A Method for Turbine Acoustic Fault Diagnosis Based on CRK Enhanced CNN-GRU and Dual-Channel Self-Attention

Abstract

This invention relates to the field of hydropower technology, specifically to a method for diagnosing acoustic faults in hydro turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention. The method includes: S1 deploying noise sensors near the turbine runner to collect acoustic signals that avoid high-frequency interference; S2 segmenting the signal into 1024 data points per segment, analyzing the signal using Fast Fourier Transform (FFT) to extract time-frequency features such as peak amplitude and energy proportion in the 400-600Hz frequency band; S3 constructing CRK based on a Tent chaotic mapping-improved Runge-Kutta optimization algorithm to enhance global search capability and convergence speed; S4 constructing a CNN-GRU-dual-channel self-attention model to sequentially achieve local time-frequency feature extraction, temporal dynamic dependency modeling, and multi-dimensional feature enhancement; S5 optimizing model hyperparameters using CRK; and S6 training the model by dividing it into training and test sets, employing transfer learning when samples are scarce. This invention exhibits strong anti-interference capabilities, efficient hyperparameter optimization, and can accurately diagnose normal, early, and extended crack states of the turbine runner. It is suitable for scenarios with scarce samples and has high engineering practicality.

Description

Technical Field

[0001] This invention relates to the field of hydropower technology, specifically to a method for diagnosing acoustic faults in hydro turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention. Background Technology

[0002] As the core power equipment of a hydroelectric power generation system, the operational stability of the turbine directly determines the power generation efficiency and safety reliability of the hydroelectric project. The runner, as a key component for energy conversion, is subjected to long-term water flow impact, pressure fluctuations, and fatigue loads, making it prone to crack-like failures. These failures, progressing from early micro-cracks to significant damage, can lead to increased turbine vibration and abnormal noise. In severe cases, they can cause runner fracture and unit shutdown, resulting in substantial economic losses and safety risks. Therefore, efficient and accurate acoustic fault diagnosis of turbine runners, especially the identification of early cracks, has become a critical technical requirement for ensuring the safe operation of hydroelectric turbines.

[0003] Currently, acoustic signal-based fault diagnosis methods for hydro turbines have become a research hotspot in this field due to their advantages such as being non-contact, convenient to monitor, and non-destructive to equipment. However, existing technologies still have the following significant shortcomings in practical engineering applications:

[0004] I. Insufficient accuracy in signal acquisition and preprocessing makes effective feature extraction difficult.

[0005] In existing acoustic fault diagnosis methods, the deployment locations of noise sensors lack targeted design and often fail to fully consider high-frequency noise interference generated by auxiliary equipment such as pressure pumps and submersible pumps in the turbine operating environment. This results in a large number of useless interference components mixed in the acquired acoustic signals, masking the effective fault signals and leading to low signal purity. Simultaneously, the segmentation methods and feature extraction strategies in the signal preprocessing stage lack adaptability to the acoustic characteristics of turbine faults. They often employ generalized spectrum analysis and feature selection methods, failing to focus on the key time-frequency domain information corresponding to runner crack faults (especially early-stage cracks). This easily leads to the extraction of redundant features or the omission of core fault features, increasing the computational burden of subsequent models and directly affecting the quality of the basic data for fault diagnosis, thus limiting diagnostic accuracy.

[0006] II. Model hyperparameter optimization is inefficient and performance is difficult to guarantee.

[0007] Hyperparameters of deep learning models (such as input feature dimension, number of neurons in fully connected layers, and number of hidden units in recurrent layers) have a crucial impact on fault diagnosis performance. Existing methods mostly employ manual trial-and-error or traditional optimization algorithms (such as particle swarm optimization and basic Runge-Kutta optimization) to optimize hyperparameters. Manual trial-and-error relies on experience, is inefficient, and struggles to find the globally optimal parameter combination. Traditional optimization algorithms generally suffer from weak global search capabilities, are prone to getting trapped in local optima, and have slow convergence speeds. This results in poor adaptability between hyperparameters and the diagnostic model, failing to fully leverage the model's feature capture and classification capabilities, further limiting diagnostic accuracy and convergence efficiency.

[0008] III. Limitations in the structural design of the diagnostic model and insufficient feature capture capability.

[0009] Existing acoustic fault diagnosis models for hydro turbines mostly employ a single deep learning module (such as only Convolutional Neural Networks (CNN) or only Gated Recurrent Units (GRU)) or a simple combination of modules, which cannot fully adapt to the complex characteristics of acoustic signals. Although CNN modules can extract local time-frequency features, they are difficult to capture the temporal dynamic dependencies in the fault development process. Although GRU modules can model temporal information, they are not sensitive enough to local key frequency anomalies. Some models that use a single attention mechanism cannot strengthen the fault feature weights from multiple dimensions, especially the ability to capture the weak and mixed acoustic features generated by early cracks. This results in low accuracy of early fault diagnosis and poor generalization ability under high interference conditions, making it difficult to adapt to the complex and ever-changing operating environment of hydro turbines.

[0010] IV. Limited sample availability leads to poor adaptability to various scenarios and restricts practical engineering applications.

[0011] In actual operation of hydro turbines, it is often difficult to obtain a large number of early crack failure samples and failure samples under new models / new operating conditions. Existing diagnostic methods mostly rely on sufficient labeled samples for model training, lacking effective strategies for efficient sample utilization. When faced with scenarios with scarce samples, models are prone to overfitting and cannot be reused under different operating conditions or different models of hydro turbines. It is necessary to collect a large number of samples again to train a new model, which not only increases time and economic costs but also narrows the engineering applicability of existing methods, making it difficult to meet actual diagnostic needs.

[0012] In summary, existing acoustic fault diagnosis methods for hydro turbines have shortcomings in terms of signal quality, parameter optimization, model performance, and scenario adaptability, making it difficult to achieve efficient and accurate diagnosis of turbine runner faults (especially early cracks). These shortcomings also fail to fully meet the engineering requirements for equipment safety monitoring in hydropower systems. Therefore, there is an urgent need to propose an improved fault diagnosis method that can solve the above problems. Summary of the Invention

[0013] The purpose of this invention is to provide a method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention, in order to solve the problems of large signal interference, low hyperparameter optimization efficiency, insufficient feature capture, and poor sample matching in the existing methods mentioned in the background.

[0014] To achieve the above objectives, the present invention provides the following technical solution:

[0015] A method for diagnosing acoustic faults in hydro-turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention includes the following steps:

[0016] S1. Acoustic fault signal acquisition: Acoustic signals during turbine operation are acquired using a noise sensor. The noise sensor is deployed near the turbine runner to avoid high-frequency noise interference generated by the operation of the pressure pump and submersible pump.

[0017] S2. Signal preprocessing: The acoustic signal acquired in step S1 is segmented, with each segment containing 1024 data points. The segmented signal is then subjected to spectral analysis and time-frequency domain features are extracted.

[0018] S3. Construction of the Chaotic Runge-Kutta Optimization Algorithm CRK: Based on the Tent chaotic mapping, the Runge-Kutta Optimization Algorithm RUN is improved to realize the population position update in the population initialization, search and development stages, so as to improve the global search capability and convergence speed of the algorithm.

[0019] S4. Construction of CNN-GRU-Dual-Channel Self-Attention Model: Construct a deep learning model consisting of a CNN module, a GRU module, a dual-channel self-attention module, and a classification module, which sequentially realizes local time-frequency feature extraction of acoustic signals, time-series dynamic dependency modeling, multi-dimensional feature enhancement, and fault category output.

[0020] S5. Model hyperparameter optimization: Using the CRK algorithm constructed in step S3, hyperparameter optimization is performed on the input feature dimension feature_num, the number of neurons in the fully connected layer conn_neuron_num, and the number of GRU hidden units GRU_hidden_num of the model in step S4.

[0021] S6. Model Training and Fault Diagnosis: The preprocessed signal from step S2 is divided into a training set and a test set according to the proportion. The training set is used to train the model optimized in step S5, and the test set is used to verify the model performance. The acoustic signal of the turbine to be diagnosed is input into the trained model, and the fault diagnosis result is output.

[0022] Preferably, the population initialization using the improved Runge-Kutta optimization algorithm (RUN) based on the Tent chaotic mapping in step S3 specifically includes:

[0023] S31. Initialize the population for the RUN algorithm using a Tent chaotic sequence, wherein the Tent chaotic sequence satisfies:

[0024]

[0025] in, The chaotic sequence value at time k is... These are the parameters for controlling chaos.

[0026] S32. Map the initial population to the solution space of the optimization problem, satisfying:

[0027]

[0028] in, For the first The first individual Solution, , The first The lower and upper boundaries of a dimensional variable. = , The values ​​of the Tent chaotic sequence after 30-50 burn-in steps are shown. D represents the dimension of the variable.

[0029] Preferably, the population location update in the search and development phase described in step S3 specifically includes:

[0030] Set a random number rand. When rand < 0.5, enter the exploration phase and update the population position according to the following formula; when rand > 0.5, enter the development phase and update the population position according to the following formula:

[0031]

[0032]

[0033] in, , , ~ The slope coefficient of Rungekuta; , where i is the current iteration number, Maxi is the maximum iteration number, r takes the value 1 or -1, and g∈[0,2]. , These are normally distributed random numbers. , ( Represents a random number within the range (0,1); It is the optimal solution; It is the optimal solution in each iteration. Individuals in the population are randomly selected; , It is another random population individual.

[0034] Preferably, the value of α is 0.5; the value of a is 2; and the value of b is 5; the hyperparameters optimized by the CRK algorithm satisfy: feature_num=40, conn_neuron_num=36, GRU_hidden_num=36.

[0035] Preferably, the CNN module in step S4 extracts local time-frequency features of the acoustic signal, specifically including: the CNN module consists of 3 convolutional layers and 3 max pooling layers stacked alternately, and includes a batch normalization layer; the number of convolutional kernels in the convolutional layers are 32, 64, and 128 respectively, the kernel size is 3×1, and the activation function is ReLU; the pooling layer uses a 2×1 pooling kernel with a stride of 2; the batch normalization layer is used to accelerate model convergence and suppress overfitting.

[0036] Preferably, the GRU module in step S4 implements temporal dynamic dependency modeling by: the GRU module receiving the feature sequence output by the CNN module, controlling the state update through update gates and reset gates, satisfying:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] in, For the Sigmoid function, For element-wise product, Input features at time t, The hidden state at time t-1 , , , , , , , are learnable parameters; and the GRU module includes a dropout layer with a dropout probability of 0.2 to prevent model overfitting.

[0042] Preferably, the dual-channel self-attention module in step S4 to achieve multi-dimensional feature enhancement specifically includes: mapping the hidden state sequence output by the GRU module into two sets of query vectors respectively. , Key vector , Sum value vector , Attention scores are calculated and combined in the following way:

[0043] ;

[0044] ;

[0045] ;

[0046] in, , They are respectively , Dimensions and ; represents the learning weight parameters, and the attention module output is · ;and , , Generated from the GRU hidden states through a linear transformation, the transformation matrix has a dimension of . , Hides the state dimension for GRU. , , The hidden states of the GRU are generated by another set of independent linear transformations, with the transformation matrix having dimension O(n). × , .

[0047] Preferably, in step S1, the noise sensor is an industrial-grade small real-time noise spectrum analyzer with the following parameters: sampling rate 48kHz, standard measurement range 25~130dBA, dynamic measurement range ≥110dBA, frequency measurement range 10~20000Hz, communication interface USBAudio+USBHID, and size φ25×115mm.

[0048] Preferably, in step S2, the spectrum analysis uses Fast Fourier Transform (FFT); the extracted time-frequency domain features include the peak amplitude of the signal, the energy proportion of the 400-600Hz frequency band, and the noise floor value of the 0-1500Hz frequency band.

[0049] Preferably, in step S6, the ratio of the number of samples in the training set to the number of samples in the test set is 7:3; the model training uses the Adam optimizer with a learning rate of 1e-4, a batch size of 32, and a cross-entropy loss function, and iterative training is performed until the loss function converges or the maximum number of iterations of 200 is reached; the fault diagnosis results include three states of the turbine runner: normal state, early crack state with 3 cracks, and crack propagation state with 8 cracks; and when turbine fault samples are scarce, transfer learning is used to transfer the trained model to the new operating conditions, specifically by freezing some of the underlying parameters of the CNN module and the GRU module, and only fine-tuning the parameters of the dual-channel self-attention module and the classification module.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] The proposed method for acoustic fault diagnosis of hydro-turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention is discussed from four dimensions: signal quality improvement, algorithm optimization efficiency, model diagnostic performance, and scene adaptability. The details are as follows:

[0052] I. Precise signal acquisition and preprocessing lay the foundation for high-quality diagnostic data.

[0053] 1. Improve the purity of the original signal

[0054] This invention deploys noise sensors near the turbine runner and employs an industrial-grade miniature real-time noise spectrum analyzer with a sampling rate of 48kHz, a dynamic range ≥110dBA, and a frequency range of 10–20000Hz. This analyzer is adapted to the acoustic signal characteristics of the turbine and can accurately acquire fault-related signals while avoiding high-frequency noise interference from equipment such as pressure pumps and submersible pumps. Under high-interference conditions, the diagnostic accuracy of sensors deployed near the runner remains at 96.5% to 97.8%, while the accuracy of the control group deployed near the submersible pump is only 79.5% to 82.3%, fully demonstrating its anti-interference advantage.

[0055] 2. Focus on extracting key fault information

[0056] In the preprocessing stage of this invention, the acoustic signal is segmented into 1024 data points. Spectral analysis is performed using Fast Fourier Transform to specifically extract peak amplitude, energy proportion in the 400–600 Hz frequency band, and noise floor value in the 0–1500 Hz frequency band. These three features are directly related to the acoustic anomalies of turbine runner cracks, filtering out useless redundant features, reducing model computation, and improving the identification accuracy of fault features.

[0057] II. The CRK optimization algorithm is efficient and reliable, solving the problem of model hyperparameter selection.

[0058] 1. Enhance global search capabilities and convergence speed

[0059] This invention improves the Runge-Kutta optimization algorithm based on the Tent chaotic mapping to form the CRK algorithm. It initializes the population with a chaotic sequence, with the chaos control parameter set to 0.5, and performs 30 to 50 burn-in steps to avoid the randomness defect of the traditional RUN algorithm in initializing the population. At the same time, it controls the entry into the exploration or development phase by random numbers, dynamically updates the population position, balances the algorithm's global search and local development capabilities, effectively avoids getting trapped in local optima, and improves convergence accuracy and speed.

[0060] 2. Implement the CRK algorithm for precise hyperparameter optimization

[0061] This invention specifically optimizes the core hyperparameters of the model, ultimately determining the input feature dimension to be 40, the number of neurons in the fully connected layer to be 36, and the number of GRU hidden units to be 36, thus solving the problems of low efficiency and poor parameter matching in traditional manual parameter tuning. The optimized model achieves a 98.5% accuracy rate in diagnosing early cracks, compared to only 89.8% for the unoptimized model. Furthermore, the optimized model converges in only 128 to 135 rounds, demonstrating significantly higher convergence efficiency.

[0062] III. Collaboration of CNN-GRU-Dual-Channel Self-Attention Model to Comprehensively Capture Fault Features

[0063] 1. Multi-module division of labor and collaboration covers multiple types of features

[0064] In this invention, the CNN module consists of three convolutional layers and three max-pooling layers stacked alternately, along with a batch normalization layer. The number of convolutional kernels is 32, 64, and 128 respectively, with a kernel size of 3×1. The activation function is ReLU, and the pooling layer uses a 2×1 pooling kernel with a stride of 2, which can effectively extract the local time-frequency features of the acoustic signal. The batch normalization layer can also accelerate convergence and suppress overfitting. The GRU module contains a dropout layer with a probability of 0.2. It controls the state update through update gates and reset gates, modeling the temporal dynamic dependence of the signal. The dropout layer further prevents overfitting. The dual-channel self-attention module maps the hidden state output by the GRU to two sets of independent query vectors, key vectors, and value vectors. It integrates dual attention scores to enhance the multi-dimensional feature weights, highlighting key fault features and solving the problem that a single attention module cannot fully capture complex features.

[0065] 2. Excellent diagnostic accuracy and generalization ability

[0066] The multi-module collaboration of this invention enables the model to simultaneously process local frequency anomalies and temporal variation patterns of acoustic signals. The diagnostic accuracy for the normal state of the turbine, the early state of 3 cracks, and the propagation state of 8 cracks all reach over 98.5%, and the accuracy remains above 96% even under high interference conditions, demonstrating strong generalization ability.

[0067] IV. Adapting to scenarios with scarce samples and improving engineering practicality

[0068] To address the engineering challenge of scarce fault samples, especially early crack samples, for hydroelectric turbines, a transfer learning strategy is proposed. This strategy freezes the first two convolutional and pooling layers of the CNN layer and the first half of the hidden unit parameters in the GRU module of the trained model, allowing for fine-tuning only the parameters of the dual-channel self-attention module and the classification module. This enables adaptation to new operating conditions, such as new models of hydroelectric turbines. With only 140 samples per state in the target domain, the transfer learning diagnostic accuracy reaches 91.2% to 93.5%, while the accuracy without transfer learning is only 72.8% to 76.2%. This significantly reduces the dependence on the amount of samples required for new operating conditions and improves the applicability of the method in practical engineering.

[0069] V. Summary of Overall Technical Results

[0070] This invention achieves high accuracy, high anti-interference, high convergence efficiency, and strong scene adaptability in hydro-turbine acoustic fault diagnosis through a full-process design including anti-interference signal acquisition, targeted feature extraction, CRK hyperparameter optimization, multi-module deep learning model construction, and transfer learning adaptation. It is particularly suitable for the accurate identification of early cracks in hydro-turbine runners, solving the problem of low sensitivity of traditional methods for early fault diagnosis. This provides reliable technical support for preventive maintenance of hydro-turbine equipment and reduces economic losses caused by downtime due to faults. Attached Figure Description

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0072] Figure 1 This is a schematic diagram of the process of the present invention;

[0073] Figure 2 This is a bar chart showing the key indicators of the sample size distribution under the three states in the standard operating conditions in Embodiment 1 of the present invention.

[0074] Figure 3 This is a bar chart comparing the diagnostic accuracy of different sensor deployment locations under high interference conditions in Embodiment 2 of the present invention.

[0075] Figure 4 This is a bar chart comparing the accuracy of transfer learning and no transfer learning under the condition of scarce samples in Embodiment 3 of the present invention. Detailed Implementation

[0076] 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.

[0077] Example 1: Acoustic Fault Diagnosis of Hydropower Turbines under Standard Operating Conditions

[0078] This embodiment verifies the diagnostic performance of the method under standard operating conditions without additional interference (no external noise sources superimposed on the turbine) with sufficient samples.

[0079] 1.1 Experimental Conditions

[0080] Turbine parameters: Model HL220, rated speed 1500r / min, rated power 500kW;

[0081] Sensor configuration: An industrial-grade small real-time noise spectrum analyzer is used and deployed near the turbine runner (to avoid interference from the pressure pump). The parameters are: sampling rate 48kHz, measurement range 25~130dBA, dynamic range ≥110dBA, frequency range 10~20000Hz, and communication interface USBAudio+USBHID.

[0082] Fault types and sample size: Three types of status signals were collected, with 1000 samples collected for each type of status (each sample contains 1024 data points, meeting the segmentation requirements):

[0083] Normal condition: The turbine has no cracks and its operating parameters are stable;

[0084] Early crack condition: The impeller has 3 cracks with a length ≤5mm;

[0085] Crack propagation status: There are 8 cracks with a length of ≥10mm in the impeller.

[0086] 1.2 Experimental Procedure

[0087] Signal acquisition (S1): Acoustic signals of 3 states are continuously acquired through a noise sensor for 2 hours per state;

[0088] Signal preprocessing (S2): Divide the data into 1024 segments, use Fast Fourier Transform (FFT) to perform spectrum analysis, and extract three types of time-frequency domain features: peak amplitude, energy proportion of the 400-600Hz band, and noise floor value of the 0-1500Hz band.

[0089] CRK algorithm construction (S3): Based on the Tent chaotic mapping, the RUN algorithm is improved. The chaotic control parameter α=0.5, and the parameters a=2 and b=5. The chaotic sequence is processed by 30 steps of burn-in.

[0090] Model building (S4):

[0091] CNN module: 3 convolutional layers (number of convolutional kernels 32→64→128, size 3×1, ReLU activation) + 3 max pooling layers (2×1 pooling kernels, stride 2) + batch normalization layer;

[0092] GRU module: contains a dropout layer (probability 0.2), controls state updates by updating and resetting gates;

[0093] Dual-channel self-attention module: Maps the GRU output to two sets of Q / K / V vectors, calculates the attention score according to the formula for multi-dimensional feature enhancement in step S4, and fuses them;

[0094] Hyperparameter optimization (S5): The CRK algorithm was used to optimize three key hyperparameters, and the final values ​​were feature_num=40, conn_neuron_num=36, and GRU_hidden_num=36.

[0095] Model training and diagnosis (S6): The training set and test set are divided into a 7:3 ratio. The Adam optimizer (learning rate 1e-4), batch size 32, and cross-entropy loss function are used to train until the loss converges (maximum 200 iterations).

[0096] 1.3 Experimental Results

[0097]

[0098] Example 2: Acoustic fault diagnosis of water turbine under high interference conditions (external noise superposition)

[0099] This embodiment addresses high-interference operating conditions by adding a submersible pump near the turbine, generating high-frequency noise, to verify the impact of sensor deployment location on diagnostic performance and the method's anti-interference capability.

[0100] 2.1 Experimental Conditions

[0101] Turbine and sensor parameters: Same as in Example 1 (HL220, 1500r / min), the sensor is still an industrial-grade small real-time noise spectrum analyzer;

[0102] Interference source: A submersible pump (rated power 7.5kW, generating 10-15kHz high-frequency noise during operation) is deployed next to the turbine.

[0103] Sensor deployment comparison: Two deployment locations were set up:

[0104] Experimental group: Sensors were deployed near the rotating wheel;

[0105] Control group: Sensors were deployed near the submersible pump (without avoiding interference);

[0106] Fault type and sample size: Same as the three states in Example 1, with 800 samples collected for each state (1024 data points per sample).

[0107] 2.2 Experimental Procedure

[0108] Signal acquisition (S1): Two sets of sensors synchronously acquire acoustic signals in three states, with a duration of 1.5 hours per state;

[0109] Signal preprocessing to model training (S2-S6): Same as Example 1 (CRK parameters, model structure, and hyperparameter optimization results are consistent), with the training set / test set divided in a 7:3 ratio (560 training states, 240 testing states).

[0110] 2.3 Experimental Results

[0111]

[0112] Example 3: Acoustic Fault Diagnosis of Hydropower Turbines under Conditions of Scarce Samples (Application of Transfer Learning)

[0113] This embodiment addresses the scenario of scarce new operating condition samples, changes in turbine model, and insufficient fault samples, verifying the effectiveness of the transfer learning strategy.

[0114] 3.1 Experimental Conditions

[0115] Water turbine parameters:

[0116] Source domain: HL220 water turbine of Example 1 (sufficient samples, 700 training samples for each of the 3 states);

[0117] Target domain: New model HL240 water turbine (operating condition difference: speed 1800 r / min), sample is scarce, only 200 samples were collected for each of the 3 states;

[0118] Sensor configuration: Same as in Example 1 (deployed near the HL240 rotor);

[0119] Transfer learning strategy: Freeze the parameters of the bottom layer of the CNN module (the first two layers of convolution + pooling) and the bottom layer of the GRU module (the first half of the hidden units), and only fine-tune the parameters of the dual-channel self-attention module and the classification module.

[0120] 3.2 Experimental Procedure

[0121] Signal Acquisition (S1): Acquire the acoustic signals of the HL240 in three states, for a duration of 40 minutes per state;

[0122] Signal preprocessing (S2): Same as Example 1 (segmentation, FFT, feature extraction);

[0123] Model transfer and training (S3-S6):

[0124] Base model: Reuse the CNN-GRU dual-channel self-attention model optimized by CRK in Example 1;

[0125] Control group: No transfer learning (a brand new model was trained directly using 200 samples from the HL240).

[0126] Training set / test set: divided in a 7:3 ratio (140 training states, 60 testing states), other training parameters are the same as in Example 1.

[0127] 3.3 Experimental Results

[0128]

[0129] The proposed method for acoustic fault diagnosis of hydro-turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention is discussed from four dimensions: signal quality improvement, algorithm optimization efficiency, model diagnostic performance, and scene adaptability. The details are as follows:

[0130] I. Precise signal acquisition and preprocessing lay the foundation for high-quality diagnostic data.

[0131] 1. Improve the purity of the original signal

[0132] This invention deploys noise sensors near the turbine runner and employs an industrial-grade miniature real-time noise spectrum analyzer with a sampling rate of 48kHz, a dynamic range ≥110dBA, and a frequency range of 10–20000Hz. This analyzer is adapted to the acoustic signal characteristics of the turbine and can accurately acquire fault-related signals while avoiding high-frequency noise interference from equipment such as pressure pumps and submersible pumps. Under high-interference conditions, the diagnostic accuracy of sensors deployed near the runner remains at 96.5% to 97.8%, while the accuracy of the control group deployed near the submersible pump is only 79.5% to 82.3%, fully demonstrating its anti-interference advantage.

[0133] 2. Focus on extracting key fault information

[0134] In the preprocessing stage of this invention, the acoustic signal is segmented into 1024 data points. Spectral analysis is performed using Fast Fourier Transform to specifically extract peak amplitude, energy proportion in the 400-600Hz frequency band, and noise floor value in the 0-1500Hz frequency band. These three features are directly related to the acoustic anomalies of turbine runner cracks, filtering out useless redundant features, reducing model computation, and improving the identification accuracy of fault features.

[0135] II. The CRK optimization algorithm is efficient and reliable, solving the problem of model hyperparameter selection.

[0136] 1. Enhance global search capabilities and convergence speed

[0137] This invention improves the Runge-Kutta optimization algorithm based on the Tent chaotic mapping to form the CRK algorithm. It initializes the population with a chaotic sequence, with the chaos control parameter set to 0.5, and performs 30 to 50 burn-in steps to avoid the randomness defect of the traditional RUN algorithm in initializing the population. At the same time, it controls the entry into the exploration or development phase by random numbers, dynamically updates the population position, balances the algorithm's global search and local development capabilities, effectively avoids getting trapped in local optima, and improves convergence accuracy and speed.

[0138] 2. Implement the CRK algorithm for precise hyperparameter optimization

[0139] This invention specifically optimizes the core hyperparameters of the model, ultimately determining the input feature dimension to be 40, the number of neurons in the fully connected layer to be 36, and the number of GRU hidden units to be 36, thus solving the problems of low efficiency and poor parameter matching in traditional manual parameter tuning. The optimized model achieves a 98.5% accuracy rate in diagnosing early cracks, compared to only 89.8% for the unoptimized model. Furthermore, the optimized model converges in only 128 to 135 rounds, demonstrating significantly higher convergence efficiency.

[0140] III. Collaboration of CNN-GRU-Dual-Channel Self-Attention Model to Comprehensively Capture Fault Features

[0141] 1. Multi-module division of labor and collaboration covers multiple types of features

[0142] In this invention, the CNN module consists of three convolutional layers and three max-pooling layers stacked alternately, along with a batch normalization layer. The number of convolutional kernels is 32, 64, and 128 respectively, with a kernel size of 3×1. The activation function is ReLU, and the pooling layer uses a 2×1 pooling kernel with a stride of 2, which can effectively extract the local time-frequency features of the acoustic signal. The batch normalization layer can also accelerate convergence and suppress overfitting. The GRU module contains a dropout layer with a probability of 0.2. It controls the state update through update gates and reset gates, modeling the temporal dynamic dependence of the signal. The dropout layer further prevents overfitting. The dual-channel self-attention module maps the hidden state output by the GRU to two sets of independent query vectors, key vectors, and value vectors. It integrates dual attention scores to enhance the multi-dimensional feature weights, highlighting key fault features and solving the problem that a single attention module cannot fully capture complex features.

[0143] 2. Excellent diagnostic accuracy and generalization ability

[0144] The multi-module collaboration of this invention enables the model to simultaneously process local frequency anomalies and temporal variation patterns of acoustic signals. The diagnostic accuracy for the normal state of the turbine, the early state of 3 cracks, and the propagation state of 8 cracks all reach over 98.5%, and the accuracy remains above 96% even under high interference conditions, demonstrating strong generalization ability.

[0145] IV. Adapting to scenarios with scarce samples and improving engineering practicality

[0146] To address the engineering challenge of scarce fault samples, especially early crack samples, for hydroelectric turbines, a transfer learning strategy is proposed. This strategy freezes the first two convolutional and pooling layers of the CNN layer and the first half of the hidden unit parameters in the GRU module of the trained model, allowing for fine-tuning only the parameters of the dual-channel self-attention module and the classification module. This enables adaptation to new operating conditions, such as new models of hydroelectric turbines. With only 140 samples per state in the target domain, the transfer learning diagnostic accuracy reaches 91.2% to 93.5%, while the accuracy without transfer learning is only 72.8% to 76.2%. This significantly reduces the dependence on the amount of samples required for new operating conditions and improves the applicability of the method in practical engineering.

[0147] V. Summary of Overall Technical Results

[0148] This invention achieves high accuracy, high anti-interference, high convergence efficiency, and strong scene adaptability in hydro-turbine acoustic fault diagnosis through a full-process design including anti-interference signal acquisition, targeted feature extraction, CRK hyperparameter optimization, multi-module deep learning model construction, and transfer learning adaptation. It is particularly suitable for the accurate identification of early cracks in hydro-turbine runners, solving the problem of low sensitivity of traditional methods for early fault diagnosis. This provides reliable technical support for preventive maintenance of hydro-turbine equipment and reduces economic losses caused by downtime due to faults.

[0149] 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 diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention, characterized in that, Includes the following steps: S1. Acoustic fault signal acquisition: Acoustic signals during the operation of the turbine are acquired using a noise sensor. The noise sensor is deployed near the turbine runner to avoid high-frequency noise interference generated by the operation of the pressure pump and submersible pump. S2. Signal preprocessing: The acoustic signal acquired in step S1 is segmented, with each segment containing 1024 data points. The segmented signal is then subjected to spectral analysis and time-frequency domain features are extracted. S3. Construction of the Chaotic Runge-Kutta Optimization Algorithm CRK: Based on the Tent chaotic mapping, the Runge-Kutta Optimization Algorithm RUN is improved to realize the population position update in the population initialization, search and development stages, so as to improve the global search capability and convergence speed of the algorithm. S4. Construction of CNN-GRU-Dual-Channel Self-Attention Model: Construct a deep learning model consisting of a CNN module, a GRU module, a dual-channel self-attention module, and a classification module, which sequentially realizes local time-frequency feature extraction of acoustic signals, time-series dynamic dependency modeling, multi-dimensional feature enhancement, and fault category output. S5. Model hyperparameter optimization: Using the CRK algorithm constructed in step S3, hyperparameter optimization is performed on the input feature dimension feature_num, the number of neurons in the fully connected layer conn_neuron_num, and the number of GRU hidden units GRU_hidden_num of the model in step S4. S6. Model Training and Fault Diagnosis: The preprocessed signal from step S2 is divided into a training set and a test set according to the proportion. The training set is used to train the model optimized in step S5, and the test set is used to verify the model performance. The acoustic signal of the turbine to be diagnosed is input into the trained model, and the fault diagnosis result is output. Step S3, which describes the improved Runge-Kutta optimization algorithm (RUN) based on the Tent chaotic map for population initialization, specifically includes: S31. Initialize the population for the RUN algorithm using a Tent chaotic sequence, wherein the Tent chaotic sequence satisfies: ； in, The chaotic sequence value at time k is... These are the parameters for controlling chaos. S32. Map the initial population to the solution space of the optimization problem, satisfying: ； in, For the first The first individual Solution, , The first The lower and upper boundaries of a dimensional variable. = , The values ​​of the Tent chaotic sequence after 30-50 burn-in steps are shown. D represents the dimension of the variable; The dual-channel self-attention module described in step S4 for multi-dimensional feature enhancement specifically includes mapping the hidden state sequence output by the GRU module into two sets of query vectors. , Key vector , Sum value vector , Attention scores are calculated and combined in the following way: ； ； ； in, , They are respectively , Dimensions and ; represents the learning weight parameters, and the attention module output is · ;and , , Generated from the GRU hidden states through a linear transformation, the transformation matrix has a dimension of . , Hides the state dimension for GRU. , , The hidden states of the GRU are generated by another set of independent linear transformations, with the transformation matrix having dimension O(n). × , 2. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 1, characterized in that, The population location update during the search and development phase specifically includes: Set a random number rand. When rand < 0.5, enter the exploration phase and update the population position according to the following formula; when rand > 0.5, enter the development phase and update the population position according to the following formula: ； ； in, , , ~ The slope coefficient of Rungekuta; , where i is the current iteration number, Maxi is the maximum iteration number, r takes the value 1 or -1, and g∈[0,2]. , These are normally distributed random numbers. , , Represents a random number within the range (0,1); It is the optimal solution; It is the optimal solution in each iteration. Individuals in the population are randomly selected; , It is another random population individual.

3. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 2, is characterized in that... The The value of is 0.5; the value of a is 2, and the value of b is 5; the hyperparameters optimized by the CRK algorithm satisfy: feature_num=40, conn_neuron_num=36, GRU_hidden_num=36.

4. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 1, characterized in that, The CNN module described in step S4 extracts local time-frequency features of acoustic signals. Specifically, the CNN module consists of three convolutional layers and three max-pooling layers stacked alternately, and includes a batch normalization layer. The number of convolutional kernels in the convolutional layers are 32, 64, and 128, respectively, with a kernel size of 3×1 and an activation function of ReLU. The pooling layers use 2×1 pooling kernels with a stride of 2. The batch normalization layer is used to accelerate model convergence and suppress overfitting.

5. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 1, characterized in that, The GRU module's implementation of temporal dynamic dependency modeling in step S4 specifically includes: the GRU module receiving the feature sequence output by the CNN module, controlling state updates through update gates and reset gates, satisfying: ； ； ； ； in, For the Sigmoid function, For element-wise product, Input features at time t, The hidden state at time t-1 , , , , , , , are learnable parameters; and the GRU module includes a dropout layer with a dropout probability of 0.2 to prevent model overfitting.

6. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 1, characterized in that, In step S1, the noise sensor is an industrial-grade small real-time noise spectrum analyzer with the following parameters: sampling rate 48kHz, standard measurement range 25~130dBA, dynamic measurement range ≥110dBA, frequency measurement range 10~20000Hz, communication interface USBAudio+USBHID, and size φ25×115mm.

7. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 1, characterized in that, The spectrum analysis employs Fast Fourier Transform (FFT); the extracted time-frequency domain features include the peak amplitude of the signal, the energy percentage of the 400–600 Hz band, and the noise floor of the 0–1500 Hz band.

8. The method for diagnosing acoustic faults in water turbines based on CRK-enhanced CNN-GRU and dual-channel self-attention as described in claim 1, characterized in that, In step S6, the ratio of the number of samples in the training set to the number of samples in the test set is 7:3; the model training uses the Adam optimizer with a learning rate of 1e-4, a batch size of 32, and a cross-entropy loss function, and iterative training is performed until the loss function converges or the maximum number of iterations of 200 is reached; the fault diagnosis results include three states of the turbine runner: normal state, early crack state with 3 cracks, and crack propagation state with 8 cracks; and when turbine fault samples are scarce, transfer learning is used to transfer the trained model to the new operating conditions, specifically by freezing some of the underlying parameters of the CNN module and the GRU module, and only fine-tuning the parameters of the dual-channel self-attention module and the classification module.

Images