A method and apparatus for monitoring a hydraulic turbine

Abstract

The application discloses a kind of water turbine monitoring method and device, wherein the method comprises: using adaptive wavelet semi-soft threshold denoising method to the vibration signal data of water turbine is denoised;Characteristic data is extracted from vibration signal data using wavelet energy coefficient analysis method combined with wavelet decomposition coefficient mean square value statistical analysis method, and water turbine unit diagnosis sample database is established;Uncertainty bayesian neural network model based on deep learning is used to mine the feature relationship between water turbine data, which solves the problem that the existing technology detects water turbine by manual reinspection in offline state, with high implementation difficulty, low coverage, easy to miss detection, and unable to monitor water turbine unit in service state in real time.

Description

Technical Field

[0001] This invention relates to the field of water turbine monitoring technology, specifically to a water turbine monitoring method and device. Background Technology

[0002] Hydropower, as a crucial component of clean energy, holds immense strategic importance for economic development. According to my country's "Medium and Long-Term Development Plan for Renewable Energy," the country's hydropower turbine manufacturing industry is poised for rapid growth. As the most critical mechanical equipment in hydropower projects, any malfunction of the turbine unit during operation can lead to not only significant economic losses but also substantial safety hazards for workers. Therefore, establishing an efficient and reliable hydropower turbine health monitoring system is of paramount importance.

[0003] Big data in the mechanical field has three main characteristics: 1. Large capacity, reaching terabytes or more, resulting in an extremely large workload for data analysis; 2. Diversity, with collected data covering a vast amount of health status information radiated from different physical sources under various operating conditions, leading to extremely complex analysis content; 3. High speed, as the various components within mechanical equipment are closely interconnected, even minor faults can cause a series of chain reactions, resulting in damage to the entire equipment. Considering these characteristics of mechanical big data, an ideal mechanical big data processing method should possess the ability to quickly and accurately extract information from big data and efficiently identify the health status of equipment. However, traditional big data identification and prediction methods have two major drawbacks: ① They require extensive signal processing techniques combined with rich engineering experience to extract fault features; ② The use of shallow models makes it difficult to characterize the complex mapping relationship between signals and health status in the context of big data. This often makes the characterization of complex, massive feature data "inadequate."

[0004] Currently, the detection of faults in key components of hydro-turbine units is mainly carried out offline through manual re-inspection. This is challenging in two ways: firstly, it is difficult to implement, especially for bearings located inside the turbine unit, which require extensive disassembly; secondly, manual inspection has low coverage and is prone to omissions. More importantly, existing detection methods are mostly indirect, such as ultrasonic non-destructive testing, which cannot perform real-time online monitoring of hydro-turbine units in service. Therefore, establishing an intelligent fault diagnosis system for hydro-turbine generator units is essential. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the existing technology of conducting turbine inspections by manual re-inspection in offline state, which is difficult to implement, has low coverage, is prone to missed inspections, and cannot perform real-time online monitoring of turbine units in service state. Thus, a turbine monitoring method and device are provided.

[0006] To address the aforementioned technical problems, the present invention discloses at least one method and apparatus for monitoring water turbines.

[0007] In a first aspect, the present invention discloses an embodiment of a water turbine monitoring method, comprising:

[0008] An adaptive wavelet semi-soft thresholding denoising method is used to denoise the vibration signal data of the water turbine.

[0009] A combination of wavelet energy coefficient analysis and wavelet decomposition coefficient mean square value statistical analysis is used to extract feature data from the vibration signal data. The feature data is used to characterize the vibration of the turbine under different fault modes and the turbine unit response, and to establish a turbine unit diagnostic sample database.

[0010] We used a deep learning-based uncertainty Bayesian neural network model to mine the feature relationships between water turbine data.

[0011] Optionally, the method of using adaptive wavelet semi-soft thresholding to denoise the turbine vibration signal data includes: using wavelet transform to perform multi-scale decomposition and adaptive denoising on the noisy vibration signal data; selecting a target wavelet basis according to the characteristics of the vibration signal, performing N-level decomposition on the signal, calculating detail coefficients from level 1 to level N and the myopia coefficient of level N to estimate the noise bias and calculate the semi-soft threshold for the corresponding level, and using the calculated semi-soft threshold to discriminate the signal coefficients, setting wavelet coefficients with amplitudes lower than the semi-soft threshold to 0, and completely retaining or performing corresponding shrinkage processing on wavelet coefficients with amplitudes higher than the semi-soft threshold;

[0012] Finally, the processed wavelet coefficients are reconstructed using inverse wavelet transform to obtain high-precision vibration signal data after denoising.

[0013] Optionally, the multi-scale decomposition and adaptive denoising of the noisy vibration signal data using wavelet transform is performed as follows: (using the formula...) and The vibration signal data of the water turbine is denoised, where thr(i) is the adaptive threshold of the i-th layer wavelet decomposition, and d i Let σ be the detail coefficients of the i-th level of wavelet decomposition. 2 Let rw be the noise variance of the signal. i For the reconstructed i-th layer wavelet, a, b, and C are constants.

[0014] Optionally, the step of combining wavelet energy coefficient analysis with wavelet decomposition coefficient mean square value statistical analysis to extract feature data from the vibration signal data includes: using basic functions to perform wavelet decomposition on the denoised vibration signal data to obtain wavelet transform functions; using the obtained wavelet transform functions to calculate the wavelet energy coefficients of each frequency band; and establishing the distribution state of the wavelet energy coefficients.

[0015] Optionally, the method of combining wavelet energy coefficient analysis with wavelet decomposition coefficient mean square value statistical analysis to extract feature data from the vibration signal data includes: decomposing the vibration signal data after semi-soft threshold denoising into J+1 frequency range components through J scales; setting a threshold for the product of the mean square value of the wavelet coefficients and a weighting coefficient less than 1 in the energy-dense frequency band, denoted by T; when the wavelet decomposition coefficient is greater than T, counting the number of wavelet decomposition coefficients greater than T, denoted by N; the mean square value of the wavelet decomposition coefficients in each feature frequency band is represented by the parameter R(N), and establishing the correspondence between the number of wavelet decomposition coefficients N greater than the threshold T and R(N).

[0016] Optionally, the formula for calculating the weighting coefficients is:

[0017]

[0018] Optionally, the step of mining the feature relationships between turbine data using a deep learning-based uncertainty Bayesian neural network model includes: combining a DNN neural network model with probabilistic modeling methods to achieve the modeling of a novel big data uncertainty Bayesian neural network model; the uncertainty Bayesian neural network model includes an uncertainty Bayesian network and an attention mechanism, utilizing variational distribution q θ (ω) replaces the posterior distribution in the Bayesian neural network, and its calculation formula is as follows:

[0019]

[0020] Where, x n For output parameters, y n For the output result, q θ (ω) is a variational distribution, and θ consists of the mean and standard deviation of each independent Gaussian. Probabilistic prediction is performed by integrating the variational distribution instead of the posterior.

[0021] The neural network model is trained by iteratively using a sample database until the divergence KL reaches its minimum. The model is then tested on a test set, and the fault types predicted by the model are compared with the actual fault types until the model's prediction accuracy meets the requirements.

[0022] Secondly, embodiments of the present invention also provide a turbine monitoring device, comprising:

[0023] The denoising module is used to denoise the turbine vibration signal data using an adaptive wavelet semi-soft threshold denoising method.

[0024] The feature extraction module is used to extract feature data from the vibration signal data by combining wavelet energy coefficient analysis and wavelet decomposition coefficient mean square value statistical analysis. The feature data is used to characterize the vibration of the turbine under different fault modes and the turbine unit response, and to establish a turbine unit diagnostic sample database.

[0025] The feature relationship determination module is used to mine the feature relationships between water turbine data using a deep learning-based uncertain Bayesian neural network model.

[0026] Thirdly, the present invention also discloses a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps of the first aspect above, or any possible implementation of the first aspect, are performed.

[0027] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the first aspect or any possible implementation thereof.

[0028] The technical solutions provided by the embodiments of the present invention can have the following beneficial effects:

[0029] This paper introduces a combination of adaptive wavelet analysis and feature statistical analysis to extract features from the vibration signals of hydro-turbines. First, an adaptive wavelet semi-soft thresholding denoising method is used to denoise the original vibration signal, obtaining a high-precision input signal. Then, a combination of wavelet energy system analysis and wavelet decomposition coefficient mean square value feature statistical analysis is employed to extract feature data of vibration and turbine unit response under different fault modes, establishing a turbine unit diagnostic sample database. Finally, a deep learning-based uncertainty Bayesian neural network model is constructed to effectively mine fault features, enabling rapid and accurate prediction of hydro-turbine faults under big data conditions.

[0030] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0031] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0032] Figure 1 A flowchart of a water turbine monitoring method provided by an embodiment of the present invention is shown;

[0033] Figure 2 A flowchart of another turbine monitoring method provided by an embodiment of the present invention is shown;

[0034] Figure 3 A schematic diagram of the uncertainty Bayesian neural network model structure in an embodiment of the present invention is shown;

[0035] Figure 4 A schematic diagram of the signals acquired by the sensor in an embodiment of the present invention is shown;

[0036] Figure 5-11 This invention discloses a comparative diagram of data before and after noise reduction in a turbine monitoring method according to an embodiment of the present invention.

[0037] Figure 12 The wavelet energy coefficient distribution state after calculation in the disclosed embodiment of the present invention is shown;

[0038] Figure 13 This diagram illustrates the structure of a water turbine monitoring device provided in an embodiment of the present invention.

[0039] Figure 14 A schematic diagram of the structure of a computer device provided by an embodiment of the present invention is shown. Detailed Implementation

[0040] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0041] Current traditional hydro-turbine abnormal condition monitoring systems, due to the large capacity and diversity of hydro-turbine signals, cannot effectively handle noise signals in the raw vibration data using conventional data preprocessing methods, thus failing to remove invalid signals. Furthermore, driven by big data, when faced with massive amounts of mechanical signals exhibiting alternating operating conditions, severe fault information coupling, and unclear and variable patterns, it is often difficult to extract typical features of the mechanical health status, resulting in poor monitoring and diagnostic capabilities and generalization performance. To address these challenges, this invention employs an adaptive wavelet semi-soft thresholding denoising module to effectively remove noise from the raw signals. Simultaneously, it combines adaptive wavelet analysis and feature statistical analysis to extract features from the hydro-turbine vibration signals. Finally, a deep learning-based uncertain Bayesian neural network model is constructed to map the vibration features to the health status, thereby achieving rapid and accurate diagnosis of hydro-turbine abnormal conditions. Compared with existing technologies, this invention provides the following embodiments:

[0042] Example 1

[0043] like Figure 1 The flowchart shown is a method for monitoring a water turbine provided by an embodiment of the present invention. The method includes:

[0044] S11: The adaptive wavelet semi-soft threshold denoising method is used to denoise the turbine vibration signal data;

[0045] S12: The wavelet energy coefficient analysis method and the wavelet decomposition coefficient mean square value statistical analysis method are combined to extract feature data from the vibration signal data. The feature data is used to characterize the vibration of the turbine under different fault modes and the turbine unit response, and to establish a turbine unit diagnostic sample database.

[0046] S13: Utilize a deep learning-based uncertainty Bayesian neural network model to mine the feature relationships between water turbine data.

[0047] It is understood that the technical solution provided in this embodiment is used for intelligent diagnosis of abnormal states of hydro-generator units in the context of big data, and to realize timely reminders of abnormal states to avoid major failures. An adaptive wavelet semi-soft threshold denoising method is proposed to preprocess the big data signals collected by the turbine sensors, achieving effective noise removal. An adaptive wavelet analysis method and a feature statistical analysis method are combined to extract features from the turbine vibration signals. First, an adaptive wavelet semi-soft threshold denoising method is used to denoise the original vibration signal to obtain a high-precision input signal. Then, a combination of wavelet energy system analysis and wavelet decomposition coefficient mean square value feature statistical analysis is employed to extract characteristic data of vibration and turbine unit response under different fault modes, establishing a turbine unit diagnostic sample database. Finally, a deep learning-based uncertainty Bayesian neural network model is constructed to effectively mine fault features, enabling rapid and accurate prediction of turbine faults under big data conditions. This method and device can adapt to the working environment requirements of the turbine unit, ensuring effective health monitoring of the turbine unit during operation, effectively avoiding safety hazards caused by turbine operation faults, and reducing manual fault diagnosis and maintenance costs.

[0048] Example 2

[0049] like Figure 2 As shown, a flowchart of another turbine monitoring method provided by an embodiment of the present invention is presented, the method comprising:

[0050] S21: Adaptive wavelet semi-soft threshold denoising method is used to denoise the turbine vibration signal data;

[0051] S22: The wavelet energy coefficient analysis method and the wavelet decomposition coefficient mean square value statistical analysis method are combined to extract feature data from vibration signal data. The feature data are used to characterize the vibration of the turbine under different fault modes and the turbine unit response, and to establish a turbine unit diagnostic sample database.

[0052] S23: Utilize a deep learning-based uncertainty Bayesian neural network model to mine the feature relationships between water turbine data.

[0053] In specific practice, such as Figure 2 As shown by the dashed line, S21 can be implemented through, but is not limited to, the following process:

[0054] S211: The vibration signal data containing noise is decomposed into multiple scales and adaptively denoised using wavelet transform;

[0055] Specifically, in some optional embodiments, a target wavelet basis is selected based on the characteristics of the vibration signal, the signal is decomposed into N layers, and the detail coefficients from 1 to N and the myopia coefficient of the Nth layer are calculated to estimate the noise deviation and calculate the semi-soft threshold of the corresponding layer. The signal coefficients are then processed by using the calculated semi-soft threshold. Wavelet coefficients with amplitudes lower than the semi-soft threshold are set to 0, while wavelet coefficients with amplitudes higher than the semi-soft threshold are completely retained or subjected to corresponding shrinkage processing.

[0056] S212: The wavelet coefficients obtained after processing are reconstructed through inverse wavelet transform to obtain high-precision vibration signal data after denoising.

[0057] In practical application, S211 can be expressed as follows: (using the formula...) and The vibration signal data of the water turbine is denoised, where thr(i) is the adaptive threshold of the i-th layer wavelet decomposition, and d i Let σ be the detail coefficients of the i-th level of wavelet decomposition. 2 Let rw be the noise variance of the signal. i For the reconstructed i-th layer wavelet, a, b, and C are constants.

[0058] In practice, not shown in the diagram, S22 can be implemented through, but is not limited to, the following processes:

[0059] S22a-1: Wavelet transform function is obtained by performing wavelet decomposition on the denoised vibration signal data using basic functions;

[0060] S22a-2: Calculate the wavelet energy coefficients of each frequency band using the obtained wavelet transform function;

[0061] S22a-3: Establish the distribution state of wavelet energy coefficients.

[0062] Furthermore, in practical applications, not shown in the figure, S22 can also be implemented through, but is not limited to, the following processes:

[0063] S22b-1: Decompose the vibration signal data after semi-soft threshold denoising into J+1 frequency range components through J scales;

[0064] S22b-1: Set the threshold for the product of the mean square value of the wavelet coefficients and a weighting coefficient less than 1 in the energy-dense frequency band, denoted by T;

[0065] S22b-1: When the wavelet decomposition coefficients are greater than T, count the number of wavelet decomposition coefficients greater than T, denoted by N;

[0066] S22b-1: The mean square value of the wavelet decomposition coefficients of each characteristic frequency band is represented by the parameter R(N). The correspondence between the number N of wavelet decomposition coefficients greater than the threshold T and R(N) is established.

[0067] In practice, the formula for calculating the weighting coefficients is as follows:

[0068]

[0069]

[0070]

[0071] In the formula, c(x)k represents the detail coefficient of the kth order, c(x)ston is the weighting coefficient of the n modes, S(x) is the squared residual of the signal (the deviation of each superimposed data point from the average value of the superimposed coefficients), and S(x) is the normalized weighting coefficient.

[0072] In practice, not shown in the diagram, S23 can be achieved through, but is not limited to, the following processes:

[0073] S231: Combining DNN neural network models with probabilistic modeling methods to realize the modeling of a novel Bayesian neural network model for big data uncertainty;

[0074] Specifically, in some alternative embodiments, see Figure 3 The uncertainty Bayesian neural network model includes an uncertainty Bayesian network and an attention mechanism, utilizing the variational distribution q. θ (ω) replaces the posterior distribution in the Bayesian neural network, and its calculation formula is as follows:

[0075]

[0076] Where, x n For output parameters, y n For the output result, q θ (ω) is the variational distribution, and θ consists of the mean and standard deviation of each independent Gaussian. Probability prediction is then performed by integrating the variational distribution instead of the posterior. An attention mechanism module is embedded in the uncertain Bayesian neural network to enhance important features and improve the accuracy of the model.

[0077] S232: Use a sample database to iterate repeatedly until the divergence of the neural network model reaches its minimum, which means the training is complete;

[0078] S233: The test set tests the neural network model by comparing the fault types predicted by the neural network model with the actual fault types until the model's prediction accuracy meets the requirements.

[0079] To facilitate understanding, the principle and implementation details of the above-mentioned turbine monitoring method are described in detail below:

[0080] (1) Noise removal of turbine vibration signal data

[0081] An adaptive wavelet semi-soft thresholding denoising method is used to denoise the turbine vibration signal data. This implementation defines it as adaptive wavelet semi-soft thresholding denoising. See [link to relevant documentation]. Figure 4 To address the noise problem in the massive amounts of raw turbine data collected by sensors, an adaptive wavelet semi-soft thresholding denoising method is employed to denoise the turbine vibration signal. Wavelet transform is used to perform multi-scale decomposition and adaptive denoising on the noisy vibration signal, followed by inverse wavelet transform to obtain the denoised high-precision input signal.

[0082] Specifically, thresholding involves two methods: hard thresholding and soft thresholding. The formulas for calculating the hard thresholding function ψ(x) and the soft thresholding function η(x) are (1) and (2), respectively:

[0083]

[0084] η(x)=sgn(x)max(|x|-T, 0) (2)

[0085] In the above formula, x is the wavelet coefficient and T is the threshold.

[0086] The hard thresholding function compares wavelet coefficients x with a threshold T, setting wavelet coefficients smaller than the threshold to 0. The soft thresholding function, on the other hand, processes wavelet coefficients larger than the threshold based on the hard threshold, subtracting the threshold to obtain new wavelet coefficients. Threshold selection is a crucial step in wavelet thresholding denoising. However, as can be seen from equation (1), the hard thresholding function is discontinuous at the threshold point, leading to problems such as ringing and pseudo-Gibbs effects. Furthermore, the original coefficients processed by the soft thresholding function always have a constant deviation from the decomposed wavelet coefficients, thus affecting the accuracy of signal reconstruction. Therefore, this paper proposes another adaptive wavelet semi-soft thresholding denoising method, with the calculation formula as follows:

[0087]

[0088]

[0089]

[0090] In the formula, thr(i) is the adaptive threshold of the i-th level wavelet decomposition, and d i Let σ be the detail coefficients of the i-th level of wavelet decomposition. 2 Let rw be the noise variance of the signal.i For the reconstructed i-th layer wavelet, a, b, and C are constants.

[0091] Figures 5-11 The data before and after noise reduction is shown in the comparison data.

[0092] Effective and high-quality turbine datasets are a crucial prerequisite for achieving turbine health monitoring. Because turbine units operate in complex environments, their sensors are susceptible to interference from complex and ever-changing external uncertainties during signal acquisition, making it difficult to obtain high-precision input signals for intelligent diagnostic models. These factors significantly impact efficient and high-speed abnormal condition monitoring of turbine units. Therefore, it is essential to address the challenge of low-quality raw signals acquired during the complex operating conditions of turbine units. The adaptive wavelet semi-soft thresholding denoising method proposed in this invention can preprocess the large datasets acquired by turbine sensors, effectively removing noise and solving the problem of low signal reconstruction accuracy in traditional wavelet denoising methods.

[0093] (2) Feature extraction

[0094] A combination of wavelet energy system analysis and wavelet decomposition coefficient mean square value characteristic statistical analysis was employed to extract characteristic data of vibration and turbine unit response under different fault modes, establishing a turbine unit diagnostic sample database. The wavelet energy coefficient analysis method utilizes fundamental functions such as sine or cosine waves to perform wavelet decomposition on the denoised vibration signal to obtain the wavelet transform function. The obtained wavelet transform function is then used to calculate the wavelet energy coefficients for each frequency band, and the distribution of the wavelet energy coefficients is established.

[0095] Figure 12 The calculated wavelet energy coefficient distribution is shown.

[0096] Considering that the wavelet energy coefficient analysis method described above cannot further refine the wavelet decomposition coefficients of characteristic frequency bands, its effectiveness in monitoring the health of water turbines is not significant. Therefore, a statistical method based on the mean square value of wavelet decomposition coefficients is combined. The technical approach is as follows: First, the vibration signal after semi-soft threshold denoising is decomposed into components of J+1 frequency ranges through J scales. Second, a threshold, denoted as T, is set for the product of the mean square value of the wavelet coefficients and a weighting coefficient less than 1 in the energy-dense frequency bands. When the wavelet decomposition coefficient is greater than T, the number of wavelet decomposition coefficients greater than T is counted, denoted as N. The mean square value of the wavelet decomposition coefficients in each characteristic frequency band is represented by the parameter R(N). Finally, the correspondence between the number of wavelet decomposition coefficients N greater than the threshold T and R(N) is established. The weighting calculation formula is as follows:

[0097]

[0098]

[0099]

[0100] In the formula, c(x)k represents the detail coefficient of the kth order, c(x)ston is the weighting coefficient of the n modes, S(x) is the squared residual of the signal (the deviation of each superimposed data point from the average value of the superimposed coefficients), and S(x) is the normalized weighting coefficient.

[0101] Currently, traditional signal extraction methods mainly include Fast Fourier Transform (FFT) and Gabor Transform. However, due to the large amount and complexity of information contained in vibration signals, traditional signal extraction methods are ineffective in vibration signal feature extraction, often resulting in low accuracy and poor flexibility of extracted features, which significantly impacts subsequent diagnostic and prediction processes. Wavelet analysis, as a novel signal extraction method that has received considerable attention in recent years, has demonstrated significantly superior feature extraction capabilities compared to traditional methods in time-frequency domain studies. Applying wavelet analysis to the extraction of vibration features from hydroelectric turbines provides an opportunity to address these challenges, helping to achieve efficient and accurate extraction of vibration features and overcoming the limitations of traditional signal extraction methods.

[0102] (3) Feature Relationship Mining

[0103] In traditional intelligent diagnostic architectures, machine learning models are primarily used to perform intelligent diagnosis based on extracted signal features. Since these diagnostic models are mainly shallow, ensuring diagnostic accuracy often requires extremely high precision in the extracted features. However, driven by massive amounts of data, and facing a vast amount of mechanical signals with alternating operating conditions, highly coupled fault information, and unclear and variable patterns, the extracted signal features often contain significant errors. This makes it difficult for shallow models in traditional models to accurately represent the complex mapping relationship between signals and health conditions, resulting in poor monitoring and diagnostic capabilities and generalization performance.

[0104] To address the aforementioned issues, this invention constructs a deep learning-based uncertainty Bayesian neural network model to effectively mine the feature relationships between turbine data. The technical approach is as follows: First, by combining a DNN neural network model with probabilistic modeling methods, a novel big data uncertainty Bayesian neural network model is built. Then, a sample database is used for iterative training until the divergence is minimized, indicating training is complete. Finally, the neural network model is tested on a test set, comparing the predicted fault types with the actual fault types. If the model's prediction accuracy meets the requirements, it is determined that the model can effectively predict turbine faults, thus enabling uncertainty prediction of turbine faults. Ultimately, this forms a rapid diagnostic system for big data-driven intelligent fault diagnosis of turbines.

[0105] With the rapid rise and development of Industrial Internet and Internet of Things technologies, sensor networks in mechanical equipment are becoming denser, the monitored equipment groups are becoming larger, and the time required for data collection is significantly increasing. Therefore, monitoring systems acquire massive amounts of data, propelling mechanical fault diagnosis into the "big data" era. Deep learning-based health monitoring of hydro-turbine units can overcome the limitations of traditional methods that overly rely on diagnostic experts and technicians, breaking the deadlock between the large volume of diagnostic data and the relative scarcity of diagnostic experts, thus greatly improving monitoring efficiency and accuracy. To this end, this patent proposes an uncertainty Bayesian neural network model based on deep learning, using the large amount of data collected by sensors to train the model, thereby achieving the monitoring of hydro-turbine units. This not only avoids economic losses caused by turbine unit malfunctions but also significantly reduces safety hazards during turbine unit operation, solving the limitations of traditional intelligent identification and diagnostic models when predicting massive amounts of mechanical data.

[0106] It is understood that the technical solution provided in this embodiment is used for intelligent diagnosis of abnormal states of hydro-generator units in the context of big data, and to realize timely reminders of abnormal states to avoid major failures. An adaptive wavelet semi-soft threshold denoising method is proposed to preprocess the big data signals collected by the turbine sensors, achieving effective noise removal. An adaptive wavelet analysis method and a feature statistical analysis method are combined to extract features from the turbine vibration signals. First, an adaptive wavelet semi-soft threshold denoising method is used to denoise the original vibration signal to obtain a high-precision input signal. Then, a combination of wavelet energy system analysis and wavelet decomposition coefficient mean square value feature statistical analysis is employed to extract characteristic data of vibration and turbine unit response under different fault modes, establishing a turbine unit diagnostic sample database. Finally, a deep learning-based uncertainty Bayesian neural network model is constructed to effectively mine fault features, enabling rapid and accurate prediction of turbine faults under big data conditions. This method and device can adapt to the working environment requirements of the turbine unit, ensuring effective health monitoring of the turbine unit during operation, effectively avoiding safety hazards caused by turbine operation faults, and reducing manual fault diagnosis and maintenance costs.

[0107] Example 3

[0108] like Figure 13 As shown, this embodiment of the invention also provides a turbine monitoring device, comprising:

[0109] Denoising module 131 is used to denoise the turbine vibration signal data using an adaptive wavelet semi-soft threshold denoising method;

[0110] The feature extraction module 132 is used to extract feature data from the vibration signal data by combining wavelet energy coefficient analysis method and wavelet decomposition coefficient mean square value statistical analysis method. The feature data is used to characterize the vibration of the turbine under different fault modes and the turbine unit response, and to establish a turbine unit diagnostic sample database.

[0111] Feature relationship determination module 133 is used to mine feature relationships between water turbine data using a deep learning-based uncertain Bayesian neural network model.

[0112] In specific practice, such as Figure 13 As shown by the dashed line, the noise reduction module 131 includes:

[0113] The decomposition and adaptive submodule 1311 is used to perform multi-scale decomposition and adaptive denoising on the noisy vibration signal data using wavelet transform.

[0114] The vibration signal data acquisition submodule 1312 is used to reconstruct the processed wavelet coefficients through inverse wavelet transform to obtain high-precision vibration signal data after denoising.

[0115] In practice, the noise reduction module 131 uses the formula and The vibration signal data of the water turbine is denoised, where thr(i) is the adaptive threshold of the i-th layer wavelet decomposition, and d i Let σ be the detail coefficients of the i-th level of wavelet decomposition. 2 Let rw be the noise variance of the signal. i For the reconstructed i-th layer wavelet, a, b, and C are constants.

[0116] In practice, as not shown in the figure, the feature extraction module 132 includes:

[0117] The function generation submodule 132a-1 is used to perform wavelet decomposition on the denoised vibration signal data using basic functions to obtain the wavelet transform function.

[0118] The energy coefficient calculation submodule 132a-2 is used to calculate the wavelet energy coefficient of each frequency band using the obtained wavelet transform function;

[0119] The distribution state establishment submodule 132a-3 is used to establish the distribution state of wavelet energy coefficients.

[0120] In practice, as not shown in the figure, the feature extraction module 132 includes:

[0121] The decomposition submodule 132b-1 is used to decompose the vibration signal data after semi-soft threshold denoising into J+1 frequency range components through J scales.

[0122] The threshold setting submodule 132b-2 is used to set the threshold of the product of the mean square value of the wavelet coefficients and the weighting coefficients less than 1 in the energy-dense frequency band, denoted by T.

[0123] The decomposition coefficient statistics submodule 132b-3 is used to count the number of wavelet decomposition coefficients greater than T when the wavelet decomposition coefficient is greater than T, denoted by N;

[0124] The submodule 132b-4, which establishes the relationship between the decomposition coefficients, uses the mean square value of the wavelet decomposition coefficients for each characteristic frequency band, represented by the parameter R(N), to establish the correspondence between the number N of wavelet decomposition coefficients greater than the threshold T and R(N).

[0125] In practice, the formula for calculating the weighting coefficients can be:

[0126]

[0127]

[0128]

[0129] In the formula, c(x)k represents the detail coefficient of the kth order, c(x)ston is the weighting coefficient of the n modes, S(x) is the squared residual of the signal, and S(x) is the normalized weighting coefficient.

[0130] In specific practice, such as Figure 13 As shown by the dashed line, the feature relationship determination module 133 includes:

[0131] Modeling submodule 1331 is used to combine DNN neural network models with probabilistic modeling methods to realize the modeling of novel big data uncertainty Bayesian neural network models;

[0132] The iterative submodule 1332 is used to repeatedly iterate using the sample database until the divergence of the neural network model reaches its minimum, which means the training is complete.

[0133] The test submodule 1333 is used to test the neural network model with a test set. It compares the fault types predicted by the neural network model with the actual fault types until the model's prediction accuracy meets the requirements.

[0134] It is understood that the technical solution provided in this embodiment is used for intelligent diagnosis of abnormal states of hydro-generator units in the context of big data, and to realize timely reminders of abnormal states to avoid major failures. An adaptive wavelet semi-soft threshold denoising method is proposed to preprocess the big data signals collected by the turbine sensors, achieving effective noise removal. An adaptive wavelet analysis method and a feature statistical analysis method are combined to extract features from the turbine vibration signals. First, an adaptive wavelet semi-soft threshold denoising method is used to denoise the original vibration signal to obtain a high-precision input signal. Then, a combination of wavelet energy system analysis and wavelet decomposition coefficient mean square value feature statistical analysis is employed to extract characteristic data of vibration and turbine unit response under different fault modes, establishing a turbine unit diagnostic sample database. Finally, a deep learning-based uncertainty Bayesian neural network model is constructed to effectively mine fault features, enabling rapid and accurate prediction of turbine faults under big data conditions. This method and device can adapt to the working environment requirements of the turbine unit, ensuring effective health monitoring of the turbine unit during operation, effectively avoiding safety hazards caused by turbine operation faults, and reducing manual fault diagnosis and maintenance costs.

[0135] Example 4

[0136] Based on the same technical concept, embodiments of this application also provide a computer device, including a memory 1 and a processor 2, such as... Figure 14 As shown, the memory 1 stores a computer program, and the processor 2 executes the computer program to implement the turbine monitoring method described above.

[0137] The memory 1 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, memory 1 can be an internal storage unit of the OTT video service monitoring system, such as a hard disk. In other embodiments, memory 1 can also be an external storage device of the OTT video service monitoring system, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Furthermore, memory 1 can include both internal storage units and external storage devices of the OTT video service monitoring system. Memory 1 can be used not only to store application software and various types of data installed in the OTT video service monitoring system, such as the code of the OTT video service monitoring program, but also to temporarily store data that has been output or will be output.

[0138] In some embodiments, processor 2 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 1 or process data, such as executing OTT video service monitoring programs.

[0139] It is understood that the technical solution provided in this embodiment is used for intelligent diagnosis of abnormal states of hydro-generator units in the context of big data, and to realize timely reminders of abnormal states to avoid major failures. An adaptive wavelet semi-soft threshold denoising method is proposed to preprocess the big data signals collected by the turbine sensors, achieving effective noise removal. An adaptive wavelet analysis method and a feature statistical analysis method are combined to extract features from the turbine vibration signals. First, an adaptive wavelet semi-soft threshold denoising method is used to denoise the original vibration signal to obtain a high-precision input signal. Then, a combination of wavelet energy system analysis and wavelet decomposition coefficient mean square value feature statistical analysis is employed to extract characteristic data of vibration and turbine unit response under different fault modes, establishing a turbine unit diagnostic sample database. Finally, a deep learning-based uncertainty Bayesian neural network model is constructed to effectively mine fault features, enabling rapid and accurate prediction of turbine faults under big data conditions. This method and device can adapt to the working environment requirements of the turbine unit, ensuring effective health monitoring of the turbine unit during operation, effectively avoiding safety hazards caused by turbine operation faults, and reducing manual fault diagnosis and maintenance costs.

[0140] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the turbine monitoring method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0141] The computer program product of the turbine monitoring method disclosed in the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the steps of the turbine monitoring method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0142] The present invention also discloses a computer program that, when executed by a processor, implements any of the methods described in the foregoing embodiments. This computer program product can be implemented specifically through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0143] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0144] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0145] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0146] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0147] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0148] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0149] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0150] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0151] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for monitoring a water turbine, characterized in that, include: An adaptive wavelet semi-soft thresholding denoising method is used to denoise the vibration signal data of the hydro turbine, including: The vibration signal data containing noise is decomposed into multiple scales and adaptively denoised using wavelet transform, including: selecting a target wavelet basis according to the characteristics of the vibration signal, decomposing the signal into N levels, calculating the detail coefficients from level 1 to level N and the approximation coefficients of level N to estimate the noise bias and calculating the semi-soft threshold for the corresponding level, and using the calculated semi-soft threshold to discriminate the signal coefficients, setting the wavelet coefficients with amplitudes lower than the semi-soft threshold to 0, and completely retaining or shrinking the wavelet coefficients with amplitudes higher than the semi-soft threshold. Finally, the processed wavelet coefficients are reconstructed using inverse wavelet transform to obtain high-precision vibration signal data after denoising. The multi-scale decomposition and adaptive denoising of the noisy vibration signal data using wavelet transform is as follows: (using the formula...) , and The vibration signal data of the water turbine is denoised, where, For the first i Adaptive thresholding for layer wavelet decomposition For wavelet decomposition of the 1st i Layer detail factor, The noise variance of the signal. For the reconstruction of the first i Layer wavelet, a , b and C It is a constant; The vibration signal data is extracted by combining wavelet energy coefficient analysis with wavelet decomposition coefficient mean square value statistical analysis, including: Wavelet transform function is obtained by performing wavelet decomposition on the denoised vibration signal data using basic functions. The wavelet energy coefficients of each frequency band are calculated using the obtained wavelet transform function; Establish the distribution state of wavelet energy coefficients; The vibration signal data after semi-soft threshold denoising is decomposed into J+1 frequency range components by J scales; In energy-dense frequency bands, a threshold is set for the product of the mean square value of the wavelet coefficients and a weighting coefficient less than 1, denoted by T. When the wavelet decomposition coefficients are greater than T, count the number of wavelet decomposition coefficients greater than T, denoted by N; The mean square value of the wavelet decomposition coefficients of each characteristic frequency band is represented by the parameter R(N). The correspondence between the number N of wavelet decomposition coefficients greater than the threshold T and R(N) is established. The feature data is used to characterize the vibration and turbine unit response under different fault modes of the turbine, and to establish a turbine unit diagnostic sample database. We used a deep learning-based uncertainty Bayesian neural network model to mine the feature relationships between water turbine data.

2. A turbine monitoring device, characterized in that, include: The denoising module is used to denoise the turbine vibration signal data using an adaptive wavelet semi-soft thresholding method, including: The vibration signal data containing noise is decomposed into multiple scales and adaptively denoised using wavelet transform, including: selecting a target wavelet basis according to the characteristics of the vibration signal, decomposing the signal into N levels, calculating the detail coefficients from level 1 to level N and the approximation coefficients of level N to estimate the noise bias and calculating the semi-soft threshold for the corresponding level, and using the calculated semi-soft threshold to discriminate the signal coefficients, setting the wavelet coefficients with amplitudes lower than the semi-soft threshold to 0, and completely retaining or shrinking the wavelet coefficients with amplitudes higher than the semi-soft threshold. Finally, the processed wavelet coefficients are reconstructed using inverse wavelet transform to obtain high-precision vibration signal data after denoising. The multi-scale decomposition and adaptive denoising of the noisy vibration signal data using wavelet transform is as follows: (using the formula...) , and The vibration signal data of the water turbine is denoised, where, For the first i Adaptive thresholding for layer wavelet decomposition For wavelet decomposition of the 1st i Layer detail factor, The noise variance of the signal. For the reconstruction of the first i Layer wavelet, a , b and C It is a constant; The feature extraction module is used to extract feature data from the vibration signal data by combining wavelet energy coefficient analysis and wavelet decomposition coefficient mean square value statistical analysis. The feature data is used to characterize the vibration under different fault modes of the turbine and the turbine unit response, and to establish a turbine unit diagnostic sample database. The feature extraction module includes: The function generation submodule is used to perform wavelet decomposition on the denoised vibration signal data using basic functions to obtain the wavelet transform function. The energy coefficient calculation submodule is used to calculate the wavelet energy coefficient of each frequency band using the obtained wavelet transform function; The distribution state establishment submodule is used to establish the distribution state of wavelet energy coefficients; The decomposition submodule is used to decompose the vibration signal data after semi-soft thresholding into J+1 frequency range components through J scales. The threshold setting submodule is used to set the threshold for the product of the mean square value of the wavelet coefficients and the weighting coefficients less than 1 in the energy-dense frequency band, denoted by T. The decomposition coefficient statistics submodule is used to count the number of wavelet decomposition coefficients greater than T when the wavelet decomposition coefficient is greater than T, denoted by N. The submodule for establishing the relationship between decomposition coefficients is used to represent the mean square value of the wavelet decomposition coefficients for each characteristic frequency band using the parameter R(N), and to establish the correspondence between the number N of wavelet decomposition coefficients greater than the threshold T and R(N). The feature relationship determination module is used to mine the feature relationships between water turbine data using a deep learning-based uncertain Bayesian neural network model.

3. A computer device, characterized in that, include: The system includes a processor, a memory, and a bus. The memory stores machine-readable instructions that the processor can execute. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the turbine monitoring method as described in claim 1 is performed.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the turbine monitoring method as described in claim 1.

Images