Method and device for monitoring bolt loosening of a hydroelectric unit

Abstract

The application provides a kind of hydroelectric generating set bolt loosening monitoring method and device, it is related to bolt loosening monitoring technical field.The method includes in the key area of hydroelectric generating set vibration sensor and ultrasonic sensor are laid out, vibration signal and ultrasonic signal are synchronously collected, and are recorded as original signal;Original signal is carried out adaptive wavelet threshold denoising processing, and obtains the sound vibration signal to be diagnosed;Condition generation is based on adversarial network and support vector machine to build diagnostic model;From the sound vibration signal to be diagnosed, time domain feature and frequency domain feature are extracted, the working condition parameter vector of hydroelectric generating set is acquired, time domain feature, frequency domain feature and working condition parameter vector are spliced, and fusion feature vector is obtained;Fusion feature vector is input into diagnostic model, and the loosening state of bolt is monitored by diagnostic model.The application can realize effective, real-time, comprehensive bolt loosening state monitoring.

Description

Technical Field

[0001] This invention relates to the field of bolt loosening monitoring technology, specifically to a method and device for monitoring bolt loosening in hydropower units. Background Technology

[0002] The core mechanical structure of hydroelectric generating units (such as the turbine top cover and rotor support) relies on high-strength bolt connections. Bolt loosening is one of the most frequent faults in hydroelectric equipment. Early loosening can lead to abnormal vibration and noise, and if not addressed in time, it may cause bolt breakage, component displacement, or even unit shutdown.

[0003] Currently, the industry mainly relies on manual monitoring: regular manual inspections and spot checks of bolt preload using torque wrenches. This method is inefficient and cannot monitor in real time. It is limited by the complexity of the unit structure (such as the top cover bolts being located in a confined space), making it difficult to cover all critical bolts. Furthermore, it heavily depends on the conscientiousness of the staff and fails to achieve effective, real-time, and comprehensive monitoring and early warning of bolt loosening.

[0004] Existing technologies include bolt loosening identification methods based on image feature point matching. However, these methods require ideal lighting and a direct viewing angle. In the dimly lit wind tunnel of a hydro turbine, where bolts are often obscured, the accuracy of image feature-based identification is low. Additionally, some methods combine acoustic signatures, vibration data, and operating condition data for fault analysis and diagnosis. However, these methods generally ignore the coupling effect of operating conditions (such as the impact of head changes on vibration; vibration is greater at lower heads, but noise is not amplified to the same extent, and the same applies when the turbine is operating at low load), which can easily lead to misjudgments. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for monitoring loose bolts in hydropower units, which can solve problems such as difficult sensor installation, small number of monitoring samples, high environmental noise, and interference from operating conditions in the process of monitoring loose bolts in hydropower units. This enables reliable, full-coverage, and real-time automated online monitoring of loose bolts in real industrial environments, and thus provides accurate early warning.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for monitoring bolt loosening in hydropower units, comprising:

[0007] Vibration sensors and ultrasonic sensors are deployed in key areas of the hydropower unit to simultaneously collect vibration and ultrasonic signals, which are recorded as raw signals. Adaptive wavelet threshold noise reduction processing is performed on the raw signals to obtain the acoustic and vibration signals to be diagnosed.

[0008] A diagnostic model is constructed based on conditional generative adversarial networks and support vector machines;

[0009] Extract time-domain and frequency-domain features from the acoustic and vibration signals to be diagnosed, obtain the operating parameter vector of the hydropower unit, and concatenate the time-domain features, frequency-domain features, and operating parameter vector to obtain a fused feature vector.

[0010] The fused feature vectors are input into the diagnostic model, which monitors the loosening status of the bolts.

[0011] According to the present invention, a method for monitoring bolt loosening in a hydropower unit is provided, and the key areas include the top cover or the spiral casing.

[0012] According to the present invention, a method for monitoring bolt loosening in a hydropower unit includes a working parameter vector comprising head and load.

[0013] According to the present invention, a method for monitoring bolt loosening in a hydropower unit involves adaptive wavelet threshold noise reduction processing of the original signal to obtain the acoustic vibration signal to be diagnosed, including:

[0014] The original signal is decomposed by discrete wavelet transform to obtain high-frequency detail coefficients and low-frequency approximation coefficients at multiple scales; the high-frequency detail coefficients are processed by a soft threshold function to obtain the processed detail coefficients; the low-frequency approximation coefficients and the processed detail coefficients are reconstructed by inverse discrete wavelet transform to obtain the acoustic vibration signal to be diagnosed.

[0015] According to the present invention, a method for monitoring bolt loosening in hydropower units is provided, which constructs a diagnostic model based on conditional generative adversarial networks and support vector machines, including:

[0016] Acquire historical normal acoustic and vibration signals and actual loosening signals of bolts in key areas, as well as corresponding historical working condition parameter vectors and historical random noise vectors.

[0017] Construct a conditional generative adversarial network, which includes a generator and a discriminator;

[0018] Historical normal acoustic and vibration signals, historical working condition parameter vectors, and historical random noise vectors are input into the generator to generate simulated loosening signals. Based on the simulated loosening signals and real loosening signals, the conditional generative adversarial network is trained by minimizing the loss functions of the generator and the discriminator.

[0019] Time-domain and frequency-domain features are extracted from multiple historical normal acoustic and vibration signals, real loosening signals, and simulated loosening signals generated by the generator after the conditional generative adversarial network is trained. These features are then concatenated with the corresponding historical working condition parameter vectors to obtain multiple samples. The corresponding bolt loosening state is used as the label for each sample to construct multiple training samples. A support vector machine is then trained using these training samples to obtain a diagnostic model.

[0020] According to the present invention, a method for monitoring bolt loosening in hydropower units has the following loss function for the generator:

[0021]

[0022] Among them, s normal Let θ be the historical normal acoustic and vibration signal, θ be the historical operating condition parameter vector, z be the historical random noise vector, G represent the generator, D represent the discriminator, log be the natural logarithm operation, and z ~ p. z (z) represents the historical random noise vector z following a prior distribution p. z (z), where E represents the mathematical expectation;

[0023] The loss function of the discriminator is:

[0024]

[0025] Among them, s real This is a genuine sign of loosening.

[0026] According to the present invention, a method for monitoring bolt loosening in hydropower units includes time-domain features such as mean, effective value, standard deviation, kurtosis, peak factor, and impulse factor, and frequency-domain features such as spectral centroid, spectral bandwidth, and harmonic factor.

[0027] According to the present invention, a method for monitoring bolt loosening in hydropower units is provided, wherein the expressions for mean, effective value, standard deviation, kurtosis, peak factor, and impulse factor are as follows:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] Where μ is the mean, σ is the standard deviation, RMS is the effective value, CF is the peak factor, and IF is the impulse factor. x i Let be the acoustic vibration signal to be diagnosed at the i-th sampling point, N be the total number of sampling points, and max represent the maximum value function.

[0034] According to the present invention, a method for monitoring bolt loosening in hydropower units extracts frequency domain features from the acoustic vibration signal to be diagnosed, including:

[0035] Perform a Fast Fourier Transform on the acoustic vibration signal to be diagnosed:

[0036]

[0037] in, This represents the complex spectrum corresponding to the frequency index k. This represents the instantaneous value of the acoustic vibration signal to be diagnosed in the frame at the (n+1)th sampling point, where j represents the imaginary unit, e represents the base of the natural logarithm, and N is the total number of sampling points;

[0038] The power spectral density at frequency index k is calculated as follows:

[0039]

[0040] Then, the spectral centroid, spectral bandwidth, and harmonic factors are extracted, and the expression is:

[0041]

[0042]

[0043]

[0044] in, For the spectral centroid, For spectrum bandwidth, M represents the harmonic factor; M represents the number of effective frequency points, M=N / 2; f k This indicates the actual physical frequency corresponding to frequency index k. , f s Indicates the sampling frequency; f base H represents the fundamental frequency of the hydroelectric generator, H represents the harmonic order, and P represents the power spectral density.

[0045] Secondly, the present invention provides a device for monitoring loose bolts in hydropower units, comprising:

[0046] The data acquisition and processing unit is used to deploy vibration sensors and ultrasonic sensors in key areas of the hydropower unit to simultaneously acquire vibration signals and ultrasonic signals, which are recorded as raw signals; adaptive wavelet threshold noise reduction processing is performed on the raw signals to obtain the acoustic and vibration signals to be diagnosed.

[0047] Building units are used to construct diagnostic models based on conditional generative adversarial networks and support vector machines;

[0048] The feature extraction and fusion unit is used to extract time-domain and frequency-domain features from the acoustic and vibration signals to be diagnosed, obtain the operating parameter vector of the hydropower unit, and concatenate the time-domain features, frequency-domain features and operating parameter vector to obtain the fused feature vector.

[0049] The monitoring unit is used to input the fused feature vector into the diagnostic model, which then monitors the loosening status of the bolts.

[0050] Compared with the prior art, the present invention has at least the following technical effects:

[0051] 1. Breakthrough in non-invasive monitoring capabilities. The use of non-invasive vibration and ultrasonic sensors solves the installation challenges in enclosed spaces such as turbine roofs, resulting in high deployment efficiency.

[0052] 2. High-precision signal extraction in noisy environments. The adaptive wavelet denoising algorithm can still achieve a signal-to-noise ratio improvement of ΔSNR ≥ 15dB even with background noise greater than 90dB.

[0053] 3. Micro-loosening samples are generated using a conditional generative adversarial network, with a feature error of less than 5%;

[0054] 4. By concatenating time-domain features, frequency-domain features, and operating condition parameter vectors to obtain a fused feature vector as input, the operating condition parameters such as head and load are effectively referenced, which can eliminate misjudgments caused by operating condition fluctuations (such as increased vibration under low head, which is prone to false alarms in existing technologies). Attached Figure Description

[0055] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0056] In the attached diagram:

[0057] Figure 1 This is a flowchart of the method for monitoring loose bolts in hydropower units according to the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0059] The following detailed description of some embodiments of the present invention will be provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0060] Please see Figure 1This invention provides a method for monitoring bolt loosening in hydropower units, which can perform online monitoring of bolt loosening conditions, including the following steps:

[0061] Step 1: Signal Acquisition and Noise Reduction. Vibration sensors and ultrasonic sensors are deployed in key areas of the hydropower unit to simultaneously acquire vibration and ultrasonic signals, which are recorded as the raw signals. Adaptive wavelet threshold noise reduction is performed on the raw signals to obtain the acoustic and vibration signals to be diagnosed.

[0062] Specifically, step 1 uses non-invasive sensing and adaptive noise reduction technology to solve the signal acquisition problem under strong noise interference. The specific process is as follows:

[0063] (1) Sensor deployment and synchronous signal acquisition

[0064] Vibration sensors and ultrasonic sensors are installed in key areas such as the top cover and spiral casing of the hydropower unit. The vibration sensors are used to collect mechanical vibration signals caused by loose bolts on the hydropower unit, and the ultrasonic sensors are used to collect ultrasonic signals generated by loose bolts on the hydropower unit.

[0065] Vibration sensors and ultrasonic sensors can be magnetically or threadedly fixed to the surface of critical areas of hydropower units without modifying the bolt body structure. This is a non-invasive monitoring solution that can avoid affecting mechanical performance.

[0066] Specifically, a time synchronization device is used to synchronize the acquisition time of the vibration sensor and the ultrasonic sensor, with a synchronization error of less than or equal to 0.8 μs. The sampling frequency can be set to 10 kHz (covering the characteristic frequency band of bolt loosening, 2 kHz–5 kHz).

[0067] (2) Wavelet thresholding noise reduction

[0068] To address the high background noise environment of hydropower stations (typically greater than 90 dB; for example, the noise level in the turbine room during hydropower station operation measured as high as 92.9 dB), this invention designs an adaptive wavelet threshold denoising algorithm to suppress noise while preserving the transient impact characteristics generated in the initial stage of bolt loosening. This adaptive wavelet threshold denoising algorithm mainly includes three steps: decomposition, threshold processing, and reconstruction.

[0069] 1. Record the collected vibration and ultrasonic signals as the original signals. x ( t The original signal is transformed using Discrete Wavelet Transform (DWT). x ( t The decomposition yields high-frequency detail coefficients and low-frequency approximation coefficients at multiple scales:

[0070]

[0071] Where j is the number of decomposition levels in the discrete wavelet transform, adaptively selected based on the original signal length (i.e., the total number of sampling points) N to ensure coverage of the characteristic frequency band of the loose bolt (2–5 kHz); where j = 1, 2, …, J; J is the maximum number of decomposition levels, usually taken as the largest integer less than or equal to log2N. k represents the discretization index of the translation parameter, representing the position of the function on the time axis. This represents the scaling function at the j-th layer and translation index k. This represents the wavelet function at the j-th layer and translation index k. This represents the inner product operation, which is the summation operation in discrete signals.

[0072] cA j [k] represents the k-th approximation coefficient of the j-th level (scale), derived from the original signal. x ( t ) and scaling function The inner product yields the low-frequency components that represent the overall characteristics of the signal. CD j [k] represents the k-th detail coefficient of the j-th layer (scale), derived from the original signal. x ( t ) and wavelet function The inner product is obtained, which represents the high-frequency detail components of the signal (noise and transient features are mainly present here).

[0073] 2. Adaptive Threshold Calculation. To effectively remove noise and retain useful features, this invention employs an adaptive threshold calculation method based on signal statistical characteristics, where the threshold is dynamically adjusted according to the noise energy.

[0074] The threshold for the detail factor of the j-th layer is dynamically determined based on the background noise energy. l j The formula for calculation is:

[0075]

[0076] in, l j The threshold value is the detail coefficient of the j-th layer. If the absolute value of the detail coefficient is less than this threshold value, it is considered noise and suppressed. s j is the standard deviation of the detail coefficients at layer j (characterizing noise intensity); N is the length of the original signal. Threshold l j Dynamically adjusting based on noise statistical characteristics avoids filtering or under-filtering caused by a fixed threshold. l j The calculation formula can be based on the noise intensity of the corresponding layer. s jThe system automatically determines the optimal threshold based on the original signal length N, avoiding problems such as filtered waveforms (signal distortion) or under-filtering (noise residue) that may occur with a fixed threshold.

[0077] 3. Soft thresholding function processing. This involves handling the detail coefficients. CD j [k] A soft thresholding function is applied for nonlinear filtering to smooth out minor fluctuations caused by noise while preserving the significant transient impact caused by bolt loosening. Applying a soft thresholding function can suppress noise while retaining signal edge characteristics (corresponding to the transient impact in the initial stage of bolt loosening). The expression is:

[0078]

[0079] in, This represents the detail coefficient of the k-th layer after processing by the soft thresholding function, which contains the retained effective detail information. The sign function is denoted by , which preserves the sign of the detail coefficients. `max` is the maximum value function, ensuring that the processed detail coefficients are not negative.

[0080] 4. Signal Reconstruction. The denoised signal is generated using Inverse Discrete Wavelet Transform (IDWT), expressed as:

[0081]

[0082] in, The reconstructed, denoised signal, i.e. the acoustic vibration signal to be diagnosed, serves as the input for subsequent feature extraction. This represents the k-th approximation coefficient of the J-th layer (i.e., the largest scale J), ​​which contains general information about the signal. This represents the scaling function at the J-th layer and translation index k.

[0083] The reconstruction process involves combining the low-frequency profile of the signal (the first term) with the effective high-frequency details preserved at all scales (the second term) to ultimately obtain a denoised signal with a high signal-to-noise ratio.

[0084] (3) Verification of noise reduction effect

[0085] To quantify the performance of the adaptive wavelet threshold denoising of this invention, the following two metrics are used for evaluation:

[0086] Signal-to-noise ratio improvement (ΔSNR):

[0087]

[0088] in, This represents the pure reference signal value at the i-th sampling point (which can be obtained by sampling in an extremely low noise environment or by advanced noise reduction methods). This represents the noise-reduced acoustic vibration signal value at the i-th sampling point. The signal-to-noise ratio (SNR) represents the original signal. It represents the SNR of the original signal directly acquired from the sensor before any noise reduction processing, and serves as a benchmark for evaluating the performance of noise reduction algorithms.

[0089] in,

[0090]

[0091] It represents the power of the useful signal component in the original signal. This represents the power of the background noise component in the original signal. The root mean square error (RMSE) metric is primarily used to measure the noise reduction algorithm's ability to suppress background noise; a higher value indicates a more significant noise reduction effect.

[0092] Root Mean Square Error (RMSE):

[0093]

[0094] RMSE measures the average deviation between the denoised signal and the clean signal; a smaller value indicates less signal distortion. This RMSE metric is primarily used to ensure that effective signal characteristics are highly preserved while denoising.

[0095] Verification shows that the adaptive wavelet threshold denoising of the present invention can achieve ΔSNR≥15dB and RMSE≤0.05, which meets the actual denoising requirements.

[0096] Step 2: Construct a diagnostic model based on conditional generative adversarial networks and support vector machines;

[0097] In some embodiments, step 2 specifically includes:

[0098] Step 21: Obtain the historical normal acoustic and vibration signals and actual loosening signals of bolts in the key area, as well as the corresponding historical working condition parameter vector and historical random noise vector;

[0099] Step 22: Construct a conditional generative adversarial network (cGAN), which includes a generator and a discriminator;

[0100] Step 23: Input the historical normal acoustic and vibration signals, historical operating condition parameter vectors, and historical random noise vectors into the generator to generate a simulated loosening signal; based on the simulated loosening signal and the real loosening signal, minimize the generator's loss function L. G The loss function L of the discriminator D Complete the training of the conditional generative adversarial network;

[0101] Wherein, the generator's loss function L G for:

[0102]

[0103] Among them, s normal Let θ be the historical normal acoustic and vibration signal, θ be the historical operating condition parameter vector, z be the historical random noise vector, G represent the generator, D represent the discriminator, log be the natural logarithm operation, and z ~ p. z (z) represents the historical random noise vector z following a prior distribution p. z (z) (usually a standard normal distribution), where E represents the expected value;

[0104] The loss function L of the discriminator D for:

[0105]

[0106] Among them, s real The signal represents a real loosening; the loss function L D The first item is to let the discriminator... D For the actual loosening signal s real The output probability is close to 1 (correctly identifying the real sample), and the second term is to make the discriminator... D For the generated simulated loosening signal s gen The output probability is close to 0 (correctly identified generated sample).

[0107] It should be noted that by training conditional generative adversarial networks (GANs), the problem of scarce real loosening samples can be solved, generating high-fidelity micro-loosening feature data. Here, the generator G is a neural network whose goal is to generate simulated loosening signals that the discriminator D cannot distinguish from real loosening signals. Mathematically, this is expressed as:

[0108]

[0109]

[0110] in, Represents historical normal acoustic vibration signals (dimension d = number of sampling points);

[0111] Represents a historical random noise vector (dimension k=64, to achieve diversity in generated signals, following a standard normal distribution N(0,1)).

[0112] This represents a vector of historical operating parameters (dimension m=2, including head / load).

[0113] Among them, the historical operating condition parameter vector (water head) ,load Both are normalized to [0,1]).

[0114] Head (unit: meters, meaning the difference between upstream and downstream water levels) and load (unit: megawatts, i.e. the power generation capacity of the hydropower unit) are the core operating parameters of the hydropower unit. They are usually obtained in real time from the monitoring system of the hydropower plant (such as the SCADA (Supervisory Control and Data Acquisition) system), which continuously monitors data such as the head height and power generation load of the hydropower unit.

[0115] In this invention, operating condition parameters are used to ensure that the generated simulated loosening signals and diagnostic models match the current operating state, thereby avoiding misjudgments caused by fluctuations in operating conditions.

[0116] Normalization is performed using the standard minimum-maximum scaling method, and the formula for the normalization parameters is as follows:

[0117]

[0118]

[0119] Where: h and l These are the actual head value and the load value, respectively; h min and h max These are the historical minimum and maximum values ​​of the water head (or the endpoint values ​​of the design range), for example, the water head may be between 30 and 45 meters; l min and l max These are the historical minimum and maximum load values ​​(e.g., 0–300 MW).

[0120] After normalization, and It is mapped to the [0,1] interval for easy processing by neural networks. In some embodiments, a head of 0.72 corresponds to an actual value of approximately 40.8 meters (assuming a range of 30–45 meters), and a load of 0.85 corresponds to an actual value of approximately 200 MW (assuming a range of 0–300 MW).

[0121] This indicates the generated simulated loosening signal (length). T Same as input).

[0122] The discriminator D is also a neural network. It receives a signal (which can be a real loosening signal or a simulated loosening signal generated by a generator), and outputs a bolt loosening probability ranging from 0 to 1, representing the probability that the input signal is a real loosening signal. Mathematically, this is expressed as:

[0123]

[0124]

[0125] in, Indicates the input signal (real loosening signal or simulated loosening signal);

[0126] probability value This represents a probability estimate that the input signal is a genuine loosening signal.

[0127] This invention uses binary cross-entropy loss to determine the loss function L of the generator. G The loss function L of the discriminator D Wherein, the generator's loss function L G Its function is to allow the generator G to generate the signal s. gen Discriminator D Misinterpreted as a "real" loosening signal. Minimize L. G In other words, it maximizes the probability of the discriminator D misclassifying the generated samples. Minimize L D This means improving the accuracy of the discriminator.

[0128] in, These are real loosening signal samples (scarce data) used to train the discriminator's "true" judgment. The signal is fabricated by the generator to train the discriminator to make "false" judgments.

[0129] The loss function L D Aimed at training the discriminator This enables it to correctly distinguish between genuine loosening signals. With the generated signal That is, to maximize the probability of recognizing genuine loosening signals. At the same time, minimize the probability of misjudging the generated signal.

[0130] Step 24: Extract time-domain and frequency-domain features from multiple historical normal acoustic and vibration signals, real loosening signals, and simulated loosening signals generated by the generator after the conditional generative adversarial network training is completed. Concatenate these features with the corresponding historical working condition parameter vectors to obtain multiple samples. Use the corresponding bolt loosening state as the label for each sample to construct multiple training samples. Use the multiple training samples to train the support vector machine to obtain the diagnostic model.

[0131] It should be noted that Support Vector Machine (SVM) is a commonly used machine learning algorithm, particularly suitable for small-sample, non-linear binary classification problems. Its basic idea is to find an optimal hyperplane in the feature space such that samples of different classes can be separated by the maximum margin. For bolt state recognition, the loose state and the normal state can be considered as two categories, and the multidimensional features collected from the bolt can be regarded as sample points. The optimal state discrimination model is constructed in the feature space using SVM. The SVM model is trained using historical data from the normal state sample library. The main steps are as follows: a certain number of samples are selected from the normal state sample library as the training set. Each sample contains multiple feature variables and a corresponding state label; the normal state label is -1, and the loose state label is +1. All feature variables are normalized to scale their numerical range to [0, 1] or [-1, +1] to avoid the influence of features with different dimensions. For linearly separable cases, the goal of SVM is to find an optimal hyperplane such that the two classes of samples can be separated by the hyperplane with the maximum margin. For linearly inseparable cases, slack variables and penalty factors can be introduced to allow for a small number of misclassifications, while controlling the cost of misclassification through the penalty factor. The above problem can be solved by finding the optimal solution to a convex quadratic programming problem, and the resulting optimal hyperplane parameters can be used to construct the decision function of the SVM. For nonlinear classification problems, kernel functions can be used to map samples from the original space to a high-dimensional feature space, making them linearly separable. Commonly used kernel functions include polynomial kernels and Gaussian kernels (RBF kernels). Using the training set data, the optimal parameters of the model are obtained by solving the convex quadratic programming problem of the SVM. For nonlinear SVMs, the parameters of the kernel function need to be selected. A common method is to combine cross-validation with grid search to select the parameter combination with the best average performance. During training, it is also necessary to balance the model's fitting ability and generalization performance, controlling the model's complexity by adjusting the penalty factor or kernel function parameters to avoid overfitting. Once the SVM is trained, it can be used for real-time diagnosis of bolt loosening status based on online monitoring data. Multiple features are extracted from the collected acoustic and vibration signals to be diagnosed, following the same method as in the training phase. The features can be normalized to align with the numerical range of the training samples. The concatenated fused feature vector is then substituted into the SVM's decision function to obtain the bolt's current state label γ (+1 or -1). Combined with the current operating parameters of the hydropower unit, the state label is further interpreted: if it is -1, the bolt is in a normal state; if it is +1, the bolt may be loose. Of course, for monitoring data of bolts identified as loose, their feature vectors can be compared with the SVM's support vectors to find the most similar loosening pattern, thus making a preliminary inference about the degree and cause of loosening.

[0132] Step 3: Extract time-domain and frequency-domain features from the acoustic and vibration signals to be diagnosed, obtain the operating parameter vector of the hydropower unit, and concatenate the time-domain features, frequency-domain features, and operating parameter vector to obtain the fused feature vector.

[0133] Specifically, extracting key features such as time-domain and frequency-domain features from the acoustic vibration signal to be diagnosed, and then concatenating them with the operating condition parameter vector to obtain a fused feature vector, can provide a highly discriminative input for subsequent diagnostic models.

[0134] It should be noted that the processes of extracting time-domain and frequency-domain features (hereinafter collectively referred to as acoustic vibration features) from the acoustic vibration signal to be diagnosed, historical normal acoustic vibration signals, real loosening signals, and simulated loosening signals generated by cGAN in steps 2 and 3 above, and then splicing them together, are similar. The only difference lies in the different signals. The following explanation uses the acoustic vibration signal to be diagnosed as an example.

[0135] In some embodiments, the time-domain features include mean, RMS value, standard deviation, kurtosis, peak factor, and impulse factor, while the frequency-domain features include spectral centroid, spectral bandwidth, and harmonic factors. The expressions for mean, RMS value (root mean square), standard deviation, kurtosis, peak factor, and impulse factor are:

[0136]

[0137]

[0138]

[0139]

[0140]

[0141] Where μ is the mean, σ is the standard deviation, RMS is the effective value, CF is the peak factor, and IF is the impulse factor. x i Let be the acoustic vibration signal to be diagnosed at the i-th sampling point, N be the total number of sampling points, and max represent the maximum value function.

[0142] It should be noted that kurtosis is suitable for detecting transient impacts or abnormal pulses in the acoustic vibration signal to be diagnosed. Under normal operating conditions, the kurtosis value of the acoustic vibration signal to be diagnosed is relatively small. Noise caused by bolt failure will cause the kurtosis value to gradually increase. Under slight loosening, the acoustic vibration signal to be diagnosed will show a small number of transient impacts, and the kurtosis value will gradually increase. Under severe loosening or even damage, the acoustic vibration signal to be diagnosed will show frequent transient impacts with a sharp distribution, and the kurtosis value will rise to the highest level.

[0143] The peak factor is an important time-domain characteristic that measures the relationship between the peak intensity of a signal and its overall energy. It is defined as the ratio of the peak value to the effective value of the acoustic / vibration signal to be diagnosed. Under normal operating conditions, the waveform of the acoustic / vibration signal to be diagnosed is relatively stable, with a small difference between the peak and effective values, resulting in a low waveform factor. When phenomena such as loosening or damage to bolts occur, the mechanical structure suddenly releases energy, generating a transient high-amplitude impact signal (such as from a collision or vibration). At this time, the peak value increases significantly, while the effective value increases more slowly due to the time averaging effect, leading to a sharp increase in the waveform factor.

[0144] Unlike the peak factor, the denominator in the formula for the impulse factor is the average value of the absolute value of the acoustic vibration signal to be diagnosed (reflecting the overall amplitude level). It is complementary to the peak factor CF: the peak factor CF is affected by the root mean square (RMS), while the impulse factor IF is affected by the mean. Using them together can reduce misjudgments.

[0145] The high-frequency component changes are often very obvious in the early stages of bolt loosening. The process of extracting frequency domain features from the acoustic vibration signal to be diagnosed is as follows:

[0146] Acoustic vibration signal to be diagnosed Perform a Fast Fourier Transform:

[0147]

[0148] This represents the complex spectrum corresponding to the frequency index k. This represents the instantaneous value of the acoustic vibration signal to be diagnosed in the frame at the (n+1)th sampling point, where j represents the imaginary unit, e represents the base of the natural logarithm, and N is the total number of sampling points;

[0149] The power spectral density at frequency index k is calculated as follows:

[0150]

[0151] Then, the spectral centroid, spectral bandwidth, and harmonic factors are extracted, and the expression is:

[0152]

[0153]

[0154]

[0155] in, For the spectral centroid, For spectrum bandwidth, is the harmonic factor; M represents the number of effective frequency points, M=N / 2 (based on the Nyquist frequency); f k This indicates the actual physical frequency (Hz) corresponding to the frequency index k. , fs Indicates the sampling frequency; f base This represents the fundamental frequency of a hydroelectric generator unit (i.e., the frequency at which the turbine generator rotates; taking a vertical hybrid axial-flow propeller turbine unit as an example, the fundamental frequency of the hydroelectric generator unit is 125 revolutions per minute, which, after conversion to the standard frequency unit Hertz, is: f base =125 / 60≈2.08Hz), H represents the harmonic order (usually taken as 5, that is, the calculation is up to the 5th harmonic), and P represents the power spectral density.

[0156] It's important to note the following: Spectral centroid (SC): This characteristic is similar to the "brightness" of sound; loosening the bolts will increase the high-frequency components. When the bolts are loose, the high-frequency components increase, and the "center of gravity" of the spectral energy, i.e., the centroid, shifts to the right. Spectral bandwidth (SB) reflects the degree of energy dispersion; the bandwidth widens when the bolts are loose. Harmonic factor (HR) reflects the proportion of fundamental frequency energy; loosening the bolts is often accompanied by nonlinear vibrations, which increases the harmonic components.

[0157] To address the interference of operating condition fluctuations on acoustic and vibration characteristics, this invention concatenates the operating condition parameter vector with the acoustic and vibration characteristics, using both as input to the diagnostic model. This allows the diagnostic model to learn fault characteristics under different operating conditions, thereby achieving adaptive diagnosis based on operating conditions. The concatenation of acoustic and vibration characteristics with the operating condition parameter vector can be represented as follows:

[0158] 9-dimensional acoustic vibration feature vector With 2D working condition parameter vector By concatenating the components, an 11-dimensional fused feature vector is obtained. :

[0159]

[0160] Step 4: Input the fused feature vector into the diagnostic model, and monitor the loosening status of the bolts through the diagnostic model.

[0161] Specifically, by using the spliced ​​fused feature vector as input, accurate identification of bolt loosening status can be achieved. To solve the interference of operating condition fluctuations on acoustic and vibration features, this invention splices the operating condition parameter vector and acoustic and vibration features together as input to SVM, enabling the diagnostic model to learn the loosening features under different operating conditions, thereby achieving operating condition adaptive diagnosis.

[0162] In practical applications, the diagnostic model can output the bolt loosening status at preset intervals, such as 10 seconds. When bolt loosening is detected, an early warning is issued, and the warning data is uploaded to the cloud, generating a bolt loosening report for subsequent analysis by staff. Thus, this invention provides an online monitoring and early warning mechanism.

[0163] Based on the same inventive concept, another embodiment of the present invention provides a device for monitoring loose bolts in hydropower units. This device corresponds to the method in the aforementioned embodiment and includes:

[0164] The data acquisition and processing unit is used to deploy vibration sensors and ultrasonic sensors in key areas of the hydropower unit to simultaneously acquire vibration signals and ultrasonic signals, which are recorded as raw signals; adaptive wavelet threshold noise reduction processing is performed on the raw signals to obtain the acoustic and vibration signals to be diagnosed.

[0165] Building units are used to construct diagnostic models based on conditional generative adversarial networks and support vector machines;

[0166] The feature extraction and fusion unit is used to extract time-domain and frequency-domain features from the acoustic and vibration signals to be diagnosed, obtain the operating parameter vector of the hydropower unit, and concatenate the time-domain features, frequency-domain features and operating parameter vector to obtain the fused feature vector.

[0167] The monitoring unit is used to input the fused feature vector into the diagnostic model, which then monitors the loosening status of the bolts.

[0168] The following are specific embodiments of the present invention.

[0169] Three vibration sensors (measuring range ±50g, frequency response 0.5-8kHz) and two ultrasonic sensors (resonant frequency 150kHz, sensitivity 80dB) were installed on the flange surface of the top cover of Unit 3 of a hydropower plant, 100mm away from the bolts to be measured. The precise time protocol IEEE 1588 was used for time synchronization, with a synchronization error less than or equal to 0.8μs. Vibration and ultrasonic signals were acquired synchronously.

[0170] The sampling frequency is 10kHz (which can cover the 2–5kHz bolt loosening characteristic frequency band), the single acquisition duration is 10 seconds, and the acquisition interval is: 5 minutes of cyclic acquisition.

[0171] The operating conditions are: load 200MW, head 85m.

[0172] Original signal before noise reduction: SNR=16.8dB (background noise 92dB).

[0173] Normal acoustic signal after noise reduction: SNR=34.5dB (ΔSNR=17.7dB).

[0174] In the adaptive wavelet thresholding denoising process, when the decomposition level j=6, l j =1.28, wavelet basis is db8.

[0175] Noise reduction performance metrics: ΔSNR≥15dB, RMSE≤0.05.

[0176] The design requirements were met, verifying that the present invention can effectively preserve the impact characteristics of loosened bolts.

[0177] During the training process of cGAN, the training samples corresponding to the historical normal acoustic vibration signal, the real loosening signal, and the generated simulated loosening signal are called real normal samples, real loosening samples, and generated samples, respectively. Together, they constitute the training dataset, as shown in Table 1. The training process is shown in Table 2, and the quality evaluation of the cGAN's production samples is shown in Table 3.

[0178] Table 1. Training Dataset

[0179]

[0180] The normalized head range is 70–120m, and the normalized load range is 0–300MW.

[0181] Table 2. Training Process

[0182]

[0183] Table 3. Quality assessment of production samples (selected characteristics)

[0184]

[0185] The test showed that the discriminator had an accuracy rate of 97.3% in identifying real loose samples and a misclassification rate of 38.6% for generated samples (that is, the probability that the discriminator would classify a generated sample as "real").

[0186] Results Analysis: After training, the discriminator maintained a high recognition rate for real loose samples, but a significant proportion of the generated samples successfully "deceived" the discriminator (the misclassification rate was much higher than the 50% random guess level). This indicates that the generator has been able to produce highly realistic loose signal samples sufficient to confuse the discriminator. Furthermore, the errors in key features between the generated samples and real loose samples are all below 5% (e.g., kurtosis K is only 3.8%), verifying that the cGAN-generated samples have high realism.

[0187] The working condition adaptability verification, namely the consistency of the characteristics of the generated samples under different working conditions, is shown in Table 4.

[0188] Table 4. Consistency of generated samples under different operating conditions

[0189]

[0190] Feature extraction and online monitoring process:

[0191] Location: Top cover bolts of Unit 3 at this hydropower station (manually loosened by 25°).

[0192] Operating conditions: water head 100m ( =0.91), load 200MW ( =0.50)

[0193] The time-domain features (6 dimensions) are shown in Table 5, and the frequency-domain features (3 dimensions) are shown in Table 6.

[0194] Table 5. Temporal Characteristics

[0195]

[0196] Table 6. Frequency Domain Characteristics

[0197]

[0198] The 9-dimensional acoustic vibration features are concatenated with the 2-dimensional operating condition parameter vector to form an 11-dimensional fused feature vector:

[0199]

[0200] During the training of the diagnostic model, the training dataset consisted of 850 normal samples (state label γ = -1), 35 real loose samples (state label γ = +1), and 5000 generated samples (state label γ = +1). After training, when the aforementioned 11-dimensional fused feature vector was input, the diagnostic model output a state label γ of +1, indicating that the bolt was loose.

[0201] In summary, this invention provides a method and device for monitoring bolt loosening in hydropower units, mainly comprising two core processes: the first is the "training process of the diagnostic model" (conducted offline and completed in one step); the second is the "online monitoring process of bolt loosening status" (continuous online monitoring). Through the deep integration of non-invasive sensor deployment, adversarial generative sample enhancement, and adaptive intelligent diagnostics based on operating conditions, the invention systematically solves problems such as difficult sensor installation, limited number of monitoring samples, high environmental noise, and operating condition interference in the monitoring of bolt loosening in hydropower units. This enables reliable, comprehensive, and real-time automated online bolt loosening monitoring in a real industrial environment, thereby providing accurate early warning.

[0202] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. It should be understood that the invention is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method of monitoring bolt looseness in a hydroelectric generating unit, characterized by, The method comprises the steps of: vibration sensors and ultrasonic sensors are arranged in key areas of the hydroelectric generating set to synchronously collect vibration signals and ultrasonic signals with an error less than or equal to 0.8 μs, and the signals are recorded as original signals; adaptive wavelet threshold denoising processing is performed on the original signals to obtain to-be-diagnosed acoustic vibration signals; a signal-to-noise ratio improvement value of the to-be-diagnosed acoustic vibration signals is greater than or equal to 15 dB; a diagnosis model is constructed based on a conditional generative adversarial network and a support vector machine; time domain features and frequency domain features are extracted from the to-be-diagnosed acoustic vibration signals, working condition parameter vectors of the hydroelectric generating set are obtained, the time domain features, the frequency domain features and the working condition parameter vectors are spliced to obtain a fusion feature vector; the fusion feature vector is input into the diagnosis model, and the loosening state of the bolt is monitored through the diagnosis model; the key areas include a top cover or a volute; the diagnosis model is constructed based on the conditional generative adversarial network and the support vector machine, and the method comprises the steps of: historical normal acoustic vibration signals and real loosening signals of the bolt in the key areas and corresponding historical working condition parameter vectors and historical random noise vectors are obtained; a conditional generative adversarial network is constructed, and the conditional generative adversarial network comprises a generator and a discriminator; the historical normal acoustic vibration signals, the historical working condition parameter vectors and the historical random noise vectors are input into the generator to generate simulated loosening signals; and the training of the conditional generative adversarial network is completed by minimizing a loss function of the generator and a loss function of the discriminator based on the simulated loosening signals and the real loosening signals; time domain features and frequency domain features are extracted from a plurality of the historical normal acoustic vibration signals, the real loosening signals and simulated loosening signals generated by the generator after the training of the conditional generative adversarial network, and the corresponding historical working condition parameter vectors are spliced to obtain a plurality of samples, and corresponding bolt loosening states are taken as labels of the plurality of samples, so that a plurality of training samples are constructed; and the support vector machine is trained by using the plurality of training samples to obtain the diagnosis model.

2. The hydroelectric generating unit bolt loosening monitoring method of claim 1, wherein, the working condition parameter vectors comprise a water head and a load.

3. The hydroelectric generating unit bolt loosening monitoring method of claim 1, wherein, the adaptive wavelet threshold denoising processing on the original signals to obtain the to-be-diagnosed acoustic vibration signals comprises the steps of: the original signals are decomposed by discrete wavelet transform to obtain high-frequency detail coefficients and low-frequency approximation coefficients of a plurality of scales; the high-frequency detail coefficients are processed by a soft threshold function to obtain processed detail coefficients; and the low-frequency approximation coefficients and the processed detail coefficients are reconstructed by inverse discrete wavelet transform to obtain the to-be-diagnosed acoustic vibration signals.

4. The hydroelectric generating unit bolt loosening monitoring method of claim 1, wherein, the loss function of the generator is: where s normal is the historical normal acoustic vibration signal, θ is the historical working condition parameter vector, z is the historical random noise vector, G represents the generator, D represents the discriminator, log is the natural logarithm operation, z~p z (z) represents that the historical random noise vector z obeys the prior distribution p z (z), and E represents the mathematical expectation; the loss function of the discriminator is: where s real is the true looseness signal.

5. The hydroelectric generating unit bolt loosening monitoring method of claim 1, wherein, the time domain features comprise a mean value, an effective value, a standard deviation, a kurtosis, a peak factor and an impulse factor, and the frequency domain features comprise a spectral centroid, a spectral bandwidth and a harmonic factor.

6. The hydroelectric generating unit bolt loosening monitoring method of claim 5, wherein, expressions of the mean value, the effective value, the standard deviation, the kurtosis, the peak factor and the impulse factor are: where μ is the mean, σ is the standard deviation, RMS is the effective value, CF is the peak factor, and IF is the impulse factor. x i is the sound and vibration signal to be diagnosed for the i th sampling point, N is the total number of sampling points, and max represents the maximum value function.

7. The hydroelectric generating unit bolt loosening monitoring method of claim 5, wherein, the frequency domain features are extracted from the to-be-diagnosed acoustic vibration signals, and the method comprises the steps of: fast Fourier transform is performed on the to-be-diagnosed acoustic vibration signals: wherein, represents a complex spectrum corresponding to the frequency index k, represents the instantaneous value of the frame of acoustic vibration signals to be diagnosed at the n+1th sampling point, j represents the imaginary unit, e represents the base of the natural logarithm, and N is the total number of sampling points. a power spectral density at a frequency index k is calculated as: furthermore, a spectral centroid, a spectral bandwidth and a harmonic factor are extracted, and expressions are: wherein, is the spectral centroid, is the spectral bandwidth, is the harmonic factor; M represents the number of effective frequency points, M = N / 2; f k represents the actual physical frequency corresponding to the frequency index k, , f s represents the sampling frequency; f base represents the fundamental frequency of the hydroelectric generating set, H represents the harmonic order, and P represents the power spectral density.

8. A bolt loosening monitoring device for a hydroelectric generating unit, characterized by, The device is used for implementing the bolt loosening monitoring method of the hydroelectric generating set as claimed in any one of claims 1-7. The acquisition processing unit is configured to arrange vibration sensors and ultrasonic sensors at key areas of the hydroelectric generating set, synchronously acquire vibration signals and ultrasonic signals, and record the acquired signals as original signals; and perform adaptive wavelet threshold denoising processing on the original signals to obtain to-be-diagnosed acoustic vibration signals. The construction unit is configured to construct a diagnosis model based on a conditional generative adversarial network and a support vector machine. The feature extraction and fusion unit is configured to extract time domain features and frequency domain features from the to-be-diagnosed acoustic vibration signals, acquire a working condition parameter vector of the hydroelectric generating set, splice the time domain features, the frequency domain features and the working condition parameter vector to obtain a fusion feature vector. The monitoring unit is configured to input the fusion feature vector into the diagnosis model, and monitor a loosening state of the bolt through the diagnosis model.

Images