phase, taking random noise and the classification labels corresponding to the actual health waveform samples during the health operation phase as input. The discriminator, constructed using a one-dimensional convolutional network, is used to determine the authenticity of the generated healthy waveforms during the healthy operation phase, taking the generated healthy waveform samples and the corresponding actual healthy waveform samples and their operating condition labels as inputs; the generator and discriminator together form a health model. The building unit is used to train the health model with the objective function of minimizing the difference between the generated healthy waveform samples and the corresponding actual healthy waveform samples, and with binary cross-entropy loss as the loss function.

9. The hydropower unit shaft system deterioration assessment system according to claim 6 or 7, characterized in that, The degraded sequence construction module includes: The unit deterioration degree calculation unit is used to calculate the deterioration index of the actual deteriorated waveform and the generated healthy waveform during the deterioration operation stage, and the difference between the two is the unit deterioration degree. The method for obtaining the healthy waveform generated during the deterioration operation phase is as follows: input the classification condition label of the deterioration waveform sample during the deterioration operation phase into the generator, and obtain the healthy waveform generated under each condition during the deterioration operation phase in sequence. The sub-deterioration sequence construction unit is used to construct a sub-deterioration sequence by sorting the degree of unit deterioration in the deterioration operation phase according to time.

Images