Hydroelectric generating set degradation trend prediction method based on multi-measuring-point fusion and CNN-Informer

Abstract

The invention relates to the technical field of hydroelectric generating set operation monitoring, in particular to a hydroelectric generating set degradation trend prediction method based on multi-measuring-point fusion and CNN-Informer, which comprises the following steps of: firstly, collecting original data of key measuring points of a hydroelectric generating set, and screening stable operation data through a sliding window and a fluctuation threshold; constructing reference probability distribution based on kernel density estimation of a healthy sample, and generating a single-measuring-point degradation degree index in combination with a negative logarithm likelihood probability; an entropy weight method is adopted to carry out objective weighted fusion on the multiple measurement point indexes to obtain a comprehensive deterioration degree index; and finally, constructing a CNN-Informer model, extracting a local time sequence mode by using the CNN, and capturing a long-time dependency relationship by the Informer to realize degradation trend prediction. Actual data verification of a certain 200MW hydroelectric generating set in China shows that the method can clearly reveal the evolution process of the set from health to deterioration, the prediction performance is better than that of CNN-LSTM, LSTM, GRU and other models, and a new path is provided for health assessment and intelligent operation and maintenance of the hydroelectric generating set.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water turbine unit operation monitoring, in particular to a water turbine unit degradation trend prediction method based on multi-measurement point fusion and CNN-Informer. BACKGROUND

[0002] As a key equipment in the power system, the operation state of the water turbine unit is directly related to the stable supply of energy and the safe service life of the equipment. With the long-time operation of the unit, the internal components of the unit will gradually produce performance degradation. If the degradation trend cannot be identified and predicted in time, the operation efficiency will decrease and even serious faults will occur. Therefore, how to scientifically and effectively construct a degradation index that can reflect the health level of the water turbine unit and realize trend prediction on this basis has become an important research direction of operation monitoring and intelligent operation and maintenance in the field of hydropower.

[0003] At present, the construction of the degradation index mainly relies on single measurement point feature extraction or traditional statistical methods. For example, common methods include threshold-based monitoring and trend line-based index fitting. These methods can represent the state of the unit to some extent, but there are two shortcomings: first, they rely on a single signal and cannot fully reflect the multi-dimensional operation characteristics of the unit; second, the index construction process is highly subjective and prone to human bias. There is an urgent need for a degradation index construction method that can fully integrate multi-measurement point information, reduce the influence of human factors, and have strong robustness and universality, so as to more objectively and comprehensively reflect the operation state of the water turbine unit.

[0004] Deep learning methods have been widely used in mechanical equipment health management due to their strong feature extraction and non-linear modeling capabilities. Classic recurrent neural networks, long short-term memory networks, and convolutional neural networks have achieved remarkable results in modeling time series data. However, these methods still have some shortcomings: for example, LSTM is less efficient in modeling long sequences, and CNN has difficulty in capturing long-range dependency information. In recent years, the Informer model based on the Transformer architecture improvement has shown higher efficiency and accuracy in long sequence prediction. Its sparse self-attention mechanism can effectively alleviate the computational complexity and improve the prediction performance. Therefore, combining CNN and Informer can achieve high-precision prediction of degradation trends while considering local feature extraction and long-term dependency modeling. SUMMARY

[0005] The present application aims to provide a water turbine unit degradation trend prediction method based on multi-measurement point fusion and CNN-Informer to solve the problems of existing single-measurement point degradation indexes that cannot fully reflect the operation state of the unit and the lack of long sequence modeling accuracy and robustness in existing prediction models.

[0006] To achieve the above object, the present application provides the following technical solutions: A hydropower unit degradation trend prediction method based on multi-measurement-point fusion and CNN-Informer, comprising the following steps: S1 stable operation data screening: the stability of the original operation data of the hydropower unit is screened through a sliding window and a fluctuation threshold, and the stable operation interval data is retained; S2 single measurement point degradation index construction: for each key measurement point, a reference probability distribution is constructed based on kernel density estimation using healthy operation data, and the negative log-likelihood probability of a new sample is calculated as a single measurement point degradation index; S3 multi-measurement-point index fusion: the entropy weight method is used to objectively weight and fuse multiple single measurement point degradation indexes to generate a comprehensive degradation degree index; S4 degradation trend prediction: a CNN-Informer model is constructed, the time series of the comprehensive degradation degree index is taken as the input of the model, the local time sequence pattern is extracted through CNN, the long-time dependence relationship is captured through Informer, and the hydropower unit degradation trend prediction result is output.

[0007] Preferably, in step S1, the length of the sliding window is set to 10 and the fluctuation threshold is set to 5% during the stable operation data screening, and when the fluctuation amplitude of the active power in the continuous 10 sampling points is not more than 5%, it is determined that the interval is in a stable operation state.

[0008] Preferably, in step S2, the probability density function calculation formula of the kernel density estimation is: ; In the formula, n represents the sample number, h represents the bandwidth parameter, x represents the test point for which the density needs to be calculated, x i represents the i th healthy sample point, is a kernel function, and the commonly selected one is a Gaussian kernel: ; In the formula, u represents the standardized distance, and exp represents the exponential function; The bandwidth parameter h > 0 is calculated by the Silverman criterion as follows: ; In the formula, h represents the calculated optimal bandwidth, sigma represents the standard deviation of the sample data, IQR represents the interquartile range of the sample data, n represents the sample number, and the constants 0.9 and 1.34 are empirical coefficients for adjusting the robustness of the bandwidth.

[0009] Preferably, in step S2, the calculation formula of the negative log-likelihood probability in the single measurement point degradation index construction is: ; In the formula, NLLP(x t ) indicates at time The negative log-likelihood probability value is used as an indicator of degradation. This represents a test sample at a specific moment. Indicates test sample The probability density value under the healthy distribution is obtained by kernel density estimation.

[0010] Preferably, in the multi-measurement point index fusion step S3, the specific implementation of the entropy weight method includes: S31: The degradation index of a single measuring point is dimensionlessly processed to obtain the standardized matrix Z. ; S32: Calculate the scaling matrix : ; In the formula, Indicates the first The measuring point at the ... The proportion of each time point This represents the standardized degradation index value, where Represents time series sample points ( ), Indicates the measurement point number n represents the total number of sample points in the time series, m represents the total number of measurement points, and ε is a very small constant; S33: Based on this, define information entropy E j : ; In the formula, Indicates the first The information entropy value of each measurement point, k is the normalization coefficient, ensuring that the entropy value is within the range of [0,1]; S34: Calculate the difference coefficient d j and weight w j : ; ; In the formula, Indicates the first The degree of difference between measurement points is such that the smaller the entropy value, the greater the difference. Indicates the first The objective weight of each measurement point; S35: Generate comprehensive degradation index F i : .

[0011] As preferred, in the deterioration trend prediction of step S4, the CNN-Informer model combines a convolutional neural network and an improved Transformer architecture, and the calculation complexity of the self-attention mechanism in the standard Transformer is , wherein is the sequence length, and the calculation formula is: ; wherein the Informer module adopts a probabilistic sparse self-attention mechanism, and the calculation formula is: ; In the formula, respectively represent a query matrix, a key matrix and a value matrix, is the feature dimension, is the sequence length, represents a subset of key query vectors obtained through probability screening, represents the transpose of the key matrix.

[0012] As preferred, in step S4, the CNN adopts a one-dimensional convolutional layer, the convolution kernel size is set to 3x1, the convolution step is 1, the padding mode is same, and the activation function adopts a ReLU function for extracting local feature mutation information in the time series of the comprehensive deterioration degree index, and the expression of the ReLU function is: .

[0013] As preferred, in step S4, the training process of the CNN-Informer model adopts an adaptive matrix estimation Adam optimizer, the initial value of the learning rate is set to 1x10 -4 , the learning rate decay strategy adopts cosine annealing, the loss function adopts a mean square error function, and the expression of the mean square error function is: , wherein N is the sample number, is the true value of the comprehensive deterioration degree index at t time, is the predicted value at t time; and the training is stopped when the loss function does not decrease for 5 consecutive rounds on the validation set.

[0014] As preferred, the key measuring points include water guide bearing X-direction vibration, upper guide bearing X-direction vibration and thrust swing peak-to-peak value, which are used for comprehensively reflecting the unit operation state.

[0015] As preferred, the time series of the comprehensive deterioration degree index is subjected to smoothing processing to reduce the influence of abnormal fluctuations on the prediction model and highlight the long-term degradation trend.

[0016] Compared with the prior art, the present application has the following beneficial effects: The multi-measurement point fusion degradation trend prediction method for hydroelectric generating units of the application utilizes stable operation data, constructs a single-point degradation index based on probability density estimation, and uses negative log likelihood probability (NLLP) to measure the deviation degree of the measurement point; then, the objective weighting of the multi-measurement point index is realized by the entropy weight method to form a comprehensive degradation degree index; finally, the CNN-Informer model is introduced to predict the time series of the fusion index, in which CNN is used to extract local change patterns and Informer is used to capture long-time dependence. The method of the application not only overcomes the limitations of traditional single-point indexes and subjective weights, but also takes into account local and global feature modeling in the prediction link, providing a new idea and tool for the operation state evaluation and health management of hydroelectric generating units.

[0017] The multi-measurement point fusion degradation trend modeling and prediction method of the application screens through stable operation data, constructs a single-measurement point degradation index by combining kernel density estimation and negative log likelihood probability, and realizes the objective fusion of multi-measurement points by using the entropy weight method to finally form a comprehensive degradation degree index. On this basis, the CNN-Informer model is introduced to realize the time series prediction of the fusion index. Experimental results show that this method can effectively depict the degradation evolution process of hydroelectric generating units from health to failure, and is superior to traditional deep learning models in terms of prediction accuracy and robustness, verifying the feasibility and effectiveness of the method of the application. The specific advantages are: (1) The fusion degradation index is superior to the single-point index. The single-measurement point NLLP can reflect the changes in the operation state of local components, but due to differences in sensitivity and response, it has the problem of incomplete description. The multi-measurement point fusion based on the entropy weight method effectively integrates different signal characteristics, avoids the limitations of single points, and can more objectively and comprehensively reflect the degradation level of the unit as a whole, and show early warning ability before failure occurs.

[0018] (2) The CNN-Informer model has the best prediction performance. In the prediction of the fusion degradation index, CNN-Informer is significantly superior to CNN-LSTM, GRU and LSTM in terms of error indicators such as MSE, MAE and MAPE, showing higher accuracy in describing trend mutations and fluctuation characteristics, and verifying its advantages and application potential in hydroelectric generating unit health prediction. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to make further detailed explanations, but do not constitute a limitation on the application.

[0020] Figure 1 The comparison chart before and after the stable screening of the hydroelectric generating unit operation data of the application; Figure 2This is a graph showing the calculation results of the degradation degree of the key measuring points in this invention; Figure 3 This is a fusion degradation index diagram constructed based on the entropy weight method in this invention; Figure 4 The graph shows the prediction results of different models of this invention for the fusion degradation index; Figure 5 This is a flowchart of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] The method proposed in this invention includes four main steps: data stability region screening, construction of measurement point degradation index, entropy weight fusion, and depth prediction. (1) Screening of stable operation data: The guide bearing and main shaft of the hydropower unit have large fluctuations during unstable operation. In order to ensure the reliability of subsequent analysis, the original operation data must first be screened for stability. Stable operation data is screened by setting a sliding window and a fluctuation threshold. The window length is set to 10 and the fluctuation threshold is set to 5%. That is, when the fluctuation of active power within 10 consecutive sampling points does not exceed 5%, the interval is determined to be in a stable operation state.

[0023] (2) Construction of the measurement point deterioration index: Based on stable operating data, a reference probability distribution is constructed using healthy samples and formula (1) for the key measurement point signals of the hydropower unit, and its corresponding probability density function is obtained. New samples during operation are input into this distribution model to calculate their probability density values. Formula The defined negative log-likelihood probability is used as an index of single-point degradation: In the formula: Represents a test sample at a specific moment; This represents the probability density of the sample under a healthy distribution; When the operational data of a certain measuring point deviates from the healthy distribution, its probability density value decreases significantly, and the corresponding NLLP value increases significantly, thus effectively reflecting the degree of degradation.

[0024] (3) Integration of multi-measurement indicators: Considering that a single measurement point cannot fully reflect the complexity of the unit's operating status, this invention further employs the entropy weight method to integrate deterioration indicators from multiple measurement points. The basic idea of ​​the entropy weight method is to use information entropy to measure the degree of uncertainty of the indicators and determine the objective weight of each indicator through differences, thereby achieving a comprehensive representation of information from multiple measurement points. The steps are as follows: First, the degradation indices at each measuring point are dimensionless to obtain a standardized matrix. ,in Represents time series sample points, This indicates the measurement point number. Next, the scale matrix is ​​calculated. Based on this, we define information entropy: Calculate the difference coefficient and weights: Finally, the overall degradation index is obtained: (4) Deterioration trend prediction: After obtaining the time series of fused deterioration indexes, a deep learning prediction model is constructed to model its future trend. The CNN-Informer model used combines a convolutional neural network with an improved Transformer architecture: CNN can extract local temporal patterns in the sequence, while Informer efficiently captures long-term dependencies through a sparse self-attention mechanism. The combination of the two ensures the sensitivity of local features and improves the efficiency and accuracy of long sequence modeling, thereby enabling high-precision prediction of the deterioration trend of hydropower units.

[0025] Kernel density estimation Kernel density estimation (KDE) is a commonly used nonparametric probability density estimation method. Its core idea is to approximate the population distribution using sample points without prior distribution assumptions. KDE is continuous, smooth, and sensitive to bandwidth parameters, enabling more refined modeling of data distribution characteristics under finite sample conditions. Due to its widespread application in anomaly detection and health status modeling, KDE is widely used to construct degradation indicators and measure the difference between data and a reference distribution.

[0026] Let random variable The observed samples are Then the probability density estimation function given by KDE is: (1) in For kernel functions, a common choice is the Gaussian kernel: (2) and The bandwidth parameter controls the smoothness of the estimate. Too large a bandwidth leads to over-smoothing, while too small a bandwidth can cause oscillations. A commonly used method for selecting bandwidth is the Silverman criterion. (3) The Informer model is an improved Transformer-based architecture proposed in recent years for time series forecasting. Its core objective is to address the high computational complexity, large memory overhead, and redundant attention distribution issues of traditional Transformers in long-sequence forecasting tasks. Compared to the classic self-attention mechanism, Informer significantly reduces computational costs while maintaining prediction accuracy, making it more suitable for large-scale long-sequence forecasting tasks. In device degradation trend prediction, Informer can effectively capture long-term dependencies in time series and improve prediction efficiency through an efficient sparse attention mechanism.

[0027] In the standard Transformer, the computational complexity of the self-attention mechanism is O(n log n). ,in Let be the sequence length. Its calculation formula is: in, These represent the query, key, and value matrices, respectively. For feature dimension, The sequence length is given.

[0028] To improve efficiency, Informer introduces a probabilistic sparse self-attention mechanism, the core idea of ​​which is to retain only the most representative terms in the attention distribution. Its formal representation is: in, This represents the "key query vector subset" obtained through probability filtering, thus reducing the complexity to... .

[0029] Example Experimental Data Introduction To verify the applicability and effectiveness of the proposed state identification method in hydropower units, actual operating data of Unit 3 of a domestic hydropower station was selected as the research object. The turbine model of this unit is ZZA315-LJ-800, with a rated speed of 107.1 r / min, a rated power of 200 MW, and a rated head of 47 m. During operation at the end of August 2015, significant abnormal sounds were observed in the turbine chamber, upper frame, spiral casing, and tailrace pipe, along with severe unit vibration. Subsequent inspection by on-site personnel confirmed that the runner ring steel plate had detached during this period. This invention collected 507 sample points, including the X-axis vibration of the water guide, the X-axis vibration of the upper guide, and the peak-to-peak value of the thrust swing, from normal operation to failure.

[0030] Figure 1 The above comparison shows the results before and after the X-axis swing peak-to-peak value and active power steady-state screening. Figure 1 (a) is the complete experimental data of the unit operation. There are many non-steady-state operating conditions in the data. If such unstable intervals are directly used for degradation modeling, noise can be easily introduced and the reliability of the indicators can be interfered with. Figure 1 (b) is the data after steady-state screening. The retained samples show a more stable trend in both active power and peak-to-peak value.

[0031] Figure 2 The evolution of the NLLP (Neutral Level Limit) degradation of the water guide bearing and the upper guide bearing is presented. Overall, the NLLP at both measuring points shows an early stable trend followed by a significant increase, revealing the transition of the unit from a healthy to a deteriorating state. However, there are differences in the timing and magnitude of degradation at different measuring points. The water guide bearing shows obvious signs of degradation at approximately 200 sample points, with the NLLP value starting to rise continuously and fluctuate. In contrast, the upper guide bearing remains relatively stable at the same stage, only showing a sudden change later. This difference indicates that a single measuring point can only reflect the degradation characteristics of a local component and is insufficient to comprehensively characterize the overall state of the unit. Therefore, it is necessary to fuse information from multiple measuring points to obtain a more robust and representative comprehensive degradation index.

[0032] Figure 3This paper demonstrates the change of the fusion degradation index constructed based on the entropy weight method over time. During the healthy operation of the unit, the index remains at a low level with small fluctuations. As the unit gradually deteriorates, the fusion index gradually increases, and a significant abrupt change occurs after approximately the 200th sample point, with a rapid increase in value accompanied by strong fluctuations. After the 300th point, it rises significantly, with a peak value close to 0.9, indicating that the unit's operating state has significantly deviated from the healthy reference distribution. Multi-point fusion can comprehensively integrate information from different measurement points, avoiding the limitations of a single signal, clearly revealing the process of the unit's evolution from healthy to deteriorating, and possessing the ability to provide early warning. The red curve in the figure represents the smoothed sequence of the fusion degradation degree, highlighting long-term degradation characteristics while preserving the main trend and inflection points. The smoothed trend helps to intuitively describe the unit's degradation process, reducing the risk of overfitting the prediction model to abnormal fluctuations, thereby improving the stability and generalization ability of subsequent modeling.

[0033] Figure 4 This paper compares the true values ​​of the fusion degradation index with the prediction results of four deep learning models. Overall, all four models effectively capture the evolutionary characteristics of the fusion degradation index, which gradually increases from a low level and then rapidly abruptly changes after approximately the 250th sample point, demonstrating the strong fitting and predictive capabilities of deep learning methods in modeling complex degradation processes. However, differences exist in the details of the performance. CNN-Informer demonstrates greater accuracy in capturing the abrupt changes and subsequent oscillations of the fusion degradation index, maintaining its ability to depict long-term trends and exhibiting better robustness in predicting short-term fluctuations.

[0034] Table 1 shows the prediction performance results of different models. It can be seen that the CNN-Informer model outperforms other methods in all three evaluation metrics, with an MSE of 1.2015, a MAE of 6.824, and a MAPE of only 13.453%, and its prediction curve is closest to the actual trend. CNN-LSTM and LSTM exhibit lag or significant amplitude deviation in their response to deterioration mutations at some stages, resulting in MAPEs of 17.377% and 48.952%, respectively, indicating relatively insufficient generalization performance. The GRU model has significantly higher MSE and MAE than other models, showing the worst performance.

[0035] Table 1 Prediction performance results of different models This invention presents a multi-measuring-point fusion degradation trend prediction method for hydropower units. It utilizes stable operating data to construct single-point degradation indices based on probability density estimation, and employs negative log-likelihood probability (NLLP) to measure the deviation of each measuring point. Subsequently, the entropy weighting method is used to objectively weight the multi-measuring-point indices, forming a comprehensive degradation index. Finally, a CNN-Informer model is introduced to perform time-series prediction of the fused index, where CNN is used to extract local change patterns, and Informer is used to capture long-term dependencies. This invention not only overcomes the limitations of traditional single-point indices and subjective weights but also considers both local and global feature modeling in the prediction stage, providing a new approach and tool for hydropower unit operating status assessment and health management.

[0036] This invention presents a multi-measurement point fusion-based degradation trend modeling and prediction method. It filters steady-state operating data, combines kernel density estimation and negative log-likelihood probability to construct a single-measurement point degradation index, and then uses the entropy weight method to achieve objective fusion of multiple measurement points, ultimately forming a comprehensive degradation index. Based on this, a CNN-Informer model is introduced to achieve time-series prediction of the fused index. Experimental results show that this method can not only effectively characterize the degradation evolution process of hydropower units from healthy to faulty, but also outperforms traditional deep learning models in terms of prediction accuracy and robustness, verifying the feasibility and effectiveness of the method. Specific advantages include: (1) The fusion of degradation indicators is superior to that of single-point indicators. Single-point NLLP can reflect the changes in the operating status of local components, but due to differences in sensitivity and response, it has the problem of incomplete characterization. Multi-point fusion based on entropy weight method effectively integrates different signal characteristics, avoids the limitations of single points, can more objectively and comprehensively reflect the overall degradation level of the unit, and demonstrates the ability to provide early warning before the occurrence of failure.

[0037] (2) The CNN-Informer model has the best prediction performance. In the prediction of fusion degradation indicators, the CNN-Informer model is significantly better than the comparative models such as CNN-LSTM, GRU and LSTM in terms of error indicators such as MSE, MAE and MAPE. It shows higher accuracy in characterizing trend changes and fluctuations, which verifies its advantages and application potential in the health prediction of hydropower units.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the degradation trend of a hydroelectric generating unit based on multi-point fusion and CNN-Informer, characterized in that, The method comprises the following steps: S1 stable operation data screening: the stability of the original operation data of the hydroelectric generating set is screened through a sliding window and a fluctuation threshold, and the data in the stable operation interval is retained; S2 single measurement point degradation index construction: for each key measurement point, a reference probability distribution is constructed based on kernel density estimation using healthy operation data, and the negative log likelihood probability of a new sample is calculated as a single measurement point degradation index; S3 multi-measurement point index fusion: the entropy weight method is used to objectively weight and fuse multiple single measurement point degradation indexes to generate a comprehensive degradation degree index; S4 degradation trend prediction: a CNN-Informer model is constructed, the time series of the comprehensive degradation degree index is used as the input of the model, the local time sequence pattern is extracted through CNN, the long-time dependence relationship is captured through Informer, and the degradation trend prediction result of the hydroelectric generating set is output.

2. The method of claim 1, wherein the method is characterized by, In step S1, the length of the sliding window is set to 10, and the fluctuation threshold is set to 5%. When the fluctuation amplitude of the active power in the continuous 10 sampling points is not more than 5%, it is determined that the interval is in a stable operation state.

3. The method of claim 1, wherein the method is characterized by, In step S2, the probability density function of the kernel density estimation is calculated as follows: ； where n represents the number of samples, h represents the bandwidth parameter, x represents the test point for which the density needs to be computed, xi represents the i-th healthy sample point, i represents the i-th healthy sample point, is the kernel function, commonly chosen to be the Gaussian kernel: ； In the formula, u represents the standardized distance, and exp represents the exponential function; The bandwidth parameter h>0 is calculated by the Silverman criterion as follows: ； In the formula, h represents the calculated optimal bandwidth, sigma represents the standard deviation of the sample data, IQR represents the interquartile range of the sample data, n represents the sample quantity, and the constants 0.9 and 1.34 are empirical coefficients for adjusting the robustness of the bandwidth.

4. The method of claim 1, wherein the method is characterized by, In step S2, the calculation formula of the negative log likelihood probability is as follows: ； In the formula, NLLP(x t ) represents a negative log-likelihood probability value at time point , as a deterioration degree index, represents a test sample at a certain time point, represents a probability density value of the test sample under a healthy distribution, obtained by kernel density estimation.

5. The method of claim 1, wherein the method is characterized by, In step S3, the specific implementation of the entropy weight method includes: S31: Dimensionless processing is performed on the single-point degradation index to obtain a standardized matrix Z, ; S32: Calculate the proportion matrix : ； In the formula, represents the proportion value of the mth measurement point at the nth time point, represents the normalized degradation index value, wherein represents a time series sample point ( ), represents the measurement point number , n represents the total number of time series sample points, m represents the total number of measurement points, and ε is a very small constant. S33: On this basis, define the information entropy E j : ； In the formula, Indicates the first The information entropy value of each measurement point, k is the normalization coefficient, ensuring that the entropy value is within the range of [0,1]; S34: Calculate the difference coefficient d j and the weight w j : ； ； In the formula, represents the difference degree of the first point, the smaller the entropy value, the greater the difference, represents the objective weight of the first point; S35: generate the comprehensive deterioration degree index F i : 。 6. The method of claim 1, wherein the method is based on multi-point fusion and CNN-Informer. In the degradation trend prediction of step S4, the CNN-Informer model combines a convolutional neural network and an improved Transformer architecture. In the standard Transformer, the calculation complexity of the self-attention mechanism is wherein is the sequence length, and the calculation formula is as follows: ； The Informer module adopts a probabilistic sparse self-attention mechanism, and its calculation formula is as follows: ； wherein, denote a query matrix, a key matrix and a value matrix, respectively, is a feature dimension, is a sequence length, denotes a subset of key query vectors obtained by probability screening, denotes a transpose of the key matrix.

7. The method of claim 1, wherein the method is characterized by, In step S4, the CNN adopts a one-dimensional convolution layer, the convolution kernel size is set to 3x1, the convolution step is 1, the padding method is same, and the ReLU function is used as the activation function to extract the local feature mutation information in the time series of the comprehensive degradation degree index. The expression of the ReLU function is as follows: 。 8. The method of claim 1, wherein the method is characterized by, The training process of the CNN-Informer model in step S4 adopts an adaptive matrix estimation Adam optimizer, and the initial value of the learning rate is set to 1x10 -4 , and the cosine annealing is used as the learning rate decay strategy; The loss function adopts a mean square error function, and the expression of the mean square error function is as follows: where N is the number of samples, is the true value of the comprehensive degradation degree index at time t, is the predicted value at time t; the iteration training stops when the loss function does not decrease for 5 consecutive rounds on the validation set.

9. The method of claim 1, wherein the method is characterized by, The key measurement points include the X-direction vibration of the water guide bearing, the X-direction vibration of the upper guide bearing, and the peak-to-peak value of the thrust runout, which are used to comprehensively reflect the operation state of the unit.

10. The method of claim 1, wherein the method is based on multi-point fusion and CNN-Informer. The time series of the comprehensive degradation degree index is smoothed to reduce the influence of abnormal fluctuations on the prediction model and highlight the long-term degradation trend.

Images