Water turbine fault diagnosis method based on acoustic signals

Abstract

The invention discloses a water turbine fault diagnosis method based on acoustic signals, and relates to the technical field of water turbine fault diagnosis. Acoustic signals of the water turbine under the normal working condition, the sediment erosion working condition and the runner blade crack fault are collected; introducing a maximum information coefficient to perform feature selection on the acoustic signals under each working condition, and selecting representative points from the acquired acoustic signals as a new data set; building a CNN-BiLSTM model by using a convolutional neural network and a bidirectional long-short term memory network BiLSTM, and optimizing model hyper-parameters by using an improved eagle optimization algorithm; and dividing the data set into a training set and a test set, inputting the training set as an input data set into the CNN-BiLSTM model for training, and inputting the test set into the trained CNN-BiLSTM model to obtain a fault diagnosis result. By constructing an intelligent diagnosis method based on acoustic signals, accurate diagnosis of different types of runner faults is realized, and valuable supplement is provided for monitoring the state of the runner of the hydroelectric generating set.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water turbine fault diagnosis, and particularly relates to a water turbine fault diagnosis method based on acoustic signals. BACKGROUND

[0002] In recent years, with the grid connection of new energy such as wind power, solar energy and tidal energy, hydropower energy undertakes more peak regulation and frequency regulation tasks, and hydropower generating units show dynamic characteristics such as high-frequency start-stop and wide-load rapid regulation. As a key flow component of hydropower generating units, the water turbine runner needs to bear severe alternating stress and cyclic load impact for a long time, as well as complex hydraulic excitation coupling effects such as silt erosion, cavitation and air erosion. The runner blade is the position with the lowest strength and the most concentrated stress in the entire runner, and long-term water erosion will lead to a decrease in blade thickness and a change in stress structure, increasing the risk of crack failure. The water turbine runner is a core equipment for energy conversion, and its state directly affects the safe and stable operation of the hydropower generating unit and the power generation efficiency. Therefore, it is of great engineering significance to develop high-precision runner fault intelligent diagnosis technology.

[0003] The current fault diagnosis method commonly used in hydropower stations is post-maintenance, which has serious hysteresis. Acoustic monitoring technology, as a non-intrusive detection method with high sensitivity, has been widely used in the field of state monitoring of large machinery. However, the operating environment of hydropower generating units is often complex, and the collected sound signals often contain a large amount of redundant noise, which often overwhelms the effective fault features in the signals. Taking appropriate measures for feature selection can help improve signal quality and thus improve the efficiency and accuracy of fault diagnosis. Feature extraction and feature selection are the core links of data-driven fault diagnosis. Feature extraction generally generates new features by reprocessing data, but when facing high-dimensional acoustic signals, it requires a large amount of calculation and has the risk of introducing irrelevant features or new noise. Feature selection, on the other hand, selects a feature subset that is strongly related to the fault state and has low redundancy from the original signal based on statistical measures, thereby achieving data dimension reduction and reducing the risk of model overfitting. In addition to obtaining a high-discriminative feature dataset, building a classification model is still a key step in achieving accurate diagnosis of water turbine runner faults. Using artificial intelligence and machine learning techniques to automatically analyze and recognize acoustic signals can effectively improve the efficiency and accuracy of fault detection.

[0004] Therefore, it is urgent to develop a water turbine fault diagnosis method based on acoustic signals that integrates feature extraction, feature selection and classification model construction to ultimately improve the efficiency and accuracy of fault detection. SUMMARY

[0005] The application aims to provide an acoustic signal-based water turbine fault diagnosis method, and solve the problem of unsatisfactory fault detection efficiency and accuracy in the prior art.

[0006] To solve the above technical problems, the application adopts the following technical scheme: an acoustic signal-based water turbine fault diagnosis method, characterized by comprising the following steps: S1. Collecting acoustic signals of a water turbine under normal working conditions, silt erosion conditions and runner blade crack faults; S2. Introducing a maximum information coefficient to select features of acoustic signals under each working condition, from Selecting representative points from the acoustic signals collected in S1 as a new data set; S3. Building a CNN-BiLSTM model using a convolutional neural network and a bidirectional long short-term memory network BiLSTM, and using an improved hawk-eagle optimization algorithm to optimize the model hyperparameters; S4. Dividing the data set in step S2 into a training set and a test set, inputting the training set as an input data set into the CNN-BiLSTM model for training, and then inputting the test set into the trained CNN-BiLSTM model to obtain a fault diagnosis result.

[0007] A further technical solution is that the maximum information coefficient MIC has the following specific steps: S1-1. Dividing variables and into grids, assuming that is divided into intervals, is divided into intervals, and a grid is formed; then the mutual information of each grid is calculated: wherein the variable is set as an original signal, the variable is a label column corresponding to the signal, is a joint probability distribution of and , and are marginal probability distributions of and ; S1-2. Finding the grid division that maximizes the mutual information by combining different grids, calculating the maximum mutual information of each and , and normalizing it: in, It is standardized mutual information. It is the standardization factor, and MIC stands for Standardized Mutual Information. The maximum value, MIC means and The maximum correlation under optimal grid partitioning, and ; S1-3. Calculate the MIC of the acoustic signal and sort the features in descending order of MIC. Select the top 1%-2.5% of features as the new dataset.

[0008] A further technical solution is the improved egret optimization algorithm in step S3, the specific steps of which are as follows: S311. Initialize population positions using Circle chaotic mapping: It is the remainder function; S312. Obtain the initial position of each individual in the heron-eagle population: in, The total number of individuals in the group. The maximum dimension of the problem being sought. and Let represent the upper and lower bounds of the i-th individual in the j-th dimension. , To represent the heron and eagle group, Let the position of the i-th heron in the population be in the j-th dimension, which is a candidate solution to the optimization problem; and calculate the position of the i-th heron. fitness value : The optimized initial position expression is: In the formula: This represents the initial optimization position of the i-th individual in the j-th dimension; Indicates the upper and lower bounds of the search space; This represents the chaotic sequence generated by iterating the Circle chaotic map t times. This represents the chaotic sequence generated by iterating through the Circle chaotic map t+1 times; K represents the parameters of the Circle chaos map; Exploration Phase: Conduct a global search by mimicking the hunting strategies of herons and eagles. in, This represents the current iteration number. The maximum number of iterations is used; the hunting process is divided into three time periods. To find prey during the hunting season, To deplete the prey's stamina for a period of time, The time period for attacking prey; and This represents a random candidate solution in the first stage of iteration. For interval The dimensions of the internally randomly generated dimension are array, This represents the new position of the i-th heron in the j-th dimension. It is the current optimal solution. , Represents randomly generated from the standard normal distribution Array; , Here is the Levy flight distribution function: in, Let i be the objective function for the i-th heron during the exploration phase. The fitness value of the objective function during the exploration phase; Development phase: Localized search is conducted by mimicking the behavior of herons and eagles when evading predators. in, Let i be the objective function for the i-th heron during the development phase. The heron eagle uses environmental camouflage strategies. Choose to escape or fly away; A dimension generated from a normal distribution. A random array, For the current iteration, a random candidate solution. It can be 1 or 2; During the development phase, cosine similarity was introduced to update the location of the fleeing heron eagle. in, , , This represents the current iteration number. This represents the maximum number of iterations; where A and B are calculated as follows: by calculating vectors... and The cosine similarity is measured by the angle between the two sides. Construct vectors within the heron-eagle population. and as follows: For the individual's current location, It is the current optimal solution. This represents the position of a random individual in the population. Let be the direction vector from the current individual's position to the local optimum. This is the direction vector from the current individual's position to the random candidate solution.

[0009] A further technical solution is that the CNN-BiLSTM model consists of an input layer, a convolutional layer, a pooling layer, a BiLSTM layer, a fully connected layer, and a Softmax classification layer.

[0010] Compared with the prior art, the beneficial effects of the present invention are: Acoustic signals from a hydro turbine under conditions of silt erosion and blade cracking were collected. First, effective features of the signals were extracted using a microphone (MIC). Then, a hybrid model consisting of a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM) was used to identify the signals. To improve the performance of the CNN-BiLSTM model, an improved Infinite Stratos Borealis Optimization Algorithm (ISBOA) was proposed to optimize the model's hyperparameters. Finally, the proposed method was used to perform binary and tri-class classification of silt erosion and blade cracking faults, respectively. Comparative verification showed that the model achieved an accuracy of 96.38% for silt erosion fault signals and 91.25% for blade cracking fault signals. This fault diagnosis method has advantages such as high accuracy, stable performance, strong generalization ability, and robustness, providing a valuable supplement to the field of real-time condition monitoring of hydro turbine runners. Attached Figure Description

[0011] Figure 1 This is a flowchart of the improved Heron Eagle optimization algorithm.

[0012] Figure 2 The image shows the collected acoustic signal curve.

[0013] Figure 3 A schematic diagram for selecting MIC features.

[0014] Figure 4 The optimal curve for ISBOA in the fault diagnosis of sediment erosion.

[0015] Figure 5 Training curves for each model in the fault diagnosis of sediment erosion.

[0016] Figure 6 A scatter plot of the results of mud and sand erosion fault identification.

[0017] Figure 7 It serves as a model evaluation index in the diagnosis of sediment erosion faults.

[0018] Figure 8 This is the fitness curve for blade crack fault diagnosis.

[0019] Figure 9 This is a scatter plot of the blade crack fault identification results.

[0020] Figure 10 It serves as a model evaluation index in blade crack fault diagnosis.

[0021] Figure 11 This is the fitness curve for multi-condition fault diagnosis.

[0022] Figure 12 This is a scatter plot of the fault identification results under multiple operating conditions.

[0023] Figure 13 It serves as a model evaluation index in multi-condition fault diagnosis.

[0024] Figure 14 The image shows the sampling wheel and its flowchart in Example 1. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0026] A method for diagnosing turbine faults based on acoustic signals, characterized by the following steps: S1. Acoustic signals of the turbine under normal operating conditions, silt erosion conditions, and runner blade crack faults; S2. Introducing the maximum information coefficient to perform feature selection on the acoustic signal under each working condition, from... Representative points are selected from the acoustic signals collected in S1 as a new dataset; MIC is a statistical measure of the relationship between two variables, particularly suitable for high-dimensional data and nonlinear signals such as acoustic signals. The specific steps for obtaining the Maximum Information Coefficient (MIC) are as follows: S1-1. Change the variables and Divide into grids respectively, assuming Classified as Each interval Classified as If there are several intervals, then a set of intervals will be formed. The grid is then used to calculate the mutual information of each grid cell. : Among them, variables Set as the original signal, variable The label column corresponding to the signal. for and The joint probability distribution, and yes and The marginal probability distribution; S1-2. Through different Grid combination, finding mutual information The largest mesh division, calculate each and Maximum mutual information, and normalize it: in, It is standardized mutual information. It is the standardization factor, and MIC stands for Standardized Mutual Information. The maximum value, MIC means and The maximum correlation under optimal grid partitioning, and ; S1-3. Calculate the MIC of the acoustic signal and sort the features in descending order of MIC. Select the top 1%-2.5% of features as the new dataset.

[0027] S3. A CNN-BiLSTM model was built using a convolutional neural network and a bidirectional long short-term memory network (BiLSTM), and the hyperparameters of the model were optimized using an improved Heron Eagle optimization algorithm.

[0028] The Secretary Bird Optimization Algorithm (SBOA) is a metaheuristic swarm optimization algorithm inspired by the survival behaviors of secretary birds (secretary birds) when hunting prey and escaping predators. SBOA consists of two main phases: the exploration phase and the exploration phase. The exploration phase involves a global search, mimicking the hunting behavior of secretary birds. The exploration phase, on the other hand, mimics the behavior of secretary birds when escaping pursuit, selecting high-potential areas through environmental observation and conducting a high-progression local search.

[0029] The improved ISBOA optimization algorithm has the following specific steps: S311. In the original SBOA algorithm, the positions of the heron and eagle are initialized randomly. This method is very prone to causing clustering of initial solutions, further affecting the accuracy and efficiency of the optimization. Therefore, a chaotic mapping method is used to improve the SBOA algorithm. Chaotic mapping has strong ergodicity and randomness, which can generate diverse and uniformly distributed initial populations, improving the algorithm's global search capability. Initializing population positions using Circle chaotic mapping: It is the remainder function; S312. Obtain the initial position of each individual in the heron-eagle population: in, The total number of individuals in the group. The maximum dimension of the problem being sought. and Let represent the upper and lower bounds of the i-th individual in the j-th dimension. , To represent the heron and eagle group, Let the position of the i-th heron in the population be in the j-th dimension, which is a candidate solution to the optimization problem; and calculate the position of the i-th heron. fitness value : The optimized initial position expression is: In the formula: This represents the initial optimization position of the i-th individual in the j-th dimension; Indicates the upper and lower bounds of the search space; This represents the chaotic sequence generated by iterating the Circle chaotic map t times. This represents the chaotic sequence generated by iterating through the Circle chaotic map t+1 times; K represents the parameters of the Circle chaos map; Exploration Phase: Conduct a global search by mimicking the hunting strategies of herons and eagles. in, This represents the current iteration number. The maximum number of iterations is used; the hunting process is divided into three time periods. To find prey during the hunting season, To deplete the prey's stamina for a period of time, The time period for attacking prey; and This represents a random candidate solution in the first stage of iteration. For interval The dimensions of the internally randomly generated dimension are array, This represents the new position of the i-th heron in the j-th dimension. It is the current optimal solution. , Represents randomly generated from the standard normal distribution Array; , Here is the Levy flight distribution function: in, Let i be the objective function for the i-th heron during the exploration phase. The fitness value of the objective function during the exploration phase; Development phase: Localized search is conducted by mimicking the behavior of herons and eagles when evading predators. in, Let i be the objective function for the i-th heron during the development phase. The heron eagle uses environmental camouflage strategies. Choose to escape or fly away; A dimension generated from a normal distribution. A random array, For the current iteration, a random candidate solution. It can be 1 or 2; During the development phase, cosine similarity was introduced to update the location of the fleeing heron eagle. in, , , This represents the current iteration number. This represents the maximum number of iterations; where A and B are calculated as follows: by calculating vectors... and The cosine similarity is measured by the angle between the two sides. Construct vectors within the heron-eagle population. and as follows: For the individual's current location, It is the current optimal solution. This represents the position of a random individual in the population. Let be the direction vector from the current individual's position to the local optimum. Let be the direction vector from the current individual's position to the random candidate solution. By introducing cosine similarity, dynamic optimization of the escape strategy is achieved, balancing the algorithm's global coverage and local search capabilities. This ensures rapid convergence in the early stages and global search in the later stages, avoiding getting trapped in local optima.

[0030] A Convolutional Neural Network (CNN) is a multi-layered neural network with high fault tolerance and robustness, mainly consisting of an input layer, a convolutional layer (CONV), a pooling layer, a fully connected layer, and an output layer.

[0031] CONV is the most important step in CNN. It extracts data features by sliding an appropriate convolutional kernel across the input data. This process is called a convolution operation, and its mathematical principle is as follows: in, For input tensors , These are the height, width, and number of channels, respectively. convolution kernel , The kernel size is [size]. The number of convolution kernels, For bias.

[0032] LSTM is a special type of recurrent neural network (RNN) structure. BiLSTM uses two LSTM units to process data, enabling the neural network to better capture temporal and causal relationships in sequences, greatly enhancing its ability to understand and analyze sequence data. The forward LSTM layer... Forward and backward LSTM layers used to process sequences The expression used to reverse a sequence is as follows: The hidden states in both directions are concatenated at each time step to obtain the BiLSTM hidden layer: , These are the weighting coefficients.

[0033] Convolutional layers in CNN models can effectively improve model processing speed and generalization ability, extract higher-level features from the dataset, and enhance the model's recognition accuracy. However, CNNs are not adept at reading global information. BiLSTM models, composed of two LSTM layers, perform simultaneous forward and backward computations on the sequence data, better capturing global information. The CNN-BiLSTM model combines the advantages of both CNN and BiLSTM models, adaptively learning both local and global features of the dataset, making it more suitable for time series. Therefore, using the CNN-BiLSTM model can reduce data complexity and effectively extract the spatial and temporal features of the acoustic signals from the operation of a water turbine runner.

[0034] The CNN-BiLSTM model consists of an input layer, a convolutional layer, a pooling layer, a BiLSTM layer, a fully connected layer, and a Softmax classification layer.

[0035] S4. Divide the dataset in step S2 into a training set and a test set. Use the training set as the input dataset to train the CNN-BiLSTM model, and then input the test set into the trained CNN-BiLSTM model to obtain the fault diagnosis results.

[0036] Example 1 The above method was applied to experiments on a micro hydroelectric power generation test bench. First, acoustic signals from the turbine runner were collected under normal operating conditions. This involved installing a properly functioning runner, filling the water tank and trough with clean water, starting the unit according to the steps described above, and collecting signals. Then, acoustic signals were collected under sediment erosion conditions. By calculating the volume of water in the tailrace and water tank, an appropriate amount of sediment was added, controlling the sediment concentration at 0.73 kg / m³. 3 After starting the unit and waiting for it to stabilize, acoustic signals of the sand-laden water flow impacting the impeller were collected; finally, acoustic signals were collected under the condition of the impeller having cracks. It is important to note that the sediment in the unit system needs to be removed before the experiment to ensure that the impeller condition is the only variable. Impeller photographs for both operating conditions are shown below. Figure 14 As shown in (a)(b). Figure 14 (c) is a schematic diagram of the experimental platform for sediment collection.

[0037] The collected acoustic signals were integrated to obtain a dataset. Data from the normal turbine runner condition was labeled "1", data from the mud and sand erosion condition was labeled "2", and data from the cracked turbine runner condition was labeled "3". The dataset is shown in Table 1, and the specific curves are shown below. Figure 2 As shown.

[0038] Table 1 In actual operation, the acoustic signals of hydroelectric generating units often contain a large amount of background noise, meaning that the sample feature points in Table 1 contain a significant amount of redundant and interfering features. Therefore, before inputting the dataset into the classification model, the Maximum Information Coefficient (MIC) is introduced to select features from the acoustic signals under each operating condition. This approach allows for the selection of representative points from a large number of sample features, improving the dataset quality to some extent while also saving considerable time.

[0039] The MIC feature selection process for acoustic signals under normal operating conditions is as follows: First, calculate the MIC of each of the 8000 features in the original dataset. Sort the features according to the MIC from largest to smallest, and select the top 100 features. Some of the selected features are shown below. Figure 3 As shown in the image. After processing, the dimensionality of the dataset was reduced to 600*100.

[0040] A CNN-BiLSTM model was built based on the collected acoustic signal dataset. This model consists of an input layer, convolutional layers, pooling layers, BiLSTM layers, fully connected layers, and a Softmax classification layer. The model structure parameters are shown in Table 4. The 1D convolutional layers use relatively small kernels to capture local features, while employing more filters to extract more information. The pooling layers use max pooling, which reduces the feature map size and computational cost while preserving basic feature information. The CNN-BiLSTM model structure parameters are shown in Table 2.

[0041] Table 2 Note: In conv(n) and maxpool(n), n refers to the optimal number of convolutional layers. Accuracy, Precision, Recall, and F1 Score are introduced as evaluation metrics to quantitatively analyze model performance. The mathematical principles are as follows: In binary classification problems, samples are typically defined as "Positives" and "Negatives". In the formula above, True Positives (TP) represents the number of samples that the model correctly identifies as positive; True Negatives (TN) represents the number of samples that the model correctly identifies as negative; False Positives (FP) represents the number of samples that the model incorrectly identifies as positive; and False Negatives (FN) represents the number of samples that the model incorrectly identifies as negative.

[0042] 1. Diagnosis of sediment erosion faults The acoustic signals from the feature-selected normal operating conditions and the sediment erosion conditions were integrated to obtain a dataset of size 400*100, which was divided into training and test sets in an 8:2 ratio, as shown in Table 3. The training set was used as the input dataset to the ISBOA-CNN-BiLSTM model. The initial population size was set to 5, the maximum number of iterations was set to 10, and the fitness function was set to... The optimization curve of ISBOA is as follows: Figure 4 As shown.

[0043] Table 3 To further illustrate the superior performance of the improved Heron-Eagle optimization algorithm, the Gray Wolf algorithm and the original Heron-Eagle algorithm were set as control groups. The optimization curves of the three algorithms are shown below. Figure 5 As shown. Comparing the three curves, it can be found that, with the same input dataset, ISBOA's optimization curve converges the fastest and has the highest convergence accuracy. Due to the introduction of Circle chaotic mapping to initialize the population position, ISBOA's initial population position distribution is more reasonable, resulting in the lowest fitness value and faster convergence in the early stages of iteration. Furthermore, the introduction of cosine similarity to dynamically update the population position makes the ISBOA optimization curve smoother and avoids getting trapped in local optima. The ISBOA optimization result is: number of convolutional kernels ( , ), Number of hidden units in BiLSTM layer ( This data is then fed back into the CNN-BiLSTM model. The CNN-BiLSTM is then trained on the training data, and the model's training curve is shown below. Figure 6 As shown.

[0044] To verify the performance improvement effect of the ISBOA optimization algorithm on the model, the original CNN-BiLSTM model and the CNN-BiLSTM models optimized by the GWO and SBOA algorithms were used as control groups. Comparing the training accuracy curves of the four models reveals that the ISBOA-CNN-BiLSTM model has the fastest convergence speed and the smoothest overall curve without severe oscillations. This indicates that the ISBOA optimization result, model design, and data conditions achieve a good balance. The test set data in Table 3 was then input into the trained model for recognition to further verify the model's generalization ability.

[0045] Figure 7 To quantify model performance and analyze its classification results on the training set data, Accuracy, Precision, Recall, and F1 score were calculated for the four models. The results are shown below. Figure 7 As shown in the figure, the ISBOA-CNN-BiLSTM model achieved the highest scores across all evaluation metrics, at 96.25%, 100%, 93.02%, and 96.38%, respectively. Accuracy was improved by 8.75%, 6.25%, and 3.75% compared to the other three models, while precision reached 100%. The scatter plot shows that the proposed model correctly identified all data under normal operating conditions, meaning that ISBOA-CNN-BiLSTM did not misidentify any data points in the diagnosis of sediment erosion faults. The model training curves also demonstrate that, compared to the other three models, the proposed method exhibits higher recognition accuracy, better robustness, and better generalization ability.

[0046] 2. Blade Crack Fault Diagnosis To further validate the model's performance, it was used for binary classification diagnosis of crack faults. Data from normal and turbine blade crack conditions obtained through MIC feature selection were integrated to obtain the dataset shown in Table 4. With an initial population size of 5 and a maximum iteration count of 10, the ISBOA optimization curve for the normal-crack dataset is shown below. Figure 8 As shown.

[0047] Table 4 Observing the optimization curves, it can be found that the ISBOA optimization algorithm converges the fastest, after 4 iterations, while GWO and SBOA converge after the 5th and 6th iterations, respectively. Furthermore, ISBOA has the lowest initial fitness function value. This is because this paper introduces chaotic mapping to optimize the SBOA algorithm. Chaotic mapping allows the initial population to be scientifically distributed in the optimization space, enabling the rapid search for the approximate region of optimal parameters globally. In contrast, GWO, SBOA, and most current optimization algorithms use random initialization of population positions. While this method can cover different regions of the search space, it also has significant uncertainty. Compared to the other two algorithms, ISBOA avoids getting trapped in local optima by dynamically updating positions using cosine similarity during the optimization process. The optimization result of ISBOA is... , , The optimization results are fed back into the model for training, and then the trained model is used to identify data on the test set. The classification scatter plot is shown below. Figure 9 As shown.

[0048] Figure 9 A scatter plot showing the classification results of the trained model for blade crack faults. Figure 10 Evaluation metrics for four different models are presented. Comparing the four models reveals that the ISBOA-CNN-BiLSTM model achieves the best results across all metrics, with Accuracy = 91.25%, representing improvements of 16.25%, 7.45%, and 4.95% respectively compared to the other four models. The proposed model has Precision = 95% and Recall = 88.37%. Using scatter plots, the model correctly identifies only two samples of Condition 1 (normal operating condition) as Condition 3 (crack fault), and five samples as Condition 1, both the lowest compared to the control group. This indicates that the ISBOA-optimized CNN-BiLSTM model significantly reduces the likelihood of false positives and false negatives in fault identification.

[0049] 3. Multi-condition fault diagnosis Sections 1 and 2 utilize ISBOA-CNN-BiLSTM for binary classification of the two fault types. To verify the model's classification ability on multi-condition datasets, this section investigates three types of signals: normal, sediment erosion, and crack faults. The three datasets are integrated to obtain the input dataset shown in Table 5. Due to the increased dataset complexity, the maximum number of iterations for the optimization algorithm is set to 15, while the initial population size remains unchanged. The optimization curve is shown below. Figure 11 As shown.

[0050] Table 5 To visually demonstrate that the improved method contributes to the performance enhancement of the optimization algorithm, GWO and SBOA algorithms were added as control groups. Observing the optimization curves of the three algorithms, it can be found that ISBOA converges after the 6th iteration and does not get trapped in local optima during the optimization process. This is because a cosine similarity dynamic optimization escape strategy is added to the heron escape decision part, and a global search is performed in the later stages of iteration to avoid local optima interfering with the optimization process. At the same time, due to the introduction of chaotic mapping to optimize the initial population position, the initial fitness value of ISBOA is lower than that of traditional optimization algorithms, which helps to achieve rapid convergence in the early stage. The optimization results of the three algorithms are shown in Table 6. They are then fed back into the CNN-BilSTM model for recognition, and the confusion matrix of the classification results is shown in Table 6. Figure 12 As shown.

[0051] Table 6 Figure 13 The evaluation metrics for four models in recognizing multi-condition datasets are shown. The ISBOA-CNN-BiLSTM model achieved the highest scores in all metrics, with Accuracy=90%, Precision=90.91%, Recall=90%, and F1 Score=90.5%. Accuracy was improved by 21.67%, 10.83%, and 6.7% compared to the other three models, respectively. The bar chart clearly shows that the CNN-BiLSTM model saw significant improvements in all metrics after the optimization algorithm was added, highlighting the importance of this algorithm. The performance difference between the optimized GWO and SBOA models was relatively small, but ISBOA showed a significantly larger improvement, indicating that the accuracy of hyperparameters has a significant impact on model performance. Figure 12 It can be observed that the confusion in the classification results of several models mainly occurs between normal operating conditions (condition 1) and crack faults (condition 3). However, it is worth mentioning that the ISBOA-CNN-BiLSTM model has a recall of 85% for the crack condition, which is 25%, 15%, and 5% higher than the other three models, respectively, indicating that this model has a relatively good recognition effect for weak fault signals. Meanwhile, the models show the best recognition effect for the sediment erosion condition (condition 2), with precision rates of 88.9%, 91.9%, and 100%, respectively, indicating that the ISBOA-CNN-BiLSTM model can perfectly identify faults with obvious features. In summary, compared to the binary classification experiment, although the indicators of each model decreased in the three-class classification experiment, the ISBOA-CNN-BiLSTM model consistently performed the best, maintaining a recognition accuracy above 90%. This demonstrates that this model has a strong ability to recognize different types of signals and exhibits good robustness and generalization ability.

[0052] Although the invention has been described herein with reference to several illustrative embodiments, it should be understood that many other modifications and implementations can be devised by those skilled in the art, which will fall within the scope and spirit of the principles disclosed herein. More specifically, various modifications and improvements can be made to the components or layout of the subject matter arrangement within the scope of the disclosure, drawings, and claims. Besides modifications and improvements to the components or layout, other uses will be apparent to those skilled in the art.

Claims

1. A method for diagnosing turbine faults based on acoustic signals, characterized in that: Includes the following steps: S1. Acoustic signals of the turbine under normal operating conditions, silt erosion conditions, and runner blade crack faults; S2. Introducing the maximum information coefficient to perform feature selection on the acoustic signal under each working condition, from... Representative points are selected from the acoustic signals collected in S1 as a new dataset; S3. A CNN-BiLSTM model was built using a convolutional neural network and a bidirectional long short-term memory network (BiLSTM), and the hyperparameters of the model were optimized using the improved Heron Eagle Optimization Algorithm. S4. Divide the dataset in step S2 into a training set and a test set. Use the training set as the input dataset to train the CNN-BiLSTM model, and then input the test set into the trained CNN-BiLSTM model to obtain the fault diagnosis results.

2. The method for diagnosing turbine faults based on acoustic signals according to claim 1, characterized in that: The specific steps for the Maximum Information Coefficient (MIC) are as follows: S1-1. Change the variables and Divide into grids respectively, assuming Classified as Each interval Classified as If there are several intervals, then a set of intervals will be formed. The grid is then used to calculate the mutual information of each grid cell. : Among them, variables Set as the original signal, variable The label column corresponding to the signal. for and The joint probability distribution, and yes and The marginal probability distribution; S1-2. Through different Grid combination, finding mutual information The largest mesh division, calculate each and Maximum mutual information, and normalize it: in, It is standardized mutual information. It is the standardization factor, and MIC stands for Standardized Mutual Information. The maximum value, MIC means and The maximum correlation under optimal grid partitioning, and ; S1-3. Calculate the MIC of the acoustic signal and sort the features in descending order of MIC. Select the top 1%-2.5% of features as the new dataset.

3. The method for diagnosing turbine faults based on acoustic signals according to claim 1, characterized in that: The specific steps of the improved egret optimization algorithm in step S3 are as follows: S311. Initialize population positions using Circle chaotic mapping: It is the remainder function; S312. Obtain the initial position of each individual in the heron-eagle population: in, The total number of individuals in the group. The maximum dimension of the problem being sought. and Let represent the upper and lower bounds of the i-th individual in the j-th dimension. , To represent the heron and eagle group, Let the position of the i-th heron in the population be in the j-th dimension, which is a candidate solution to the optimization problem; and calculate the position of the i-th heron. fitness value : The optimized initial position expression is: In the formula: This represents the initial optimization position of the i-th individual in the j-th dimension; Indicates the upper and lower bounds of the search space; This represents the chaotic sequence generated by iterating the Circle chaotic map t times. This represents the chaotic sequence generated by iterating through the Circle chaotic map t+1 times; K represents the parameters of the Circle chaos map; Exploration Phase: Conduct a global search by mimicking the hunting strategies of herons and eagles. in, This represents the current iteration number. The maximum number of iterations is used; the hunting process is divided into three time periods. To find prey during the hunting season, To deplete the prey's stamina for a period of time, The time period for attacking prey; and This represents a random candidate solution in the first stage of iteration. For interval The dimensions of the internally randomly generated dimension are array, This represents the new position of the i-th heron in the j-th dimension. It is the current optimal solution. , Represents randomly generated from the standard normal distribution Array; , Here is the Levy flight distribution function: in, Let i be the objective function for the i-th heron during the exploration phase. The fitness value of the objective function during the exploration phase; Development phase: Localized search is conducted by mimicking the behavior of herons and eagles when evading predators. in, Let i be the objective function for the i-th heron during the development phase. The heron eagle uses environmental camouflage strategies. Choose to escape or fly away; A dimension generated from a normal distribution. A random array, For the current iteration, a random candidate solution. It can be 1 or 2; During the development phase, cosine similarity was introduced to update the location of the fleeing heron eagle. in, , , This represents the current iteration number. This represents the maximum number of iterations; where A and B are calculated as follows: by calculating vectors... and The cosine similarity is measured by the angle between the two sides. Construct vectors within the heron-eagle population. and as follows: For the individual's current location, It is the current optimal solution. This represents the position of a random individual in the population. Let be the direction vector from the current individual's position to the local optimum. This is the direction vector from the current individual's position to the random candidate solution.

4. The method for diagnosing turbine faults based on acoustic signals according to claim 1, characterized in that: The CNN-BiLSTM model consists of an input layer, a convolutional layer, a pooling layer, a BiLSTM layer, a fully connected layer, and a Softmax classification layer.

Images