Hydroelectric generating set fault diagnosis method based on chaos Kepler optimization algorithm

Abstract

The invention discloses a hydroelectric generating set fault diagnosis method based on a chaos Kepler optimization algorithm, and belongs to the technical field of hydroelectric generating set fault diagnosis. A Tent chaos function is introduced to replace an initial population of a Kepler optimization algorithm (KOA), and by improving the uniformity and uncertainty of the population, the convergence speed of the KOA algorithm is increased, and the global search capability is improved. An improved Kepler optimization algorithm is adopted to optimize hyper-parameters of the BiLSTM, a BiLSTM prediction model CKOA-BiLSTM of the hydroelectric generating set is established, the robustness and generalization ability of an existing model are improved, and the model is made to be suitable for processing complex nonlinear fault signals of the hydroelectric generating set. The model shows the optimal diagnosis effect on recognition of complex fault types, and the research can be used as a beneficial supplement of an existing hydroelectric generating set state monitoring and fault diagnosis system.

Description

Technical Field

[0001] The invention belongs to the technical field of hydropower unit fault diagnosis, and discloses a hydropower unit fault diagnosis method based on a chaotic Kepler optimization algorithm. Background Art

[0002] As the core equipment for energy conversion in the hydropower energy system, the safe and stable operation of the hydropower unit is the cornerstone of the entire hydropower generation system. Compared with other forms of power generation, hydropower units often encounter various faults in their complex and severe operating environment. At the same time, with the accelerated evolution of the power system towards adapting to large-scale and high-proportion new energy, wind power and photovoltaic power are connected to the grid on a large scale, and traditional hydropower units and pumped storage units are large-capacity regulating power sources. Frequent and rapid large-scale load adjustment will inevitably become the normal working condition of their operation. Rapid and frequent large-scale load adjustment will cause the unit to frequently pass through the vibration zone, and the sudden change of load will induce problems such as shaft system and runner fatigue, increasing the safety risk of hydropower unit operation, especially during the peak flood period, when a large amount of silt and large stones are carried in the water flow and continuously wear and hit the hydropower unit. These continuous impact and instantaneous impact failures not only lead to huge maintenance costs and economic losses, but also cause unpredictable accidents, posing a threat to the long-term stable and safe operation of the hydropower unit. Therefore, it is very important to diagnose and identify faults of hydropower units.

[0003] The essence of fault diagnosis of hydropower units is pattern recognition, which uses monitoring data to identify the location and severity of the fault, and provides guidance for formulating corresponding maintenance strategies. Traditional fault diagnosis requires a lot of prior knowledge and rich expert experience, and subjective factors have a great influence on the diagnosis results. In addition, it is difficult to analyze the cause of the fault from the massive unit monitoring data by relying solely on manual methods. With the development of machine learning, a number of classic learning algorithms have been applied to the research of vibration fault identification of hydropower units. In fault diagnosis based on machine learning, the model can reliably, efficiently and accurately identify the fault type according to the reasoning mechanism learned during training, replace the expert reasoning process, and reduce the dependence on manual experience. However, there are still great limitations in practical applications, such as the convergence of data processing, the generalization ability of model calculation, computing power and computing efficiency, etc., which need to be further optimized. At the same time, the accuracy and stability of the prediction results need to be improved. Therefore, it is urgent to develop a more reliable fault diagnosis method for hydropower units. Summary of the invention

[0004] The purpose of the present invention is to provide a hydropower unit fault diagnosis method based on the chaotic Kepler optimization algorithm to solve the problems of large calculation amount and low precision in the existing method, resulting in poor stability and unsatisfactory prediction results.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solution: a hydropower unit fault diagnosis method based on chaotic Kepler optimization algorithm, characterized by comprising the following steps:

[0006] (1) Collect the acoustic and vibration signals of the hydropower unit under normal operating conditions, sediment conditions, and bolt impact conditions;

[0007] (2) Generate an initial population through chaotic mapping and use it as the initial population of the Kepler optimization algorithm to obtain an improved Kepler optimization algorithm;

[0008] (3) Use the improved Kepler optimization algorithm to optimize the hyperparameters of the BiLSTM model and obtain the optimized BiLSTM prediction model;

[0009] (4) The acoustic vibration signals are used as feature sets and divided into training sets and test sets. The training set is input into the optimized BiLSTM prediction model to obtain the fault diagnosis model. The test set is input into the fault diagnosis model for state recognition, thus completing the fault diagnosis of the acoustic vibration signals of the hydropower unit.

[0010] A further technical solution is that the sediment concentration in the sediment working condition in step (1) is 1.4 kg / m 3 .

[0011] A further technical solution is that the specific steps of step (2) are as follows:

[0012] 2-1. Using Tent chaos mapping to obtain the initial population of planets in the Kepler optimization algorithm:

[0013]

[0014] Among them, xk is the value of the Tent map at the kth iteration, representing an individual position or value in the current population, which is used to generate the value of the next step. k+1 is the value of the Tent map at the k+1th iteration, used to generate the next individual position or value in the population.

[0015] 2-2. Find the gravitational force of the sun:

[0016]

[0017] Among them, e i is the eccentricity of the planetary orbit, ranging from 0 to 1; μ(t) is a function that controls the search accuracy, which decreases exponentially with time t; and is the normalized value of M s and m i , respectively represent the sun X s and target planet Xi The mass of the object; ε is a very small value; r1 is a random value between 0 and 1 to provide more variation to the gravity value during the optimization process; YesR i The normalized value of the sun X s and target planet X i The Euclidean distance between i The calculation formula is as follows:

[0018]

[0019] Among them, ||X S (t)-X i (t)||2 represents X S With X i The Euclidean distance between them; d represents the spatial dimension; X Sj (t) represents the coordinate of the sun in the jth dimension, which changes with time t; X ij (t) represents the coordinate of the target planet in the jth dimension, which changes with time t.

[0020] The masses of the Sun and planet i at time t are calculated by suitability assessment:

[0021]

[0022] Among them, fit s (t) represents the fitness value of the sun at time t; fit i (t) represents the fitness value of planet i at time t; fit k (t) represents the fitness value of the kth individual in the population at time t; worst(t) represents the worst fitness value of all individuals at time t, that is, the maximum fitness value; N represents the total number of individuals in the population; r2 is used to diverge the mass of the planet, and its value ranges from 0 to 1.

[0023] 2-3. Calculate the target planet velocity:

[0024] The velocity of a planet is calculated as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] Among them, V i (t) represents planet X i The speed at time t; r3, r4, r5 and r6 are randomly generated values ​​in the interval [0, 1]; is a random vector between 0 and 1; and represents a solution randomly selected from the population; R i (t) is the sun X at time t s and target planet X i The distance between i-norm (t) is the sun X at time t s and target planet X i The normalized value of the Euclidean distance between l represents random variables, vector directions and factors The coefficient of correlation; Represents a very small value to avoid division by zero errors; It is a symbolic variable with a value of 111 or -1-1-1, used to control the change of direction; is a binary variable used to control the direction; and Represent the target individual X i The upper and lower limits in each dimension are used to calculate the speed adjustment range; Defined as a random factor.

[0036] 2-4. Update the position of the target planet and its distance from the sun

[0037] The new position of each object moving away from the Sun:

[0038]

[0039] in, The target planet is X i The new position at time t+1; It's Planet X i The speed required to reach the new position; is the best position of the Sun found so far; h is an adaptive factor used to control the distance between the Sun and the current planet at time t.

[0040] A further technical solution is that the specific steps of step (3) are as follows

[0041] 3-1. Constructing a single-layer LSTM unit for a hydropower unit

[0042] The forget gate is used to limit the amount of fault feature information of the hydropower unit and filter out invalid information. The expression is as follows:

[0043] f t =σ[W f (h t-1 , x t )+b f ]

[0044] Among them, σ is the activation function; W is the weight coefficient matrix; h t-1 is the output state at the previous moment; x is the input feature vector; b is the bias vector;

[0045] The new fault feature information of the hydropower unit is put into the current cell state by the input gate and the information in the cell state is updated through the σ and tanh layers:

[0046]

[0047] In the above, I t is the output of the input gate; is the output of the tanh function; C t is the memory unit output;

[0048] The output gate transmits the fault characteristic information of the hydropower unit at the current moment through the output unit state C t , and get the unit output h t :

[0049]

[0050] 3-2. Use the state information of the two-layer bidirectional LSTM unit as the final output value y(t) of the BiLSTM model:

[0051] y(t)=W y h f (t)+x y h b (t)+b y

[0052] Among them, h f (t) and h b (t) represents the unit output to the forward layer and the backward layer respectively;

[0053] 3-3. Use the improved Kepler optimization algorithm to optimize the BiLSTM model hyperparameters bilstmLayerSizes and dropout Layer.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention proposes a hydropower unit fault diagnosis method based on the chaotic Kepler optimization algorithm, introduces the Tent chaotic function to replace the initial population of the Kepler optimization algorithm (KOA), accelerates the convergence speed of the KOA algorithm by improving the uniformity and uncertainty of the population, and improves the global search capability. The improved Kepler optimization algorithm is used to optimize the hyperparameters of BiLSTM, and a BiLSTM prediction model CKOA-BiLSTM of the hydropower unit is established to improve the robustness and generalization ability of the existing model, making it suitable for processing complex nonlinear fault signals of the hydropower unit. The experimental results show that the training accuracy and test accuracy of the CKOA-BiLSTM model are 97.6% and 98.3% respectively, which are better than the LSTM, BiLSTM and KOA-BiLSTM models. The CKOA-BiLSTM model shows the best diagnostic effect in the identification of complex fault types. This research can serve as a beneficial supplement to the existing hydropower unit state monitoring and fault diagnosis system. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the present invention.

[0057] Figure 2 It is a flow chart of the CKOA algorithm in the present invention.

[0058] Figure 3 is the model training curve in the embodiment.

[0059] Figure 4 It is a confusion matrix diagram of the model training in the embodiment.

[0060] Figure 5 2 is a diagram of the model test results in the embodiment. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0062] Figure 1A method for fault diagnosis of a hydropower unit based on a chaotic Kepler optimization algorithm is shown, which is characterized by comprising the following steps:

[0063] (1) Collect the acoustic and vibration signals of the hydropower unit under normal operating conditions, sediment conditions, and bolt impact conditions.

[0064] The signal is collected using a noise sensor (CRY2301) to collect data of normal turbine operation and sediment concentration of 1.4 kg / m 3 The acoustic vibration signals of three working conditions of bolt impact. The data collected from each working condition are sorted and segmented, and each sample contains 1024 data points. Each working condition contains 70 groups of samples, a total of 280 groups. The samples are randomly selected for division, 70% of the data is used as the training set and 30% of the data is used as the test set to form the final data set.

[0065] (2) An initialization population is generated through chaotic mapping and used as the initial population of the Kepler optimization algorithm to obtain the improved Kepler optimization algorithm CKOA.

[0066] The Kepler Optimization Algorithm (KOA) shows excellent search performance when dealing with optimization problems, but it has a potential problem that the initial population is generated in a random manner, which may cause the initial solutions to be unevenly distributed in the search space and lack diversity.

[0067] The chaotic mapping Tent function is a nonlinear dynamic system whose behavior is highly sensitive to initial conditions. This sensitivity means that even small initial differences can lead to extremely different results in the evolution of the system over time. The sequence generated by the chaotic mapping Tent function has good randomness and aperiodicity, which can effectively prevent the algorithm from falling into the local optimum, thereby enhancing the global search capability. In addition, the chaotic sequence generated by the Tent function is evenly distributed in the search space, which is conducive to more comprehensive global exploration. Its simple formula structure not only makes the calculation efficient, but also can flexibly adjust the balance between exploration and development during the optimization process, enhancing the adaptability and robustness of the algorithm.

[0068] Specific steps are as follows Figure 2 As shown, specifically:

[0069] 2-1. Using Tent chaos mapping to obtain the initial population of planets in the Kepler optimization algorithm:

[0070]

[0071] Among them, xk is the value of the Tent map at the kth iteration, representing an individual position or value in the current population, which is used to generate the value for the next step. k+1is the value of the Tent map at the k+1th iteration, used to generate the next individual position or value in the population.

[0072] 2-2. Find the gravitational force of the sun:

[0073]

[0074] Among them, e i is the eccentricity of the planetary orbit, ranging from 0 to 1; μ(t) is a function that controls the search accuracy, which decreases exponentially with time t; and is the normalized value of M s and m i , respectively represent the sun X s and target planet X i The mass of the object; ε is a very small value; r1 is a random value between 0 and 1 to provide more variation to the gravity value during the optimization process; YesR i The normalized value of the sun X s and target planet X i The Euclidean distance between i The calculation formula is as follows:

[0075]

[0076] Among them, ||X S (t)-X i (t)||2 represents X S With X i The Euclidean distance between them; d represents the spatial dimension; X Sj (t) represents the coordinate of the sun in the jth dimension, which changes with time t; X ij (t) represents the coordinate of the target planet in the jth dimension, which changes with time t.

[0077] The masses of the Sun and planet i at time t are calculated by suitability assessment:

[0078]

[0079] Among them, fit s (t) represents the fitness value of the sun at time t; fit i (t) represents the fitness value of planet i at time t; fit k (t) represents the fitness value of the kth individual in the population at time t; worst(t) represents the worst fitness value of all individuals at time t, that is, the maximum fitness value; N represents the total number of individuals in the population; r2 is used to diverge the mass of the planet, and its value ranges from 0 to 1.

[0080] 2-3. Calculate the target planet velocity:

[0081] The velocity of a planet is calculated as follows:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092] Among them, V i (t) represents planet X i The speed at time t; r3, r4, r5 and r6 are randomly generated values ​​in the interval [0, 1]; is a random vector between 0 and 1; and represents a solution randomly selected from the population; R i (t) is the sun X at time t s and target planet X i The distance between i-norm (t) is the sun X at time t s and target planet X i The normalized value of the Euclidean distance between l represents random variables, vector directions and factors The coefficient of correlation; Represents a very small value to avoid division by zero errors; It is a symbolic variable with a value of 111 or -1-1-1, used to control the change of direction; is a binary variable used to control the direction; and Represent the target individual X i The upper and lower limits in each dimension are used to calculate the speed adjustment range; Defined as a random factor.

[0093] 2-4. Update the position of the target planet and its distance from the sun

[0094] The new position of each object moving away from the Sun:

[0095]

[0096] in, The target planet is X i The new position at time t+1; It's Planet X i The speed required to reach the new position; is the best position of the Sun found so far; h is an adaptive factor used to control the distance between the Sun and the current planet at time t.

[0097] (3) Use the improved Kepler optimization algorithm to optimize the hyperparameters of the BiLSTM model and obtain the optimized BiLSTM prediction model CKOA-BiLSTM. The specific steps are as follows:

[0098] 3-1. Constructing a single-layer LSTM unit for a hydropower unit

[0099] The forget gate is used to limit the amount of fault feature information of the hydropower unit and filter out invalid information. The expression is as follows:

[0100] f t =σ[W f (h t-1 , x t )+b f ]

[0101] Among them, σ is the activation function; W is the weight coefficient matrix; h t-1 is the output state at the previous moment; x is the input feature vector; b is the bias vector;

[0102] The new fault feature information of the hydropower unit is put into the current cell state by the input gate and the information in the cell state is updated through the σ and tanh layers:

[0103]

[0104] In the above, I t is the output of the input gate; is the output of the tanh function; C t is the memory unit output;

[0105] The output gate transmits the fault characteristic information of the hydropower unit at the current moment through the output unit state C t , and get the unit output ht :

[0106]

[0107] 3-2. Use the state information of the two-layer bidirectional LSTM unit as the final output value y(t) of the BiLSTM model:

[0108] y(t)=W y h f (t)+x y h b (t)+b y

[0109] Among them, h f (t) and h b (t) represents the unit output to the forward layer and the backward layer respectively. CKOA is used to optimize the hyperparameters of the created BiLSTM model. The optimized BiLSTM hyperparameters are: bilstmLayerSizes = [11,52], dropoutLaye = 0.097127.

[0110] bilstmLayerSizes=[11,52] represents the number of units in each layer of the BiLSTM network. In this example, the first layer has 11 units and the second layer has 52 units. This structure allows the model to learn features at different levels. Generally, increasing the number of layers and units can improve the expressiveness of the model. dropoutLaye=0.097127 refers to the dropout rate applied during training. The dropout rate refers to the proportion of neurons randomly discarded in each training step. 0.097127 means that approximately 9.7% of the neurons will be randomly "turned off" in each iteration to prevent overfitting. In this way, the model can better generalize to unseen data.

[0111] Through this bidirectional structure, the memory unit of BiLSTM at each moment contains the past and future fault feature information, which can more fully learn the time dependency of the feature sequence and thus effectively diagnose the fault.

[0112] (4) The acoustic vibration signals are used as feature sets and divided into training sets and test sets. The training set is input into the optimized BiLSTM prediction model to obtain the fault diagnosis model. The test set is input into the fault diagnosis model for state recognition, thus completing the fault diagnosis of the acoustic vibration signals of the hydropower unit.

[0113] 4-1. Model training

[0114] The sorted data is imported into the CKOA-BiLSTM model for training. In order to highlight the effectiveness of the method proposed in this study, LSTM, BiLSTM and KOA-BiLSTM are selected as comparison models. The training accuracy curves of the four models are shown in Figure 2. Figure 3 shown.

[0115] from Figure 3 It can be seen that each model shows different degrees of accuracy improvement during the training process, but its final convergence level and fluctuation range are different. The accuracy of the LSTM model finally stabilized at 72.3%, showing large fluctuations and slow convergence speed. This is because LSTM cannot fully capture the bidirectional temporal dependencies in the data, resulting in limited generalization performance. In contrast, the BiLSTM model improves the accuracy to 76.8% by bidirectionally learning the two temporal information before and after, which is better than the LSTM model. However, there are still fluctuations during the training process, indicating that the model may be affected by the local optimal solution or poor hyperparameter selection. The Kepler Optimization Algorithm (KOA) is introduced to optimize the hyperparameters of the BiLSTM model. The accuracy of the KOA-BiLSTM model is significantly improved to 85.2%, showing better convergence speed and stability, indicating that KOA as a meta-heuristic algorithm can avoid the problem that the traditional gradient descent method is prone to fall into the local optimal solution, thereby optimizing the model performance. Furthermore, the accuracy of the CKOA-BiLSTM model reaches 95.7%, becoming the best performing model. Tent mapping enhances KOA's global search capability, improves optimization efficiency, and avoids the problem of local optimal solutions. The red curve shows that the CKOA-BiLSTM model has minimal fluctuations during training, showing high stability and generalization ability.

[0116] The loss rate curve reflects the gradual reduction of the model error during the training process and is an important indicator for measuring model performance. Figure 3It can be seen that the initial loss of the LSTM model is relatively high, close to 1.5. As the number of training rounds increases, the loss rate gradually decreases but eventually converges to about 0.8, and there are fluctuations in the later stage of training, indicating that its performance is limited in capturing the fault characteristics of hydropower units. In contrast, the initial loss rate of the BiLSTM model is lower than that of the LSTM, at 1.3, and decreases faster, and finally converges to about 0.6, indicating that BiLSTM effectively reduces the loss rate and improves learning ability by capturing bidirectional time series information. The initial loss rate of the KOA-BiLSTM model is further reduced to 1.2, and then quickly drops to about 0.5, indicating that the KOA algorithm accelerates the convergence of the model and reduces the fluctuation of the loss rate by optimizing the hyperparameters of the BiLSTM model. The initial loss rate of the CKOA-BiLSTM model is 0.9, which is significantly lower than that of other models. During the training process, the loss rate drops rapidly and stabilizes at about 0.2, with minimal fluctuations, showing excellent convergence and stability.

[0117] At the same time, we analyze the confusion matrices of the four models, because confusion matrix analysis can also reveal the performance differences of the models when dealing with common faults and rare faults, thereby optimizing the reliability and accuracy of the fault diagnosis system and ensuring higher fault tolerance in real applications. Figure 4 shown.

[0118] according to Figure 4 (a), For normal operating conditions, the LSTM model misclassified 41 normal operating condition samples as wear faults (category 2). For wear faults, the LSTM model only correctly classified 2 samples, and the remaining 41 samples were incorrectly identified as normal operating conditions. The impact fault classification effect was better, and the model correctly classified all 50 samples. This shows that LSTM performs poorly in distinguishing normal operating conditions from wear faults.

[0119] according to Figure 4 (b) Compared with LSTM, the number of BiLSTM misclassifications of normal conditions is reduced from 41 to 37. The performance of wear faults is also improved, with 6 samples correctly classified, but there are still 37 wear fault samples that are misclassified as normal conditions. The classification of impact faults remains correct. Although BiLSTM is better than LSTM in capturing bidirectional time series information, its ability to distinguish between wear faults and normal conditions is still limited.

[0120] according to Figure 4(c) The correct classification rate of KOA-BiLSTM under normal working conditions is significantly improved to 70.1%, and the number of misclassifications is reduced to 23. The classification performance of wear faults is also improved, with 20 samples correctly classified, but 23 samples are still misclassified as normal working conditions. Impact faults continue to maintain a classification accuracy of 100%. Overall, KOA optimization significantly improves the performance of BiLSTM under normal working conditions and wear faults, especially in reducing misclassification of normal working conditions.

[0121] according to Figure 4 (d) The correct classification rate of the CKOA-BiLSTM model for normal working conditions is as high as 94.6%, and only 5.4% of the samples are misclassified as wear faults. The classification performance of wear faults has also been greatly improved, with a correct classification rate of 97.6%, and only a few samples are misclassified as other categories. The classification accuracy of impact faults remains 100%. This result shows that CKOA greatly improves the classification ability of the BiLSTM model in fault diagnosis tasks, especially in distinguishing normal working conditions from wear faults, reducing misclassification and improving the stability and generalization ability of the model.

[0122] In summary, the performance of LSTM and BiLSTM in fault diagnosis tasks has obvious limitations, especially when distinguishing normal working conditions from wear faults. After the introduction of the Kepler optimization algorithm, the classification performance of the model is significantly improved, especially in reducing misclassification. After further combining the chaos Kepler optimization algorithm, the classification accuracy of the model in various working conditions and faults has reached a high level, demonstrating the powerful potential of the chaos optimization algorithm in complex classification tasks and significantly improving the reliability of the fault diagnosis system.

[0123] 4-2. Model testing

[0124] Test and verify the trained model, and the test results of the test set Figure 5 As shown. When analyzing the corresponding confusion matrix, we have mentioned the performance differences of different models in hydropower unit fault diagnosis. The test accuracy of the LSTM model is 70%. LSTM is relatively reliable in the diagnosis of normal working conditions and impact faults, but there is a significant deviation in the diagnosis of wear faults. Although the model can capture the fault characteristics to a certain extent, the diagnosis of wear faults is not accurate enough and there is a large error.

[0125] The test accuracy of the BiLSTM model is 74.4%. In contrast, the BiLSTM model has improved in the diagnosis of wear faults. Its ability to learn time series information bidirectionally makes BiLSTM more advantageous in fault diagnosis tasks, especially when dealing with complex fault types. However, despite the improved diagnosis effect, there is still a certain deviation.

[0126] The KOA-BiLSTM model has significantly improved performance in all working conditions, with a test accuracy of 82.6%. Especially in the diagnosis of wear faults, the model's error has been significantly reduced, showing better diagnostic accuracy and stability. The optimization algorithm effectively adjusts the model's parameters, making the model's performance more consistent under different working conditions.

[0127] The CKOA-BiLSTM model further enhances the diagnostic capability of the BiLSTM model, with a test accuracy of 98.3%. The diagnostic results under all working conditions almost completely match the actual working conditions. In particular, the CKOA-BiLSTM model shows excellent accuracy in the identification of wear and impact faults, verifying the effectiveness of chaos optimization in improving the global search capability of the model. With the introduction of the optimization algorithm and the increase in model complexity, the performance of the fault diagnosis model has been significantly enhanced, especially in the identification of complex fault types, the CKOA-BiLSTM model has shown the best diagnostic effect.

[0128] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hydropower unit fault diagnosis method based on chaotic Kepler optimization algorithm, characterized in that The steps include: (1) Collect the acoustic and vibration signals of the hydropower unit under normal operating conditions, sediment conditions, and bolt impact conditions; (2) Generate an initial population through chaotic mapping and use it as the initial population of the Kepler optimization algorithm to obtain an improved Kepler optimization algorithm; (3) Using the improved Kepler optimization algorithm to optimize the hyperparameters of the BiLSTM model, an optimized BiLSTM model is obtained; (4) The acoustic vibration signals are used as feature sets and divided into training sets and test sets. The training set is input into the optimized BiLSTM model to obtain the fault diagnosis model. The test set is input into the fault diagnosis model for state recognition, thus completing the fault diagnosis of the acoustic vibration signals of the hydropower unit.

2. The method according to claim 1, characterized in that: The sediment concentration in the sediment condition described in step (1) is 1.4 kg / m 3 .

3. The method according to claim 1, characterized in that: The specific steps of step (2) are as follows: 2-1. Using Tent chaos mapping to obtain the initial population of planets in the Kepler optimization algorithm: Among them, x k is the value of the Tent map at the kth iteration, representing an individual position or value in the current population, used to generate the value for the next step, x k+1 is the value of the Tent map at the k+1th iteration, which is used to generate the next individual position or value in the population; 2-2. Find the gravitational force of the sun: Among them, e i is the eccentricity of the planetary orbit, ranging from 0 to 1; μ(t) is a function that controls the search accuracy, which decreases exponentially with time t; and is the normalized value of M s and m i , respectively represent the sun X s and target planet X i The mass of the object; ε is a very small value; r1 is a random value between 0 and 1 to provide more variation to the gravity value during the optimization process; YesR i The normalized value of the sun X s and target planet X i The Euclidean distance between i The calculation formula is as follows: Among them, ||X S (t)-X i (t)||2 represents X S With X i The Euclidean distance between them; d represents the spatial dimension; X Sj (t) represents the coordinate of the sun in the jth dimension, which changes with time t; X ij (t) represents the coordinate of the target planet in the jth dimension, which changes with time t; The masses of the Sun and planet i at time t are calculated by suitability assessment: Among them, fit s (t) represents the fitness value of the sun at time t; fit i (t) represents the fitness value of planet i at time t; fit k (t) represents the fitness value of the kth individual in the population at time t; worst(t) represents the worst fitness value of all individuals at time t, that is, the maximum fitness value; N represents the total number of individuals in the population; r2 is used to diverge the mass of the planet, and its value ranges from 0 to 1. 2-3. Calculate the target planet velocity: The velocity of a planet is calculated as follows: Among them, V i (t) represents planet X i The speed at time t; r3, r4, r5 and r6 are randomly generated values ​​in the interval [0,1]; is a random vector between 0 and 1; and represents a solution randomly selected from the population; R i (t) is the sun X at time t s and target planet X i The distance between i-norm (t) is the sun X at time t s and target planet X i The normalized value of the Euclidean distance between Representation and random variables, vector directions, and factors The coefficient of correlation; Represents a very small value to avoid division by zero errors; It is a symbolic variable with a value of 111 or -1-1-1, used to control the change of direction; is a binary variable used to control the direction; and Represent the target individual X i The upper and lower limits in each dimension are used to calculate the speed adjustment range; Defined as a random factor; 2-4. Update the position of the target planet and its distance from the sun The new position of each object moving away from the Sun: in, The target planet is X i New position at time t+1; It's Planet X i The speed required to reach the new position; is the best position of the Sun found so far; h is an adaptive factor used to control the distance between the Sun and the current planet at time t.

4. The method according to claim 1, characterized in that: The specific steps of step (3) are as follows 3-1. Constructing a single-layer LSTM unit for a hydropower unit The forget gate is used to limit the amount of fault feature information of the hydropower unit and filter out invalid information. The expression is as follows: f t =σ[W f (h t-1 ,x t )+b f )] Among them, σ is the activation function; W is the weight coefficient matrix; h t-1 is the output state at the previous moment; x is the input feature vector; b is the bias vector; The new fault feature information of the hydropower unit is put into the current cell state by the input gate and the information in the cell state is updated through the σ and tanh layers: In the above,i t is the output of the input gate; is the output of the tanh function; C t is the memory unit output; The output gate transmits the fault characteristic information of the hydropower unit at the current moment through the output unit state C t , and get the unit output h t : 3-2. Use the state information of the two-layer bidirectional LSTM unit as the final output value y(t) of the BiLSTM model: y(t)=W y h f (t)+x y h b (t)+b y Among them, h f (t) and h b (t) represents the unit output to the forward layer and the backward layer respectively; 3-3. Use the improved Kepler optimization algorithm to optimize the BiLSTM model hyperparameters bilstmLayerSizes and dropoutLayer.

Images