Hydroelectric generating set bolt state dynamic monitoring method and system

Abstract

The present application relates to the technical field of hydroelectric generating set bolt monitoring, and particularly relates to a hydroelectric generating set bolt state dynamic monitoring method and system. The method comprises the following steps: performing resonance frequency prediction on the bolt, and performing adaptive resonance excitation to obtain a bolt resonance characteristic spectrum; performing bolt acoustic signal collection according to the bolt resonance characteristic spectrum to obtain a multi-channel original acoustic waveform set; performing working condition noise preprocessing on the multi-channel original acoustic waveform set to obtain an acoustic propagation time-domain-frequency domain matrix; performing multi-dimensional stress state mapping analysis according to the acoustic propagation time-domain-frequency domain matrix, and performing bolt loosening state recognition to obtain a bolt structure stress distribution atlas; and performing multi-source heterogeneous data fusion and edge computing according to the bolt structure stress distribution atlas to obtain a connection reliability index. The present application realizes high-precision, real-time and online monitoring of the bolt state through acoustic characteristic change analysis under a complex stress state.

Description

Dynamic monitoring method and system for the condition of bolts in hydropower units Technical Field

[0001] This invention relates to the field of bolt monitoring technology for hydropower units, and in particular to a method and system for dynamic monitoring of bolt status in hydropower units. Background Technology

[0002] Currently, the following technologies are mainly used for monitoring bolts in hydropower units:

[0003] Torque wrench: This is the most common method for bolt pre-tightening, indirectly controlling the pre-tightening force by controlling the applied torque. The relationship between torque and pre-tightening force is greatly affected by the coefficient of friction, which is influenced by various factors such as thread surface condition and lubrication, making it difficult to control accurately.

[0004] Tapping method: This method involves tapping the bolt and judging its tightness by the crispness of the sound. It is highly subjective, relies on experience, and is difficult to quantify.

[0005] Manual inspection: This involves visually inspecting and touching the bolts to check for looseness, breakage, or other defects. It is labor-intensive and inefficient.

[0006] Elongation measurement method: The preload is calculated by measuring the elongation of the bolt. This method requires specialized tools and is complex to operate.

[0007] Resistance strain gauge method: Strain gauges are attached to the bolts, and the preload is calculated by measuring the bolt strain. This method is complex to install and prone to damage.

[0008] Existing dynamic monitoring methods for the condition of bolts in hydropower units generally suffer from problems such as low accuracy, poor sensitivity, difficulty in quantification, and inability to monitor online. They cannot accurately reflect the actual preload and loosening state of the bolts, and therefore cannot meet the needs of safe operation of hydropower units.

[0009] The operating environment of hydropower units is extremely complex, with the following main interference factors: Rotating components such as turbines and generators generate strong mechanical vibrations, which are transmitted through the structure to the bolts, affecting monitoring signals. During operation, the temperature of hydropower units fluctuates significantly, impacting sensor performance and measurement results. Traditional methods (such as torque wrenches, impact methods, and resistance strain gauges) have poor anti-interference capabilities and are easily affected by the aforementioned environmental factors, leading to distorted measurement results or even malfunction.

[0010] In summary, existing technical methods suffer from insufficient measurement accuracy and difficulty in handling complex hydropower environments, problems that urgently need to be addressed. Summary of the Invention

[0011] Therefore, it is necessary to provide a method and system for dynamic monitoring of the condition of bolts in hydropower units to solve at least one of the above-mentioned technical problems.

[0012] To achieve the above objectives, a method for dynamic monitoring of the condition of bolts in a hydropower unit includes the following steps:

[0013] Step S1: Predict the resonant frequency of the bolt to obtain an initial excitation waveform library; perform adaptive resonant excitation on the bolt based on the initial excitation waveform library to obtain the bolt resonant characteristic spectrum.

[0014] Step S2: Acquire the bolt acoustic signal based on the bolt resonance characteristic spectrum diagram to obtain a multi-channel original acoustic waveform set; preprocess the multi-channel original acoustic waveform set for working condition noise to obtain the acoustic propagation time-frequency domain matrix;

[0015] Step S3: Extract bolt acoustic projectile parameters based on the acoustic propagation time-frequency domain matrix to obtain a segmented acoustic projectile basic model, a set of higher-order nonlinear coefficients, and a temperature-compensated acoustic projectile coefficient matrix; perform interface contact nonlinear modeling on the bolt based on the segmented acoustic projectile basic model to obtain an interface contact acoustic characteristic model; perform multidimensional stress state mapping analysis on the interface contact acoustic characteristic model to obtain a multidimensional stress acoustic mapping matrix; identify bolt loosening state based on the set of higher-order nonlinear coefficients, the temperature-compensated acoustic projectile coefficient matrix, and the multidimensional stress acoustic mapping matrix to obtain a bolt structural stress distribution map.

[0016] Step S4: Perform multi-source heterogeneous data fusion and edge calculation based on the stress distribution spectrum of the bolt structure to obtain the connection reliability index;

[0017] Step S5: Based on the connection reliability index, perform intelligent early warning of bolt status for hydropower unit bolts to obtain bolt reliability prediction results.

[0018] This invention achieves adaptive resonant excitation for bolts of different models and specifications by combining bolt characteristic parameter acquisition, finite element resonant frequency prediction, multiple excitation waveform design, excitation-response test loops, excitation parameter optimization, and FPGA-based hardware-accelerated resonant characteristic spectrum construction. This maximizes the signal-to-noise ratio of bolt vibration response, providing a high-quality signal source for subsequent high-precision acoustic signal acquisition and improving the sensitivity and reliability of the entire monitoring system. By combining optimized deployment of multi-channel receiver arrays, multi-dimensional acoustic signal acquisition, operating condition noise feature extraction and classification, signal wavelet decomposition and reconstruction, spatiotemporal synchronous signal integration, and time-domain-frequency domain dual-dimensional feature extraction, this invention achieves high signal-to-noise ratio, multi-dimensional, and spatiotemporally synchronous acquisition of bolt resonant response signals. This effectively suppresses noise interference under complex operating conditions of hydropower units, providing accurate and comprehensive acoustic data for subsequent mechanical characteristic analysis. By combining acoustic parameter extraction, interface contact nonlinear modeling, multidimensional stress state mapping analysis, nonlinear acoustic correction model construction, preload inversion calculation, bolt loosening state identification, three-dimensional reconstruction of mechanical parameters, and generation of structural stress distribution maps, accurate conversion from acoustic signals to key mechanical parameters such as bolt preload, loosening state, and stress distribution was achieved. This overcame the limitations of traditional linear models under complex working conditions, improved the accuracy and reliability of monitoring results, and provided a comprehensive mechanical basis for assessing the bolt connection status of hydropower units. Furthermore, by combining multi-source heterogeneous data acquisition, dynamic allocation of edge computing resources, spatiotemporal alignment and standardization of multi-domain data, feature dimensionality reduction and key parameter extraction, construction of a dynamic Bayesian network fusion model, construction of a multidimensional health assessment index system, and comprehensive calculation of the connection reliability index, real-time processing and fusion analysis of monitoring data at the edge were achieved. This reduced data transmission latency and the computational burden on the central server, improved the real-time performance and efficiency of monitoring, and generated a reliability index that comprehensively reflects the health status of bolt connections, providing a quantitative basis for operation and maintenance decisions. By combining historical data analysis, residual learning network model construction, micro-loosening feature identification, multi-level state classification, degradation trend time-series prediction model construction, multi-scenario life prediction simulation, and reliability prediction result output, this invention achieves intelligent diagnosis, early warning, and life prediction of bolt conditions. It can promptly detect micro-loosening of bolts and predict their future degradation trends, providing a scientific basis for predictive maintenance of hydropower units, reducing the risk of unplanned downtime, and improving equipment safety and economic efficiency. Therefore, this invention provides a dynamic monitoring method for the condition of bolts in hydropower units, addressing the shortcomings of traditional methods such as insufficient measurement accuracy and difficulty in handling complex hydropower environments. By constructing a nonlinear acoustoelastic correction model, considering higher-order acoustoelastic coefficients and temperature compensation factors, it can more accurately describe the changes in the acoustic characteristics of bolts under complex stress states. Through multi-dimensional stress state mapping analysis, considering the stress distribution of bolts under complex stress conditions, it can more comprehensively assess the stress state of bolts.By generating adaptive frequency ultrasonic waves for different types of bolts using an intelligent high-frequency high-voltage exciter, optimal penetration is achieved. This effectively overcomes various interferences in the complex environment of hydropower units, enabling high-precision, real-time, and online monitoring of bolt conditions.

[0019] Preferably, the bolt acoustic spring parameter extraction in step S3 includes:

[0020] Acoustic feature components are analyzed and identified from the acoustic propagation time-frequency matrix to obtain the acoustic feature parameter set;

[0021] Based on the acoustic characteristic parameter set, the acoustic elastic coefficient of the bolt material is obtained by performing acoustic elastic analysis on the material.

[0022] Based on the acoustic elastic coefficient of the bolt material, a basic acoustic elastic relationship model is constructed to obtain a segmented acoustic elastic basic model.

[0023] The higher-order nonlinear coefficients are calibrated based on the acoustoelastic coefficient of the bolt material to obtain a set of higher-order nonlinear coefficients.

[0024] Based on the segmented acoustic missile basic model and the set of higher-order nonlinear coefficients, the temperature-acoustic missile coupling correction factor is extracted to obtain the temperature-compensated acoustic missile coefficient matrix.

[0025] This invention effectively filters key features strongly correlated with bolt preload from the high-dimensional acoustic propagation time-frequency domain matrix by calculating the Pearson correlation coefficient and employing principal component analysis (PCA). Dimensionality reduction is then performed to eliminate redundant information, reduce the complexity of subsequent calculations, and improve the efficiency and accuracy of acoustoelastic parameter extraction. Combining material handbooks, databases, and ultrasonic pulse-echo measurements, the acoustoelastic coefficients of the bolt material at room temperature, such as Young's modulus, Poisson's ratio, and Lamé constant, are obtained. This provides an accurate material parameter basis for establishing a model of the relationship between sound velocity and stress, ensuring the accuracy of the subsequent acoustoelastic model. A piecewise linear model is used to describe the relationship between sound velocity and stress, and the acoustoelastic coefficients for each interval are determined through experimental calibration. Compared to traditional single linear models, this model more accurately reflects the changes in the acoustoelastic properties of bolts within different preload ranges, improving the model's accuracy and applicability. A third-order nonlinear acoustoelastic model was introduced, and experimental data were fitted using the least squares method to obtain higher-order nonlinear coefficients. This more accurately describes the higher-order nonlinear relationship between sound velocity and stress, further improving the accuracy of the acoustoelastic model, especially under high preload or with minor loosening. By conducting sound velocity measurements and nonlinear coefficient calibration experiments at different temperatures, and calculating the temperature coefficient of each acoustoelastic coefficient using linear regression, a temperature compensation model was established. This effectively eliminated the influence of temperature changes on the acoustoelastic parameters, improving the model's applicability and robustness under different ambient temperatures.

[0026] Preferably, the interface contact nonlinear modeling in step S3 includes:

[0027] Roughness fractal characterization of the bolt connection interface is performed to obtain a fractal feature description set of the contact surface;

[0028] Based on the fractal feature description set of the contact surface, the risk of contact area evolution under pressure is determined, and a set of pressure-contact area response curves is obtained.

[0029] Based on the pressure-contact area response curve set, the interface equivalent stiffness nonlinear mapping is performed to obtain the interface stiffness nonlinear equation set.

[0030] Based on the set of nonlinear equations for interface stiffness and the basic model of segmented acoustic elasticity, a discrete model of the spring-mass network is performed to obtain the interface acoustic propagation network model.

[0031] The ultrasonic transmission characteristics are calculated based on the interface acoustic propagation network model to obtain the interface transmittance spectrum characteristic set.

[0032] Phase jump and time delay effects were analyzed on the interface transmittance spectrum characteristic set to obtain the interface acoustic phase characteristic mapping table.

[0033] The interface acoustic phase characteristic mapping table and the interface acoustic propagation network model are integrated to obtain the interface contact acoustic characteristic model.

[0034] This invention utilizes a white-light interferometer to measure the three-dimensional morphology of the bolt connection interface and employs variable-scale fractal dimension calculation methods, structure function methods, and power spectral density methods to obtain fractal characteristic parameters such as the fractal dimension, characteristic length, and root-mean-square roughness of the contact surface. This achieves a quantitative description of the roughness of the bolt connection interface, providing key parameters for establishing a more realistic contact model. Using a fractal theory-based contact model (such as the Majumdar-Bhushan model), the change in the actual contact area under different pressures was calculated, yielding pressure-contact area response curves. Compared to the traditional Hertz contact theory, this model better reflects the contact characteristics of a truly rough surface, providing more accurate predictions and a reliable basis for subsequent calculations of interface stiffness. By calculating the ratio of the normal load increment to the normal displacement increment and fitting it using an exponential function, a nonlinear function of the normal equivalent stiffness of the contact interface as a function of pressure was obtained. This function accurately describes the nonlinear characteristics of the contact interface stiffness as a function of preload, providing key parameters for establishing a sound propagation model. The finite difference method was used to simulate the propagation of ultrasonic waves in a spring-mass network model, and the transmission coefficients at different frequencies were calculated, resulting in an interface transmittance spectrum characteristic set. This spectrum characteristic reflects the energy attenuation and frequency selectivity of ultrasonic waves passing through the contact interface, providing important information for subsequent analysis of bolt loosening status. By calculating the phase spectrum of the transmission coefficients, and performing unwrapping processing and calculating group time delay, an interface acoustic-phasic characteristic mapping table was obtained. This mapping table reflects the phase change and time delay that occur when ultrasonic waves pass through the contact interface, providing supplementary information for a more comprehensive description of the acoustic characteristics of the contact interface. By introducing the phase delay factor and time delay factor into the spring-mass network model, a comprehensive model that can simultaneously reflect the stiffness nonlinearity, transmission characteristics, and acoustic-phasic characteristics of the contact interface was established. This model can more accurately simulate the interaction between ultrasonic waves and the bolt connection interface, providing a more reliable theoretical basis for subsequent preload inversion and loosening status identification. Attached Figure Description

[0035] Figure 1 is a flowchart illustrating the steps of a method for dynamic monitoring of bolt status in a hydropower unit.

[0036] Figure 2 is a detailed flowchart of step S2 in this invention.

[0037] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0038] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0039] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0040] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0041] Referring to Figure 1, which is a flowchart illustrating the steps of the dynamic monitoring method for the condition of bolts in a hydropower unit according to the present invention, the method includes the following steps:

[0042] Step S1: Predict the resonant frequency of the bolt to obtain an initial excitation waveform library; perform adaptive resonant excitation on the bolt based on the initial excitation waveform library to obtain the bolt resonant characteristic spectrum.

[0043] In this embodiment of the invention, firstly, the geometric parameters and material type of the bolt to be tested are obtained using a coordinate measuring machine and an eddy current flaw detector, and a bolt characteristic parameter set is constructed. Then, based on the finite element analysis method (ANSYS) and combined with the material mechanical parameters, the theoretical resonant frequency of the bolt is calculated, forming a theoretical resonant frequency mapping table. Next, using an arbitrary waveform generator, various excitation waveforms such as swept-frequency sine waves, linearly modulated chirp signals, and Gaussian-modulated sine pulses are designed according to the theoretical resonant frequency, forming an initial excitation waveform library. Subsequently, a piezoelectric ceramic transducer is driven by a power amplifier to sequentially apply the waveforms in the initial excitation waveform library, and the vibration response of the bolt surface is measured using a laser vibrometer to establish an excitation-response relationship dataset. Then, the response dataset is analyzed, the signal-to-noise ratio is calculated, and orthogonal experimental design and a genetic algorithm are used to optimize the excitation parameters to obtain the optimal excitation parameter configuration. Finally, based on ARM... The control system of the Cortex-M4 microcontroller controls an arbitrary waveform generator chip via the SPI interface to generate excitation signals of specified frequency, waveform, and amplitude, which drive the piezoelectric ceramic transducer to generate ultrasonic waves. A laser vibrometer is used to monitor the vibration response of the bolt surface in real time. An FPGA-based hardware acceleration module is used for signal processing, including sliding window processing and windowing. The FFT algorithm is used to calculate the spectrum of each window of data, calculate the amplitude spectrum and phase spectrum of the spectrum, and select the peak value in the amplitude spectrum that is higher than a set threshold as the resonant frequency to obtain the bolt resonance characteristic spectrum.

[0044] Step S2: Acquire the bolt acoustic signal based on the bolt resonance characteristic spectrum diagram to obtain a multi-channel original acoustic waveform set; preprocess the multi-channel original acoustic waveform set for working condition noise to obtain the acoustic propagation time-frequency domain matrix;

[0045] In this embodiment of the invention, the propagation of ultrasonic waves in the bolt and surrounding structure is simulated using sound field simulation software (COMSOL) to determine candidate areas for receiver placement. A genetic algorithm is then used to optimize the receiver array layout (8 piezoelectric ceramic ultrasonic receivers) to obtain an optimized spatial distribution scheme for the receivers. Next, according to the optimized scheme, the piezoelectric ceramic ultrasonic receivers are fixed to the bolt and surrounding structure using magnetic chucks. Raw acoustic signals are acquired using an 8-channel synchronous data acquisition card to form a multi-channel raw acoustic waveform set. Then, under normal operating conditions of the hydroelectric generator, background noise signals are acquired, short-time Fourier transform (STFT) is performed, power spectral density (PSD) is calculated, and K-means clustering algorithm is used. Noise classification was performed, and a noise characteristic map of the monitored environment was plotted. Subsequently, Discrete Wavelet Transform (DWT) was performed using the Symlet8 wavelet basis function, and a hard threshold denoising method was used to reconstruct the signal, obtaining an acoustic feature-enhanced waveform group. Then, the signal received by the first receiver was selected as the reference signal, and the cross-correlation function between the signals received by the other 7 receivers and the reference signal was calculated. The signals were shifted according to the time delay to align all signals in time, resulting in an acoustic propagation synchronization dataset. Finally, time-domain analysis (calculating 13 time-domain features such as peak value and root mean square value) and frequency-domain transformation (FFT, calculating 10 frequency-domain features such as centroid frequency) were performed on the synchronization dataset to obtain the acoustic propagation time-domain-frequency-domain matrix.

[0046] Step S3: Extract bolt acoustic projectile parameters based on the acoustic propagation time-frequency domain matrix to obtain a segmented acoustic projectile basic model, a set of higher-order nonlinear coefficients, and a temperature-compensated acoustic projectile coefficient matrix; perform interface contact nonlinear modeling on the bolt based on the segmented acoustic projectile basic model to obtain an interface contact acoustic characteristic model; perform multidimensional stress state mapping analysis on the interface contact acoustic characteristic model to obtain a multidimensional stress acoustic mapping matrix; identify bolt loosening state based on the set of higher-order nonlinear coefficients, the temperature-compensated acoustic projectile coefficient matrix, and the multidimensional stress acoustic mapping matrix to obtain a bolt structural stress distribution map.

[0047] In this embodiment of the invention, acoustic feature parameters are extracted from the acoustic propagation time-frequency matrix. Features with a cross-correlation greater than 0.6 are selected using the Pearson correlation coefficient, and key acoustic feature parameter sets are obtained through PCA dimensionality reduction. Subsequently, material databases are consulted, and the longitudinal and transverse wave velocities of the bolts are measured using the ultrasonic pulse-echo method. The Lamé constant is calculated, and a piecewise linear relationship model between sound velocity and stress is constructed. The preload range (0-200 MPa) is divided into five intervals, and the acoustoelastic coefficient is experimentally calibrated within each interval. A third-order nonlinear acoustoelastic model (CL = CL0 + k1σ + k2σ) is employed. 2 +k3σ 3The experimental data were fitted using the least squares method to obtain a set of high-order nonlinear coefficients. A temperature compensation coefficient matrix was established through constant temperature chamber experiments (-20℃ to 60℃, 10℃ intervals). Next, the three-dimensional morphology of the bolt contact interface was measured using a white light interferometer (0.1nm vertical resolution), and the fractal dimension D, characteristic length Lc, and root mean square roughness Rq were calculated. Based on the Majumdar-Bhushan model, the evolution of the actual contact area under different pressures (0-200MPa) was simulated, deriving the nonlinear equation for the interface equivalent stiffness: Kn(P) = a(1-exp(-bP)). The bolt and connector were discretized into a spring-mass network, and the ultrasonic propagation characteristics were calculated using the finite difference method to obtain the interface transmittance spectrum characteristics and phase delay mapping table, forming a complete interface contact acoustic characteristic model. Subsequently, the multiaxial stress state was calculated using the finite element method. Based on the Christoffel equation and the acoustoelastic tensor theory, a sound velocity-stress tensor mapping function set was constructed, and the transmission path was analyzed using the stress streamline method to generate a multidimensional stress acoustic mapping matrix. Finally, based on the isolated forest algorithm to detect abnormal ultrasonic propagation characteristics, combined with the analysis of acoustic impedance changes at the contact surface, the feature weights were determined by the analytic hierarchy process (AHP) to construct a multi-dimensional loosening index system. The K-means clustering algorithm was used to establish a four-level loosening state classification standard. The SVM classifier was used to judge the bolt loosening state in real time. Based on the stress distribution in the key areas of the bolt (thread root and head transition area), a multi-level, interactive structural stress distribution map was generated to provide early warning reference.

[0048] Step S4: Perform multi-source heterogeneous data fusion and edge calculation based on the stress distribution spectrum of the bolt structure to obtain the connection reliability index;

[0049] In this embodiment of the invention, the edge computing device receives stress distribution maps of bolt structures, vibration data from hydropower unit operation, and historical monitoring records, forming a multi-source monitoring data stream. Based on the characteristics of the data stream, an embedded platform based on NVIDIA Jetson TX2 is used to dynamically allocate computing resources (GPU for image processing, ARM Cortex-A57 CPU for vibration data processing, and Denver CPU for data processing). The system employs a 2-CPU process to process historical data, resulting in an intelligent computing scheme for monitoring. It preprocesses multi-source data (image enhancement, time-frequency analysis, data standardization, etc.) and performs spatiotemporal alignment to obtain a monitoring aligned dataset. The mutual information method is used to evaluate the correlation between each feature and the bolt health status. Features with mutual information greater than a set threshold are selected as candidate features. Principal component analysis (PCA) is used to reduce the dimensionality of the candidate features, yielding a key state feature vector. A dynamic Bayesian network model is constructed, using a conditional probability table (CPT) to describe the probabilistic relationships between nodes. A structural learning algorithm (such as the K2 algorithm) is used to learn the network's topology, and a parameter learning algorithm (such as maximum likelihood estimation) is used to learn the network's CPT parameters, resulting in a state fusion probability map. A Bayesian inference algorithm is used to calculate the probability of bolts being in different health states, constructing a multi-dimensional health assessment index system. A weighted average method is used to calculate the connection reliability index.

[0050] Step S5: Based on the connection reliability index, perform intelligent early warning of bolt status for hydropower unit bolts to obtain bolt reliability prediction results;

[0051] In this embodiment of the invention, historical bolt monitoring data and historical connection reliability indices are read from the database, preprocessed, and divided into training, validation, and test sets. A one-dimensional convolutional residual network model is designed, containing multiple residual blocks (each residual block contains two convolutional layers and one skip connection). The network is trained using the Adam optimizer to obtain a deep model for bolt micro-variation recognition. Real-time collected data is input into the deep model to calculate the rate of change of predicted probabilities, extract micro-loosening features, and obtain an early anomaly feature map. Combining the output of the deep model and the early anomaly feature map, a more refined state classification is performed (adding two levels: "early slight loosening" and "early moderate loosening"). The classification results are used as the refined health status classification results and stored in real time in JSON format. Based on the refined health status classification... Based on the detailed classification results, a Long Short-Term Memory (LSTM) network model (containing one LSTM layer and one fully connected layer) is constructed to predict the degradation trend of bolts. Mean Squared Error (MSE) is used as the loss function, and the network is trained using the Adam optimizer to obtain the bolt degradation trend prediction model. Based on this model, considering different load conditions, ambient temperature, vibration intensity, and other factors, lifespan prediction simulations are performed under various working conditions. The detailed classification results of the health status are converted into numerical representations. When the predicted health status value reaches a set threshold, the bolt is considered to have failed. The failure time of the bolt under different working conditions is recorded to obtain the lifespan distribution, represented in the form of probability density function (PDF) and cumulative distribution function (CDF), yielding the bolt reliability prediction results, which are stored in JSON format. The prediction results are then transformed into a visualized time-reliability curve, marking key warning points and maintenance time windows to provide decision-making references.

[0052] Preferably, step S1 includes the following steps:

[0053] Step S11: Collect the characteristic parameters of the bolt to obtain the bolt characteristic parameter set;

[0054] Step S12: Perform resonant frequency prediction analysis based on the bolt characteristic parameter set to obtain the theoretical resonant frequency mapping table;

[0055] Step S13: Design high-frequency excitation waveforms based on the theoretical resonant frequency mapping table to obtain the initial excitation waveform library;

[0056] Step S14: Perform an excitation-response test loop on the bolt according to the initial excitation waveform library to obtain the response characteristic dataset;

[0057] Step S15: Set the excitation parameters according to the response characteristic dataset to obtain the excitation parameter configuration;

[0058] Step S16: Based on the excitation parameter configuration, drive the high-frequency high-voltage exciter to perform ultrasonic excitation on the bolt to obtain a real-time resonant response data stream;

[0059] Step S17: Construct the resonant characteristic spectrum of the real-time resonant response data stream to obtain the bolt resonant characteristic spectrum diagram.

[0060] In this embodiment of the invention, a coordinate measuring machine (CMM) is used to perform non-contact scanning of the bolts to be tested in a hydroelectric generator. The laser probe of the CMM moves along the axial and radial directions of the bolt with an accuracy of 0.01 mm to acquire point cloud data of the bolt's outline. Simultaneously, an eddy current flaw detector is used to perform non-destructive testing on the bolt material at a frequency of 50 kHz to acquire the electrical conductivity and magnetic permeability data of the bolt material. The point cloud data is imported into reverse engineering software, and a three-dimensional geometric model of the bolt is reconstructed through a surface fitting algorithm to accurately measure the bolt's diameter, length, thread angle, pitch, and other geometric parameters. The electrical conductivity and magnetic permeability data are compared with a material database to determine the bolt's material type (e.g., 40CrNiMoA). The geometric parameters and material type data are integrated to form a bolt characteristic parameter set, which is stored in a database in XML format. This database is embedded in an edge computing device.

[0061] Based on the bolt characteristic parameter set, the resonant frequency is predicted using the finite element method (FEM). First, the three-dimensional geometric model of the bolt is imported into the finite element analysis software ANSYS Mechanical. Then, according to the bolt material type, the corresponding mechanical parameters such as elastic modulus, Poisson's ratio, and density are selected from the material library. In the finite element model, one end of the bolt is set as a fixed constraint to simulate the boundary conditions of the bolt under actual working conditions. The Block Lanczos modal extraction algorithm is used to calculate the first 10 natural frequencies and mode shapes of the bolt. For each natural frequency, the proportions of longitudinal vibration, bending vibration, and torsional vibration components are calculated. The three frequencies with the highest proportion of longitudinal vibration are selected as candidate theoretical resonant frequencies of the bolt. The candidate theoretical resonant frequencies of bolts of different models and specifications are compiled into a table, forming a theoretical resonant frequency mapping table, stored in a CSV file format, and stored in the local FPGA memory.

[0062] Based on the theoretical resonant frequency mapping table, a set of high-frequency excitation waveforms was designed. An arbitrary waveform generator was used to generate the excitation signals. For each candidate theoretical resonant frequency, a swept sine wave signal, a linearly modulated chirp signal, and a Gaussian-modulated sine pulse signal were generated, each with the same center frequency and a bandwidth of 10% of the center frequency. The start and end frequencies of the swept signal were set to cover ±5% of the theoretical resonant frequency. The start and end frequencies of the chirp signal were set to be the same as those of the swept signal, and the pulse width was set to 1 ms. The center frequency of the Gaussian-modulated sine pulse signal was set to be the same as the theoretical resonant frequency, the pulse width was set to 0.5 ms, and the repetition period was 10 ms. The generated excitation waveform signals of different frequencies and types were stored in binary format to form an initial excitation waveform library, which was stored in DDR4 memory.

[0063] The high-frequency, high-voltage exciter probe is brought into close contact with the bolt under test, and an ultrasonic coupling agent is applied to the contact surface to improve energy transmission efficiency. Excitation waveforms are selected sequentially from the initial excitation waveform library, amplified by a power amplifier, and used to drive a piezoelectric ceramic transducer to generate ultrasonic waves, thus exciting the bolt vibration. Simultaneously, a laser vibrometer is used to measure the vibration velocity on the bolt surface at a sampling rate of 1 MHz. The laser vibrometer spot diameter is set to 0.1 mm, scanning five different locations on the bolt surface to acquire multi-point vibration response signals. The steady-state vibration response signal corresponding to each excitation waveform is recorded for 1 second. The vibration velocity signal is then subjected to a Fast Fourier Transform (FFT) to obtain the frequency domain response spectrum. The time-domain signal and frequency domain response spectrum data are packaged to form a response characteristic dataset, which is stored on an SSD in HDF5 format.

[0064] Analyze the response characteristic dataset to evaluate the bolt vibration response under different excitation waveforms. Calculate the signal-to-noise ratio (SNR) at the peak resonant frequency in each frequency domain response spectrum. Select the excitation waveform with the highest SNR as the optimal excitation waveform type for that bolt. For the optimal excitation waveform type, further analyze the influence of its parameters (such as the start and end frequencies of the sweep signal, the pulse width of the chirp signal, etc.) on the SNR. Use an orthogonal experimental design method, selecting 3-5 key parameters, each with 3 levels, to conduct parameter optimization experiments. Based on the experimental results, establish a regression model between SNR and excitation parameters. Use a genetic algorithm to optimize the excitation parameters with the goal of maximizing SNR. Combine the optimal excitation waveform type and the corresponding optimized parameters to form the excitation parameter configuration, store it as a JSON file in the Flash memory of the embedded system.

[0065] The output parameters of the high-frequency high-voltage exciter are set according to the excitation parameter configuration. The high-frequency high-voltage exciter employs a control system based on an ARM Cortex-M4 microcontroller. The microcontroller reads the excitation parameter configuration from Flash memory and controls an arbitrary waveform generator chip (such as AD9959) via an SPI interface to generate an excitation signal with a specified frequency, waveform, and amplitude. The excitation signal is amplified by a power amplifier (using an H-bridge circuit built with MOSFETs) and then drives a piezoelectric ceramic transducer to generate ultrasonic waves. A laser vibrometer is used to monitor the vibration response of the bolt surface in real time, with a sampling rate of 2MHz. The acquired vibration signal is transmitted to an edge computing device via a USB 3.0 interface, forming a real-time resonant response data stream. The data stream is transmitted using the UDP protocol.

[0066] The real-time resonant response data stream is processed. In the edge computing device, an FPGA-based hardware acceleration module is used for signal processing. First, a sliding window is applied to the data stream, with a window length of 1024 sampling points and an overlap rate of 50%. Windowing (Hanning window) is applied to the data within each window to reduce spectral leakage. Then, a radix-2 FFT algorithm is used to calculate the spectrum of the data in each window. The amplitude and phase spectra are calculated. The amplitude spectrum is smoothed (using a Savitzky-Golay filter, order 3, window length 11) to suppress noise. Peak values ​​in the amplitude spectrum above a set threshold (set to 10% of the maximum amplitude) are selected as resonant frequencies. The amplitude and phase corresponding to each resonant frequency are recorded. The resonant frequency, amplitude, and phase information are plotted into a three-dimensional graph, i.e., a bolt resonant characteristic spectrum. The horizontal axis represents frequency, the vertical axis represents amplitude, and color represents phase. The spectrum data is stored in PNG format and displayed in real-time on the monitoring interface.

[0067] As an example of the present invention, referring to FIG2, step S2 in this example includes:

[0068] Step S21: Optimize the deployment of the multi-channel receiver array based on the bolt resonance characteristic spectrum diagram to obtain the receiver spatial distribution optimization scheme;

[0069] In this embodiment of the invention, ultrasonic wave propagation is simulated using the sound field simulation software COMSOLMultiphysics based on the bolt resonance characteristic spectrum. A 3D CAD model of the bolt and surrounding structures (such as flanges, gaskets, etc.) is imported, and material properties (density, Young's modulus, Poisson's ratio) are set. The frequency and location of the excitation source are set according to the resonance characteristic spectrum. The excitation source model is a point source located at the center of the bolt head. The propagation sound field distribution of ultrasonic waves in the bolt and surrounding structures is calculated. Virtual receiving points are set on the bolt surface and surrounding space, with a spacing of 1 mm. The sound pressure level data of each virtual receiving point is extracted. Regions with sound pressure levels greater than -6 dB are selected as candidate regions for receiver placement. Within the candidate regions, a genetic algorithm is used to optimize the receiver array layout. The number of receivers is set to 8, and the array form is a circular array. The optimization objective function is set to maximize the average signal-to-noise ratio of the received signal, with the constraint that the receiver spacing is greater than 10 mm to avoid mutual interference. Based on the optimization results, the specific location coordinates of the 8 ultrasonic receivers are determined, forming a receiver spatial distribution optimization scheme, which is saved in TXT text format.

[0070] Step S22: Based on the receiver spatial distribution optimization scheme, perform multi-dimensional acoustic signal acquisition on the bolt to obtain a multi-channel original acoustic waveform set;

[0071] In this embodiment of the invention, according to the receiver spatial distribution optimization scheme, eight piezoelectric ceramic ultrasonic receivers (center frequency of 2MHz, bandwidth of 500kHz) are fixed to bolts and surrounding structures using magnetic chucks. Ultrasonic coupling agent is applied between the receivers and the bolt surfaces. Each receiver is connected to an independent preamplifier (gain of 40dB, bandwidth of 100kHz-5MHz). The output signals of the eight preamplifiers are connected to an 8-channel synchronous data acquisition card (sampling rate of 20MHz, resolution of 16 bits). The acquisition card is connected to an edge computing device via a PCIe interface. In the edge computing device, a data acquisition program is started to synchronously acquire ultrasonic signals from eight channels. The acquisition duration is set to 1 second, the trigger mode is set to rising edge trigger, and the trigger level is set to 0.1V. The acquired raw acoustic signal data from the eight channels are stored in binary format to form a multi-channel raw acoustic waveform set in TDMS file format.

[0072] Step S23: Extract and classify the operating condition noise features from the multi-channel raw acoustic waveform set to obtain the monitoring environment noise feature map;

[0073] In this embodiment of the invention, a multi-channel raw acoustic waveform set is analyzed. Under normal operating conditions of the hydropower unit, a 10-second background noise signal is acquired. A short-time Fourier transform (STFT) is performed on the noise signal of each channel, using a Hanning window with a window length of 1024 sampling points and an overlap rate of 75%. The power spectral density (PSD) at each time-frequency point is calculated. The probability density distribution of the PSD at all time-frequency points is statistically analyzed. Peak values ​​in the PSD probability density distribution are identified, corresponding to different types of noise sources. For example, a peak near 100Hz corresponds to harmonics of the turbine's rotation frequency, and a peak near 50Hz corresponds to power grid frequency interference. The K-means clustering algorithm is used to classify noise components of different frequencies. Based on the classification results, a noise feature map of the monitored environment is plotted. The horizontal axis of the map represents frequency, the vertical axis represents power spectral density, and the color represents the probability density of different noise types. The noise feature map is stored in JPG format.

[0074] Step S24: Perform wavelet decomposition and reconstruction of the original acoustic waveform set of the multi-channel according to the monitoring environmental noise feature map to obtain the acoustic feature enhanced waveform group;

[0075] In this embodiment of the invention, wavelet denoising is performed on the multi-channel original acoustic waveform set based on the monitored environmental noise characteristic spectrum. The Symlet8 wavelet basis function is selected for Discrete Wavelet Transform (DWT). The number of decomposition layers is set to 5. For each layer of wavelet coefficients, a hard threshold denoising method is used. The threshold is set to 3σ, where σ is the standard deviation of the wavelet coefficients in that layer, estimated using the Median Absolute Deviation (MAD) method. Wavelet coefficients above the threshold are retained, while those below the threshold are set to zero. The processed wavelet coefficients are then subjected to Inverse Discrete Wavelet Transform (IDWT) to reconstruct the signal. Wavelet denoising is performed on the signals of each of the 8 channels, resulting in 8 sets of denoised signals, i.e., acoustic feature-enhanced waveform sets. The waveform sets are stored in MAT format.

[0076] Step S25: Perform spatiotemporal synchronization integration of the acoustic feature enhancement waveform group to obtain an acoustic propagation synchronization dataset;

[0077] In this embodiment of the invention, the acoustic feature enhancement waveform group is synchronized. The signal received by the first receiver is selected as the reference signal. The cross-correlation function between the signals received by the other 7 receivers and the reference signal is calculated. The peak position of the cross-correlation function corresponds to the time delay between the two signals. The signals are shifted according to the time delay to align all signals in time. Since the spatial position of the receiver is known, spatial position information can be obtained through spatial geometry and time delay. The time-aligned signal data of the 8 channels are merged into a matrix. The number of rows in the matrix is ​​the number of sampling points, and the number of columns is the number of channels (8). This matrix is ​​used as the acoustic propagation synchronization dataset, and the data is stored in HDF5 format.

[0078] Step S26: Extract time-domain and frequency-domain dual-dimensional features from the acoustic propagation synchronization dataset to obtain the acoustic propagation time-domain and frequency-domain matrix;

[0079] In this embodiment of the invention, feature extraction is performed on the acoustic propagation synchronization dataset. In the time domain, 13 time-domain features are calculated for each channel signal, including peak value, root mean square value, kurtosis, skewness, waveform factor, impulse factor, and margin factor. In the frequency domain, a Fast Fourier Transform (FFT) is performed on each channel signal to obtain the spectrum. Ten frequency-domain features are calculated, including the centroid frequency, root mean square frequency, frequency variance, and frequency standard deviation. The time-domain and frequency-domain features of each channel are combined into a feature vector (length 23). The feature vectors of the eight channels are concatenated column-wise to form a 23x8 matrix. This matrix is ​​used as the acoustic propagation time-domain-frequency domain matrix and stored in CSV format.

[0080] Preferably, the bolt acoustic spring parameter extraction in step S3 includes:

[0081] Acoustic feature components are analyzed and identified from the acoustic propagation time-frequency matrix to obtain the acoustic feature parameter set;

[0082] Based on the acoustic characteristic parameter set, the acoustic elastic coefficient of the bolt material is obtained by performing acoustic elastic analysis on the material.

[0083] Based on the acoustic elastic coefficient of the bolt material, a basic acoustic elastic relationship model is constructed to obtain a segmented acoustic elastic basic model.

[0084] The higher-order nonlinear coefficients are calibrated based on the acoustoelastic coefficient of the bolt material to obtain a set of higher-order nonlinear coefficients.

[0085] Based on the segmented acoustic missile basic model and the set of higher-order nonlinear coefficients, the temperature-acoustic missile coupling correction factor is extracted to obtain the temperature-compensated acoustic missile coefficient matrix.

[0086] In this embodiment of the invention, feature selection and dimensionality reduction are performed based on the acoustic propagation time-frequency domain matrix. First, the Pearson correlation coefficient between each feature in the matrix and the bolt preload is calculated. Features with an absolute correlation coefficient greater than 0.6 are selected as candidate features. Then, Principal Component Analysis (PCA) is used to reduce the dimensionality of the candidate features. The covariance matrix of the candidate features is calculated, and eigenvalue decomposition is performed. Principal components with a cumulative variance contribution rate greater than 95% are selected. The original features are projected onto the principal component space to obtain the dimensionality-reduced feature vectors. These principal components mainly include: ultrasonic propagation time (TOF), attenuation coefficient, dominant frequency, and bandwidth. The dimensionality-reduced feature vectors are combined into an acoustic feature parameter set and stored in XML format.

[0087] Based on the bolt material type (e.g., 40CrNiMoA), consult material handbooks and databases to obtain the Young's modulus (E), Poisson's ratio (ν), and density (ρ) of the material at room temperature. The longitudinal wave velocity (CL) and transverse wave velocity (CT) of the bolt material are measured using the ultrasonic pulse-echo method. Specifically, an ultrasonic transducer (center frequency 5MHz) is placed vertically on the bolt end face, and an ultrasonic pulse is emitted. The time (t) for the ultrasonic pulse to travel one round trip inside the bolt is measured using an oscilloscope. Given the bolt length (L), the longitudinal wave velocity CL = 2L / t. The transverse wave transducer is replaced, and the above process is repeated to measure the transverse wave velocity CT. Based on elasticity theory, the Lamé constants (λ and μ) are calculated: λ = ρ(CL). 2 -2CT 2 ), μ=ρCT 2 Young's modulus (E), Poisson's ratio (ν), and Lamé constants (λ and μ) are used as the acoustic elastic coefficients of the bolt material and stored in JSON format.

[0088] Based on the acoustoelastic coefficient of the bolt material obtained in the previous step, a model relating sound velocity and stress is established. Since the acoustoelastic coefficient of the bolt varies under different stress states, a piecewise linear model is used to describe this nonlinear relationship. The bolt preload range (e.g., 0-200 MPa) is divided into five intervals, each with a length of 40 MPa. Within each interval, a linear relationship between sound velocity and stress is assumed: CL = CL0 + kσ, where CL0 is the longitudinal wave velocity under zero stress, k is the acoustoelastic coefficient, and σ is the stress. The acoustoelastic coefficient k for each interval is determined through experimental calibration. Specifically, a hydraulic servo testing machine is used to apply different axial tensile forces to the bolt, while simultaneously measuring the longitudinal wave velocity under different tensile forces. The experimental data are linearly fitted to obtain the acoustoelastic coefficient k for each interval. The linear models of the five intervals are combined into a piecewise acoustoelastic basic model, stored as a piecewise function in TXT format.

[0089] To more accurately describe the nonlinear relationship between sound velocity and stress, higher-order nonlinear coefficients are introduced. A third-order nonlinear acoustoelastic model is adopted: CL = CL0 + k1σ + k2σ 2 +k3σ 3 Here, k1, k2, and k3 are higher-order nonlinear coefficients. The experimental data are nonlinearly fitted using the least squares method to obtain these higher-order nonlinear coefficients. To verify the model's accuracy, the fitting results are compared with the experimental data, and the root mean square error (RMSE) is calculated. If the RMSE exceeds a set threshold (e.g., 0.5 m / s), the order of the nonlinear model is increased, and the fitting is repeated until the RMSE meets the requirement. The final determined higher-order nonlinear coefficients k1, k2, k3… are grouped into a higher-order nonlinear coefficient set and stored in TXT text format.

[0090] Considering the influence of temperature on the acoustoelastic coefficient, a temperature compensation model was established. Bolts were placed in a constant temperature chamber with a temperature range of -20℃ to 60℃, in 10℃ intervals. The sound velocity measurement and nonlinear coefficient calibration experiments were repeated at each temperature. The piecewise acoustoelastic basic model and high-order nonlinear coefficient set were obtained at different temperatures. The variation of the acoustoelastic coefficient with temperature was analyzed. It was found that the acoustoelastic coefficient and temperature have an approximately linear relationship: k(T) = k0 + αT, where k0 is the acoustoelastic coefficient at the reference temperature, α is the temperature coefficient, and T is the temperature. The temperature coefficient α for each acoustoelastic coefficient was calculated using linear regression. All temperature coefficients α were combined into a matrix, i.e., the temperature-compensated acoustoelastic coefficient matrix. The rows of the matrix correspond to different acoustoelastic coefficients, and the columns correspond to different temperature ranges, stored in CSV file format.

[0091] Preferably, the interface contact nonlinear modeling in step S3 includes:

[0092] Roughness fractal characterization of the bolt connection interface is performed to obtain a fractal feature description set of the contact surface;

[0093] Based on the fractal feature description set of the contact surface, the risk of contact area evolution under pressure is determined, and a set of pressure-contact area response curves is obtained.

[0094] Based on the pressure-contact area response curve set, the interface equivalent stiffness nonlinear mapping is performed to obtain the interface stiffness nonlinear equation set.

[0095] Based on the set of nonlinear equations for interface stiffness and the basic model of segmented acoustic elasticity, a discrete model of the spring-mass network is performed to obtain the interface acoustic propagation network model.

[0096] The ultrasonic transmission characteristics are calculated based on the interface acoustic propagation network model to obtain the interface transmittance spectrum characteristic set.

[0097] Phase jump and time delay effects were analyzed on the interface transmittance spectrum characteristic set to obtain the interface acoustic phase characteristic mapping table.

[0098] The interface acoustic phase characteristic mapping table and the interface acoustic propagation network model are integrated to obtain the interface contact acoustic characteristic model.

[0099] In this embodiment of the invention, a white light interferometer is used to measure the three-dimensional topography of the bolt connection interface (threaded surface and the surface of the connected parts). The white light interferometer has a vertical resolution of 0.1 nm and a horizontal resolution of 1 μm. The measurement area is 5 mm × 5 mm, containing at least 10 thread teeth. The obtained three-dimensional topography data is preprocessed to remove tilt and curvature. The fractal dimension D of the rough surface (between 2 and 3) is calculated using a variable-scale fractal dimension calculation method (such as box counting). The characteristic length Lc of the rough surface is calculated using the structure function method. Lc is defined as the minimum lateral distance when the structure function value reaches saturation. The root mean square roughness Rq of the rough surface is calculated using the power spectral density (PSD) method. The fractal dimension D, characteristic length Lc, and root mean square roughness Rq are used as a fractal feature description set of the contact surface and stored in TXT text format.

[0100] Based on the fractal feature description set obtained in the previous step, a fractal-based contact model (such as the Majumdar-Bhushan model) is used to calculate the actual contact area of ​​the contact surface under different pressures. This model treats the rough surface as a collection of micro-protrusions of different sizes, and the deformation of each micro-protrusion follows Hertz contact theory. The Young's modulus and Poisson's ratio of the bolt material and the connected parts material are input. The normal load range is set to 0-200 MPa, with an increment of 1 MPa. The actual contact area A under each normal load is calculated. The normal load and the corresponding actual contact area are plotted as a curve to obtain the pressure-contact area response curve. For different combinations of roughness parameters (D, Lc, Rq), the above calculation is repeated to obtain a set of pressure-contact area response curves, which are stored in MAT format.

[0101] Based on the pressure-contact area response curves obtained in the previous step, the normal equivalent stiffness of the contact interface is calculated. The normal equivalent stiffness is defined as the ratio of the normal load increment to the normal displacement increment: Kn = ΔP / Δδ, where ΔP is the normal load increment and Δδ is the normal displacement increment. The normal displacement increment Δδ can be calculated from the change in the actual contact area A. The normal equivalent stiffness Kn at each pressure point is calculated using a numerical differential method. The relationship between the normal equivalent stiffness Kn and the normal pressure P is fitted to obtain the nonlinear stiffness function Kn(P). An exponential function is used for fitting: Kn(P) = a(1-exp(-bP)), where a and b are fitting parameters. The pressure-contact area response curves corresponding to different roughness parameters are fitted separately to obtain a set of nonlinear stiffness functions, i.e., the interface stiffness nonlinear equation set, stored in MAT format.

[0102] The bolt and the connected components are discretized into a series of point masses, connected by springs. The stiffness of the springs is determined by the nonlinear equation set of interface stiffness obtained in the previous step. Inside the bolt and the connected components, the spring stiffness between the point masses is determined by the elastic modulus and Poisson's ratio of the material. At the contact interface, the spring stiffness between the point masses is determined by the normal equivalent stiffness Kn(P). The value of Kn(P) is dynamically updated according to the current contact pressure P. The connection relationship between the point masses is established, forming a three-dimensional spring-point network. The mass of each point mass in the network is determined by the material density and control volume. The topology, point mass, spring stiffness, and other parameters of the spring-point network are stored in matrix form to obtain the interface sound propagation network model, stored in HDF5 format.

[0103] Based on the interface acoustic propagation network model obtained in the previous step, the finite difference method is used to calculate the propagation of ultrasound in the model. An ultrasonic excitation signal (such as a Gaussian modulated sinusoidal pulse) is applied to one side of the bolt. The time step and spatial step are set to meet the numerical stability condition. The time step is set to 1 ns and the spatial step is set to 0.1 mm. An explicit central difference scheme is used for time iteration calculation. A receiving point is set on the other side of the connected parts to record the received ultrasonic signal. Fourier transforms are performed on the excitation signal and the received signal to obtain the spectrum. The transmission coefficient is calculated as: T(f) = |Ar(f) / Ai(f)|, where Ar(f) is the spectrum of the received signal, Ai(f) is the spectrum of the excitation signal, and f is the frequency. The frequency of the excitation signal is changed, and the above calculation is repeated to obtain the transmission coefficient at different frequencies. The relationship between the transmission coefficient and frequency is plotted as a curve to obtain the interface transmittance spectrum characteristic set, which is stored in CSV format.

[0104] The interface transmittance spectrum characteristic set obtained in the previous step is analyzed. The phase spectrum φ(f) of the transmittance coefficient T(f) is calculated. The phase spectrum represents the phase change that occurs when the ultrasonic wave passes through the contact interface. The phase spectrum is unwrapped to eliminate the 2π phase jump. The slope of the phase spectrum is calculated: τ(f) = -dφ(f) / dω, where ω = 2πf is the angular frequency. τ(f) represents the group delay of the ultrasonic wave passing through the contact interface. The variation of the group delay τ(f) with frequency f is analyzed. The phase change φ(f) and group delay τ(f) at different frequencies are recorded in a table to form an interface acoustic phase characteristic mapping table, which is stored in TXT text format.

[0105] The interface acoustic phase characteristic mapping table obtained in the previous step is combined with the interface acoustic propagation network model. Phase delay and time delay effects are introduced into the spring-mass network model. Specifically, a phase delay factor exp(iφ(f)) and a time delay factor exp(-iωτ(f)) are added to the spring at the contact interface. The phase delay factor and time delay factor are obtained from the interface acoustic phase characteristic mapping table based on the current excitation frequency f. The modified spring-mass network model is used as the interface contact acoustic characteristic model. This model can simultaneously reflect the stiffness nonlinearity, transmission characteristics, and acoustic phase characteristics of the contact interface, and is stored in HDF5 format.

[0106] Preferably, the multidimensional stress state mapping analysis in step S3 includes:

[0107] Multiaxial stress was determined on the acoustic property model of the interface contact to obtain multiaxial stress data of the acoustoelastic model.

[0108] Tensor-type acoustoelasticity analysis was performed based on the multiaxial stress data of the acoustoelastic model, and the fundamental equations of tensor-type acoustoelasticity were obtained.

[0109] Based on the fundamental equations of tensor-type acoustoelasticity, the sound velocity-stress tensor relationship is derived, and the sound velocity-stress tensor mapping function set is obtained.

[0110] Principal stress decomposition and extraction are performed based on the sound velocity-stress tensor mapping function set to obtain the principal stress acoustic response characteristic table.

[0111] The acoustic elastic matrix transformation data is obtained by performing acoustic elastic matrix transformation based on the principal stress acoustic response characteristic table.

[0112] Bolt stress transmission path analysis is performed based on acoustoelastic matrix transformation data to obtain bolt stress distribution path data.

[0113] A multidimensional stress acoustic mapping matrix is ​​obtained by integrating the bolt stress distribution path, acoustoelastic matrix transformation data, and principal stress acoustic response characteristic table.

[0114] In this embodiment of the invention, the stress distribution of the bolt under preload is calculated using the finite element method based on the interface contact acoustic characteristic model. The three-dimensional CAD models of the bolt and the connected parts are imported into the finite element software ANSYS. Material properties (Young's modulus, Poisson's ratio) are set. The contact surface is defined, and the contact stiffness is set; the contact stiffness data is derived from the interface contact acoustic characteristic model. A preload load is applied to the bolt head, the magnitude of which is determined according to the actual working conditions (e.g., 100kN). The model is meshed using hexahedral elements with a mesh size of 1mm. Static analysis is performed to calculate the stress distribution in the bolt and the connected parts. The stress tensor components (σx, σy, σz, τxy, τyz, τzx) of each node in the bolt are extracted. The stress tensor components of each node are used as multiaxial stress data for the acoustoelastic model and stored in CSV format.

[0115] Based on the multiaxial stress data of the acoustoelastic model obtained in the previous step, the influence of the stress tensor on the sound velocity is considered. Using second-order acoustoelastic theory, the sound velocity is expressed as a function of the stress tensor. For isotropic materials, the relationship between the sound velocity tensor Cij and the stress tensor σij can be expressed as: Where C0 is the speed of sound under zero stress, δij is the Kronecker symbol, and Aijkl, Let Aijkl be the acoustic elastic coefficient tensor. The acoustoelastic coefficient tensor is determined by the second and third elastic constants of the material. Due to material symmetry, the acoustoelastic coefficient tensor can be simplified into several independent parameters. Expanding the above equation yields a set of equations concerning the sound velocity and stress tensor components, namely the fundamental equations of tensor-based acoustoelasticity. These equations are stored in TXT text format.

[0116] Based on the fundamental tensor-based acoustoelastic equations obtained in the previous step, the relationship between the ultrasonic velocity and the stress tensor for different propagation and polarization directions is derived. Consider longitudinal waves and transverse waves with two mutually perpendicular polarization directions. For each wave type, the sound velocity under a given stress state is calculated using the Christoffel equation. The Christoffel equation is: (Γik-ρv) 2 δik)uj=0, where Γik is the Christoffel tensor, ρ is the density, v is the speed of sound, and uj is the polarization direction vector. The Christoffel tensor Γik is determined by the acoustoelastic coefficient tensor and the propagation direction vector. Solving the Christoffel equation for eigenvalues ​​and eigenvectors yields the speed of sound v and the polarization direction uj. The speed of sound v is expressed as a function of the stress tensor components: v=f(σx,σy,σz,τxy,τyz,τzx). For different propagation and polarization directions, a set of speed-stress tensor mapping functions is obtained, i.e., the set of speed-stress tensor mapping functions, which is stored in MATLAB function form.

[0117] Based on the sound velocity-stress tensor mapping function set obtained in the previous step, calculate the principal stresses and their corresponding sound velocities at each node in the bolt. For the stress tensor of each node, calculate its eigenvalues ​​and eigenvectors. The eigenvalues ​​are the principal stresses (σ1, σ2, σ3), and the eigenvectors are the directions of the principal stresses. Substitute the principal stresses into the sound velocity-stress tensor mapping function to calculate the longitudinal wave velocity and transverse wave velocity in each principal stress direction. Record the magnitude and direction of the principal stresses and their corresponding sound velocities at each node. Organize these data into a table to form a principal stress acoustic response characteristic table, which is stored in CSV format.

[0118] Based on the principal stress acoustic response characteristic table obtained in the previous step, the acoustoelastic matrix of each node in the bolt is calculated. The acoustoelastic matrix describes the relationship between stress variation and sound velocity variation. For a given propagation direction and polarization direction, the elements of the acoustoelastic matrix can be calculated using the partial derivatives of the sound velocity-stress tensor mapping function. For example, matrix elements... This represents the change in sound velocity in the first direction caused by the stress change in the first principal stress direction. For each node, all elements of the acoustoelastic matrix are calculated. The acoustoelastic matrix is ​​then transformed from the principal stress coordinate system to the global coordinate system. The transformed acoustoelastic matrix data is stored in matrix form, resulting in the transformed acoustoelastic matrix data, stored in HDF5 format.

[0119] Based on the acoustoelastic matrix transformation data obtained in the previous step, the stress transmission path in the bolt is analyzed. The stress streamline method is employed. Starting from the preload application area at the bolt head, a series of starting points are selected. The principal stress directions are calculated based on the stress tensor at each starting point. The stress streamline is traced along the direction of maximum principal stress at fixed step sizes (e.g., 0.1 mm). At each step, the stress tensor and principal stress direction at that point are calculated based on the acoustoelastic matrix transformation data. The tracing direction is adjusted according to the new principal stress directions, and the tracing of the stress streamline continues. This process is repeated until the stress streamline reaches the bottom of the bolt or the stress value is below a set threshold. The trajectory data of all stress streamlines are recorded to obtain bolt stress distribution path data, which is stored in VTK format for visualization.

[0120] The bolt stress distribution path data, acoustoelastic matrix transformation data, and principal stress acoustic response characteristic table obtained in the previous step are integrated. For each node in the bolt, a mapping relationship is established between its spatial location, stress state (magnitude and direction of principal stresses), acoustic response (sound velocity in different directions and polarizations), and acoustoelastic matrix. This data is organized into a multidimensional matrix. The dimensions of the matrix include: spatial coordinates (x, y, z), principal stresses (σ1, σ2, σ3), principal stress directions (θ1, φ1, θ2, φ2, θ3, φ3), sound velocity (v1, v2, v3), and acoustoelastic matrix elements (M11, M12, ..., Mnn). This matrix is ​​the multidimensional stress acoustic mapping matrix, stored in HDF5 format.

[0121] Preferably, the bolt loosening status identification in step S3 includes:

[0122] Based on the high-order nonlinear coefficient set, temperature-compensated acoustoelastic coefficient matrix, interface contact acoustic characteristic model, and multidimensional stress acoustic mapping matrix, a nonlinear acoustoelastic correction model based on bolt geometry and stress characteristics is constructed to obtain the nonlinear correction model; the preload distribution data is obtained by inverting the preload based on the acoustoelastic nonlinear correction model.

[0123] The temporal fluctuation characteristics of the real-time preload distribution data are extracted to obtain the temporal feature set of preload;

[0124] Anomaly patterns of ultrasonic propagation are identified by performing ultrasonic propagation anomaly features on the preload timing feature set.

[0125] Based on the abnormal characteristics of ultrasonic propagation, the dynamic change of acoustic impedance at the contact surface is analyzed to obtain the characteristic vector of the contact surface state.

[0126] A multidimensional loosening index system was constructed by using the preload timing feature set, ultrasonic propagation anomaly features, and contact surface state feature vector to obtain a multidimensional bolt loosening index system.

[0127] Based on the multidimensional index system of bolt loosening, a loosening degree grading standard was designed to obtain a loosening level assessment system.

[0128] Bolt loosening status data is obtained by detecting bolt loosening status based on the preload timing feature set, ultrasonic propagation anomaly features, and contact surface state feature vector.

[0129] Based on bolt loosening status data and loosening level assessment system, connection status characteristics are judged to obtain connection status characteristic spectrum;

[0130] Based on the characteristic spectrum of the connection state, the mechanical parameters are reconstructed in three dimensions, and the structural stress distribution spectrum is generated to obtain the stress distribution spectrum of the bolt structure.

[0131] In this embodiment of the invention, a complete nonlinear acoustoelastic correction model is constructed based on a high-order nonlinear coefficient set, a temperature-compensated acoustoelastic coefficient matrix, an interface contact acoustic characteristic model, and a multidimensional stress acoustic mapping matrix. First, a detailed finite element model of the bolt is established, including structural features such as the head, bolt, and threads. The element type is selected as SOLID185 hexahedral elements, with a mesh size of 0.5 mm and a total of approximately 25,000 elements. The bolt material properties are specified as the previously measured elastic modulus and Poisson's ratio. Considering the geometric nonlinearity of the bolt, a hyperelastic material model (Mooney-Rivlin model) is used to describe large deformations, and the model parameters are obtained by fitting experimental data. Combined with stress concentration factor analysis, key regions such as the bolt head transition zone and thread root are identified, and a local mesh refinement strategy is established, with the element size refined to 0.1 mm. Based on the piecewise acoustoelastic basic model, a nonlinear mapping relationship between sound velocity and stress is constructed: v(σ,T)=v0+∑k i σ i +α(T-T0), where v0 is the speed of sound under zero stress, k i Here, α represents the high-order nonlinear coefficient, T represents the temperature compensation coefficient, and T0 represents the current temperature. The relationship between interface stiffness and preload is extracted from the interface contact acoustic characteristic model: K(P)=K0(1-e^(-βP)), where K0 is the saturation stiffness and β is the material correlation coefficient. By integrating the tensor transformation relationship in the multidimensional stress-acoustic mapping matrix, a complete acoustoelastic nonlinear correction model is established: v(σ,T,P)=v0+∑k i σ i +α(T-T0)+γ·K(P), where γ is the interface coupling coefficient. This model is implemented in C++ and compiled into a dynamic link library, running on an edge computing device with a computational efficiency of 5ms / time. Based on the established nonlinear acoustic bullet correction model, the preload inversion calculation is performed. First, the time of arrival (ToF) data of the first ultrasonic wave packet is extracted from the acoustic propagation time-frequency domain matrix. The measured ToF is compared with the theoretical calculation value to establish the deviation function: E(P)=|ToF_measured - ToF_theoretical(P)|, where ToF_theoretical(P) is the theoretical arrival time under the preload P. The gradient descent method is used to minimize the deviation function, and the iterative formula is: Where η is the learning rate, initially set to 0.01, and an adaptive adjustment strategy is adopted. The convergence threshold is set to 0.1 μs, and the maximum number of iterations is 200. For cases with difficult convergence, a simulated annealing algorithm is used to avoid getting trapped in local optima. The preload values ​​at 8 measuring points on the bolt are calculated in parallel using CUDA acceleration to achieve real-time calculation. Spatial interpolation (using radial basis functions) is performed on the calculation results to generate a continuous preload distribution field. The preload distribution data is stored in HDF5 format, including spatial coordinates, preload values, and uncertainty estimates. The system updates the preload distribution data every second and retains 24 hours of historical data for trend analysis.

[0132] Based on real-time preload distribution data, the variation law of preload over time is analyzed. Time-domain analysis is performed on the preload data at each monitoring point (different positions on the bolt). Statistical characteristics such as mean, variance, standard deviation, peak value, root mean square value, kurtosis, skewness, waveform factor, impulse factor, and margin factor of the preload time series are calculated. A Fast Fourier Transform (FFT) is performed on the preload time series to obtain the spectrum. Frequency domain characteristics such as centroid frequency, root mean square frequency, and frequency variance are calculated. The time-domain and frequency-domain characteristics are combined into a feature vector. The feature vectors of all monitoring points are combined to form a preload time series feature set, stored in matrix form in CSV format.

[0133] Based on the preload time-series feature set obtained in the previous step, the Isolation Forest algorithm is used to detect outlier data points. The Isolation Forest algorithm is an unsupervised anomaly detection method based on binary trees. This algorithm constructs a set of binary trees by randomly selecting features and split values. Outlier data points are typically located in the shallower layers of the tree, while normal data points are located in the deeper layers. An anomaly score is calculated for each data point; a higher anomaly score indicates that the data point is more likely to be an anomaly. An anomaly score threshold is set (e.g., 0.6). Data points with anomaly scores higher than the threshold are marked as anomalies. Ultrasonic propagation features corresponding to the anomalies are extracted, such as propagation time, attenuation coefficient, and frequency shift. These anomaly features are combined into ultrasonic propagation anomaly features and stored in vector form.

[0134] Based on the ultrasonic propagation anomaly characteristics obtained in the previous step, the change in acoustic impedance at the contact surface is analyzed. Acoustic impedance is the product of material density and sound velocity. Loosening of the bolts leads to a decrease in pressure at the contact surface, thereby changing the equivalent acoustic impedance of the contact surface. According to the reflection coefficient formula: R = (Z2 - Z1) / (Z2 + Z1), where Z1 and Z2 are the acoustic impedances of the bolt and the connected parts, respectively, and R is the reflection coefficient. The reflection coefficient is measured using the ultrasonic pulse-echo method. Based on the change in the reflection coefficient, the change in the equivalent acoustic impedance of the contact surface is calculated. The change in the equivalent acoustic impedance of the contact surface, including the amount, rate of change, and frequency of change, is combined into a contact surface state feature vector and stored in vector form.

[0135] The preload timing feature set, ultrasonic propagation anomaly features, and contact surface state feature vectors obtained in the previous step are integrated to construct a comprehensive bolt loosening index system. The Analytic Hierarchy Process (AHP) is used to determine the weight of each feature. A judgment matrix is ​​constructed to compare the relative importance of different features. The maximum eigenvalue and eigenvector of the judgment matrix are calculated. The eigenvector represents the weight of each feature. The value of each feature is multiplied by its weight, and then summed to obtain a comprehensive loosening index. The loosening indices of all monitoring points are combined into a vector, i.e., the multidimensional bolt loosening index system, and stored in vector form.

[0136] Based on the multi-dimensional index system of bolt loosening obtained in the previous step, a grading standard for the degree of bolt loosening is established. The K-means clustering algorithm is used to cluster the loosening indexes of historical monitoring data (containing bolts with different degrees of loosening). According to the clustering results, the degree of bolt loosening is divided into four levels: normal, slightly loose, moderately loose, and severely loose. The loosening index range corresponding to each level is determined. For example: normal (0-0.2), slightly loose (0.2-0.4), moderately loose (0.4-0.7), and severely loose (0.7-1.0). The grading standard is stored in tabular form, forming a loosening level assessment system, and the table is stored in JSON format.

[0137] Based on the preload timing feature set, ultrasonic propagation anomaly features, and contact surface state feature vector, the current monitoring data is analyzed in real time. Various features of the current monitoring data are calculated and compared with historical data. A Support Vector Machine (SVM) classifier is used to determine the current bolt loosening status. The training data for the SVM classifier is historical monitoring data, labeled as loosening level. The features of the current monitoring data are input into the SVM classifier to obtain the classification results (normal, slightly loose, moderately loose, severely loose). The classification results are stored in real time in JSON format as bolt loosening status data.

[0138] Based on the bolt loosening status data and loosening level assessment system obtained in the previous step, a comprehensive assessment of the bolt connection status is performed. The bolt loosening status data (normal, slightly loose, moderately loose, severely loose) is converted into numerical representations (e.g., normal = 0, slightly loose = 1, moderately loose = 2, severely loose = 3). Combining information such as bolt location and importance, a connection status feature vector is constructed. For example, a hydroelectric generator unit with 10 bolts can have a connection status feature vector represented as [0, 0, 1, 0, 0, 2, 0, 0, 0, 0], indicating that the 3rd bolt is slightly loose, the 6th bolt is moderately loose, and the other bolts are normal. The connection status feature vector is then graphically represented to form a connection status feature spectrum. The feature spectrum can be presented in the form of heatmaps, bar charts, etc., to visually display the connection status of each bolt in the hydroelectric generator unit, and stored as a PNG image.

[0139] The qualitative loosening states (normal, slightly loose, moderately loose, and severely loose) in the connection state characteristic spectrum are converted into quantitative preload loss rates, with the conversion relationships as follows: Normal (0-5%), Slightly loose (5-20%), Moderately loose (20-50%), and Severely loose (>50%). A parametric finite element model is established based on the bolt's geometric characteristics and material parameters. In the model, the contact surface between the bolt head and the connected parts uses surface-to-surface contact elements (CONTA174 / TARGE170), with a friction coefficient set to 0.15. For bolts with different degrees of loosening, the friction characteristics and contact stiffness of the contact surface are adjusted to simulate interface changes during the loosening process. A reverse calculation strategy is adopted, using a numerical optimization method (Powell's method) to find the optimal combination of material parameters and boundary conditions, ensuring the best match between the simulation results and the measured acoustic characteristics. The optimization objective function is set as the weighted mean square error of the measured and simulated acoustic characteristics, with the weighting coefficients determined through sensitivity analysis. Parallel computing technology was employed to simultaneously calculate four different parameter combinations on an 8-core CPU, accelerating the convergence process. The resulting three-dimensional parametric field, encompassing elastic modulus distribution, interface stiffness distribution, and friction coefficient distribution, was stored in VTK format. Based on the reconstructed mechanical parametric field, a stress distribution map of the bolt structure was generated. In the finite element software environment, using the aforementioned parametric model, load conditions under actual system operating conditions were applied, including axial preload, transverse shear force, and bending moment. Load data was acquired in real-time from the hydropower unit monitoring system and transmitted to the computing device via a CAN bus. Static analysis was performed for different loosening states to calculate the stress distribution field. The calculated three-dimensional stress field included von Mises stress, principal stresses, and stress components, with a spatial resolution of 0.5 mm. The stress field was sliced ​​to generate stress distribution maps for key cross-sections and longitudinal sections. An improved rainbow color spectrum (blue-green-yellow-red) was used to represent stress levels, with blue representing minimum stress and red representing maximum stress. Close-up views were generated for stress concentration areas (such as the thread root and head transition zone) to highlight potential danger zones. The local stress concentration factor is calculated and displayed on the stress contour map using contour lines. Simultaneously, the safety factor distribution is calculated; the safety factor is defined as the ratio of the material's yield strength to the local equivalent stress. The generated stress distribution map, contour map, and safety factor distribution map are combined into an interactive atlas of the multi-layered structure in HTML5+JavaScript format, supporting zooming, rotation, and layer cropping. The system updates the structural stress distribution atlas every 5 minutes and triggers an update immediately upon detecting a change in bolt condition. The generated atlas is stored in both PNG and interactive HTML formats and simultaneously pushed to the monitoring terminal for display.

[0140] Preferably, step S4 includes the following steps:

[0141] Step S41: Receive the stress distribution map of the bolt structure, and simultaneously acquire the vibration data of the hydropower unit operation and historical monitoring records to obtain a multi-source monitoring data stream;

[0142] Step S42: Dynamically allocate edge computing resources based on the monitored multi-source data streams to obtain a monitoring intelligent computing solution;

[0143] Step S43: Perform multi-domain spatiotemporal alignment and standardization on the multi-source monitoring data streams according to the monitoring intelligent computing scheme to obtain the monitoring aligned dataset;

[0144] Step S44: Perform feature dimensionality reduction and key parameter extraction on the monitoring alignment dataset to obtain the state key feature vector;

[0145] Step S45: Construct a dynamic Bayesian network fusion model based on the key state feature vectors to obtain the state fusion probability map;

[0146] Step S46: Construct a multidimensional health assessment index system based on the state fusion probability diagram to obtain a multidimensional health assessment matrix;

[0147] Step S47: Calculate the connection reliability index by performing a comprehensive calculation based on the health multidimensional assessment matrix.

[0148] In this embodiment of the invention, the edge computing device receives the stress distribution map (PNG format) of the bolt structure via an Ethernet interface. Simultaneously, it connects to the vibration sensor (accelerometer, 1kHz sampling rate) of the hydropower unit via a CAN bus interface to acquire triaxial vibration acceleration data. The vibration data is sent in data packets, each containing one second of data. The edge computing device also reads historical monitoring records via a local database interface (SQLite), including historical stress distribution maps, historical vibration data, and historical maintenance records. These historical monitoring records are stored in CSV format. The received stress distribution map data, real-time vibration data, and historical monitoring record data are integrated to form a multi-source monitoring data stream. The data stream is synchronized using timestamps with an accuracy of 1 millisecond.

[0149] The edge computing device utilizes an embedded platform based on the NVIDIA Jetson TX2. This platform includes a dual-core Denver2 CPU, a quad-core ARM Cortex-A57 CPU, and a 256-core Pascal GPU. Computational resources are dynamically allocated based on the characteristics of the multi-source data streams being monitored. For stress distribution map data, the GPU is used for image processing and feature extraction. For real-time vibration data, the ARM Cortex-A57 CPU is used for time-domain and frequency-domain analysis. For historical monitoring records, the Denver2 CPU is used for data querying and statistical analysis. A priority scheduling algorithm is employed to prioritize the processing of real-time vibration data and stress distribution map data, ensuring real-time monitoring performance. Information such as the computational task allocation scheme and resource usage is recorded to form a monitoring intelligent computing scheme, stored in JSON format.

[0150] According to the intelligent computing scheme for monitoring, the multi-source monitoring data stream is preprocessed. For stress distribution spectrum data, it is converted into grayscale images and image enhancement processing (histogram equalization) is performed. Image features, such as edge features and texture features, are extracted. For real-time vibration data, time-domain and frequency-domain analysis is performed to extract time-domain features (mean, variance, peak value, etc.) and frequency-domain features (spectrum, power spectral density, etc.). For historical monitoring records, key information, such as historical maximum stress and historical average vibration acceleration, is extracted. Different types of data are time-aligned and sampled at the same time interval (e.g., 1 second). Different types of data are standardized to have the same dimensions and numerical range. The Z-score standardization method is used: x' = (x - μ) / σ, where x is the original data, μ is the mean, σ is the standard deviation, and x' is the standardized data. The standardized data are combined into a matrix to form a monitoring aligned dataset, which is stored in HDF5 format.

[0151] Feature selection and dimensionality reduction are performed on the monitoring aligned dataset. Mutual information is used to evaluate the correlation between each feature and the bolt health status. Mutual information is calculated between each feature and the bolt loosening level (normal, slightly loose, moderately loose, severely loose). Features with mutual information greater than a set threshold (e.g., 0.1) are selected as candidate features. Principal component analysis (PCA) is used to reduce the dimensionality of the candidate features. The covariance matrix of the candidate features is calculated, and eigenvalue decomposition is performed. Principal components with a cumulative variance contribution rate greater than 90% are selected. The original features are projected onto the principal component space to obtain the dimensionality-reduced feature vectors. The dimensionality-reduced feature vectors are used as state-critical feature vectors and stored in vector form.

[0152] Based on the key state feature vectors obtained in the previous step, a dynamic Bayesian network model is constructed. A dynamic Bayesian network is a probabilistic graphical model used to describe the dependencies between variables and their changes over time. Bolt health status, stress characteristics, and vibration characteristics are used as nodes. The connections between nodes are determined based on domain knowledge and data analysis. For example, there are connections between the bolt health status node and the stress characteristic and vibration characteristic nodes. A conditional probability table (CPT) is used to describe the probabilistic relationships between nodes. The parameters of the CPT are learned from historical data. A structural learning algorithm (such as the K2 algorithm) is used to learn the network's topology. A parametric learning algorithm (such as maximum likelihood estimation) is used to learn the network's CPT parameters. The learned dynamic Bayesian network model is represented in graph form, i.e., a state fusion probability graph. In the graph, nodes represent variables, edges represent dependencies between variables, and numbers on the nodes represent conditional probabilities. The state fusion probability graph is stored in XML format.

[0153] Based on the state fusion probability map obtained in the previous step, the probability distribution of the bolt's health state is calculated. Using a Bayesian inference algorithm, the probability of the bolt being in different health states (normal, slightly loose, moderately loose, severely loose) is calculated based on currently observed features (stress features, vibration features, etc.). These probability values ​​are used as health assessment indicators. A multidimensional health assessment indicator system is constructed, including: probability of normal state, probability of slightly loose, probability of moderately loose, and probability of severely loose. These indicators are combined into a vector, i.e., the multidimensional health assessment matrix. Each row of the matrix corresponds to a time point, and each column corresponds to a health assessment indicator, stored in matrix form.

[0154] Based on the multidimensional health assessment matrix obtained in the previous step, a comprehensive connection reliability index is calculated. A weighted average method is used to sum the probabilities of different health states. The weights are determined according to the severity of each health state. For example: normal state has a weight of 1, slight loosening has a weight of 0.8, moderate loosening has a weight of 0.5, and severe loosening has a weight of 0.1. Connection reliability index = 1 * normal state probability + 0.8 * slight loosening probability + 0.5 * moderate loosening probability + 0.1 * severe loosening probability. The connection reliability index ranges from 0 to 1; a higher value indicates a more reliable connection. The connection reliability index is stored in time series format and displayed in real-time on the monitoring interface.

[0155] Preferably, step S5 includes the following steps:

[0156] Step S51: Obtain historical bolt monitoring data; construct a training set based on the connection reliability index and historical bolt monitoring data to obtain a bolt state evolution training set;

[0157] Step S52: Design a residual learning network model based on the bolt state evolution training to obtain a bolt micro-variation recognition depth model;

[0158] Step S53: Use the bolt micro-change recognition depth model to identify the characteristics of minor bolt loosening and obtain an early abnormal feature map;

[0159] Step S54: Perform multi-level state classification based on the early abnormal feature map to obtain a fine classification result of the health status;

[0160] Step S55: Based on the results of the fine classification of health status, construct a time-series prediction model for the degradation trend to obtain the bolt degradation trend prediction model;

[0161] Step S56: Perform multi-scenario life prediction simulation based on the bolt degradation trend prediction model to obtain bolt reliability prediction results.

[0162] In this embodiment of the invention, historical bolt monitoring data, including historical connection reliability indices, historical stress distribution maps, and historical vibration data, is read from the database of an edge computing device. The data is stored in CSV format. At least 1000 historical data points are selected, with a time span of at least 6 months. The historical connection reliability index is used as the label, and the historical stress distribution map and historical vibration data are used as features. The data is preprocessed, including missing value imputation (using mean imputation), outlier removal (using the 3σ principle to remove outliers), and standardization (using Z-score standardization). The dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set is used to train the model, the validation set is used to adjust model parameters, and the test set is used to evaluate model performance. The training set, validation set, and test set are stored in HDF5 format.

[0163] Based on the bolt state evolution training set obtained in the previous step, a one-dimensional convolutional residual network model is designed. The network input is time series data (connection reliability index, stress characteristics, vibration characteristics, etc.). The network contains multiple residual blocks. Each residual block contains two convolutional layers and one skip connection. The convolutional layers use one-dimensional convolutional kernels with a kernel size of 3, a stride of 1, and "same" padding. The activation function is ReLU. The skip connection directly adds the input of the residual block to the output. Before the first residual block, a convolutional layer is added to extract initial features. After the last residual block, a global average pooling layer and a fully connected layer are added. The output of the fully connected layer is the predicted probability of the bolt health state (normal, slightly loose, moderately loose, severely loose). The network is trained using the Adam optimizer with cross-entropy loss. The initial learning rate is set to 0.001, the batch size to 64, and the number of training epochs to 100. During training, the loss on the validation set is monitored, and training is stopped when the loss no longer decreases. The trained model is stored in HDF5 format to obtain the bolt micro-variation recognition depth model.

[0164] A deep model for identifying bolt micro-variation is applied to real-time monitoring data. The real-time acquired data is preprocessed (using the same preprocessing method as the training data). The preprocessed data is then input into the deep model to obtain the predicted probability of the bolt's health status. The rate of change of the predicted probability is calculated. When the rate of change of the predicted probability exceeds a set threshold (e.g., 0.05 / second), a minor loosening is considered detected. The stress and vibration characteristics at this point are extracted as minor loosening features. These features are then visualized, for example, by plotting stress distribution maps and vibration spectra. The visualization results are compared with the features under normal conditions, highlighting abnormal areas. The abnormal feature maps are stored in PNG format to obtain early abnormal feature maps.

[0165] Based on the early abnormal feature map obtained in the previous step, and combined with the output of the deep model (the predicted probability of bolt health status), a more refined status classification is performed. In addition to the four levels of normal, slightly loose, moderately loose, and severely loose, two more levels, "early slightly loose" and "early moderately loose," are added. When the deep model predicts slight looseness, and the early abnormal feature map shows a small abnormal area and a mild degree of abnormality, it is classified as "early slightly loose." When the deep model predicts moderate looseness, and the early abnormal feature map shows a small abnormal area and a mild degree of abnormality, it is classified as "early moderately loose." The classification results (normal, early slightly loose, slightly loose, early moderately loose, moderately loose, severely loose) are used as the refined health status classification results and stored in real time in JSON format.

[0166] Based on the refined health status classification results obtained in the previous step, a Long Short-Term Memory (LSTM) network model is constructed to predict the degradation trend of bolts. LSTM is a special type of recurrent neural network suitable for processing time series data. The model input is the historical refined health status classification results (time series). The model contains one LSTM layer and one fully connected layer. The LSTM layer is used to extract long-term dependencies in the time series. The fully connected layer is used to map the output of the LSTM layer to the predicted health status. The mean squared error (MSE) is used as the loss function, and the Adam optimizer is used to train the network. The initial learning rate is set to 0.001, the batch size is 32, and the number of training epochs is 50. During training, historical data is used for training, and a sliding time window (e.g., window size of 30 days) is set. Data within the window is input each time to predict the health status at the next time point. The trained model is stored in HDF5 format to obtain the bolt degradation trend prediction model.

[0167] Based on the bolt degradation trend prediction model obtained in the previous step, life prediction simulations are performed under different working conditions. Different load conditions, ambient temperature, vibration intensity, and other factors are considered. For each working condition, a set of parameters is set. For example, the load condition can be set to 80%, 100%, and 120% of the rated load. The ambient temperature can be set to -10℃, 25℃, and 50℃. The vibration intensity can be set to 1, 2, and 3 times the normal vibration intensity. For each working condition, the degradation trend prediction model is used to simulate and predict the change in bolt health status over time. The refined classification results of health status are converted into numerical representations (e.g., normal = 0, early slight loosening = 1, slight loosening = 2, early moderate loosening = 3, moderate loosening = 4, severe loosening = 5). When the predicted health status value reaches a set threshold (e.g., 4, corresponding to moderate loosening), the bolt is considered to have failed. The failure time of the bolt under different working conditions is recorded. The failure time is taken as the bolt life. The bolt life data under different working conditions are statistically analyzed to obtain the life distribution. The life distribution is represented in the form of probability density function (PDF) and cumulative distribution function (CDF) to obtain the bolt reliability prediction results, which are stored in JSON format.

[0168] Preferably, the present invention also provides a dynamic monitoring system for the condition of bolts in hydropower units, used to execute the dynamic monitoring method for the condition of bolts in hydropower units as described above, the dynamic monitoring system for the condition of bolts in hydropower units comprising:

[0169] The adaptive resonant excitation module is used to predict the resonant frequency of the bolt and obtain an initial excitation waveform library; based on the initial excitation waveform library, the bolt is subjected to adaptive resonant excitation to obtain the bolt resonant characteristic spectrum.

[0170] The acoustic signal acquisition module is used to acquire the acoustic signal of the bolt based on the bolt resonance characteristic spectrum diagram to obtain a multi-channel original acoustic waveform set; the multi-channel original acoustic waveform set is subjected to working condition noise preprocessing to obtain the acoustic propagation time-frequency domain matrix;

[0171] The nonlinear mechanical property conversion module is used to extract bolt acoustoelastic parameters based on the acoustic propagation time-frequency domain matrix, obtaining a segmented acoustoelastic basic model, a set of higher-order nonlinear coefficients, and a temperature-compensated acoustoelastic coefficient matrix. Based on the segmented acoustoelastic basic model, it performs interface contact nonlinear modeling on the bolt, obtaining an interface contact acoustic property model. It then performs multidimensional stress state mapping analysis on the interface contact acoustic property model, obtaining a multidimensional stress acoustic mapping matrix. Based on the higher-order nonlinear coefficient set, the temperature-compensated acoustoelastic coefficient matrix, the interface contact acoustic property model, and the multidimensional stress acoustic mapping matrix, it constructs a nonlinear acoustoelastic correction model based on the bolt's geometry and stress characteristics, obtaining a nonlinear correction model. Based on the acoustoelastic nonlinear correction model, it performs preload inversion calculation, obtaining preload distribution data. It then identifies the bolt loosening state based on the preload distribution data, obtaining a connection state feature spectrum. Finally, it performs three-dimensional reconstruction of mechanical parameters based on the connection state feature spectrum and generates a structural stress distribution map, obtaining a bolt structural stress distribution map.

[0172] The edge computing module is used to perform multi-source heterogeneous data fusion and edge computing based on the stress distribution map of the bolt structure to obtain the connection reliability index.

[0173] The intelligent early warning module is used to provide intelligent early warning of the bolt status of hydropower unit bolts based on the connection reliability index, and obtain bolt reliability prediction results.

[0174] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0175] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for dynamic monitoring of the condition of bolts in a hydropower unit, characterized in that, The steps include: Step S1: Predict the resonant frequency of the bolt to obtain an initial excitation waveform library; perform adaptive resonant excitation on the bolt according to the initial excitation waveform library to obtain the bolt resonant characteristic spectrum. Step S2: Acquire the bolt acoustic signal based on the bolt resonance characteristic spectrum diagram to obtain a multi-channel original acoustic waveform set; preprocess the multi-channel original acoustic waveform set for working condition noise to obtain the acoustic propagation time-frequency domain matrix; Step S3: Extract bolt acoustic elastic parameters from the acoustic propagation time-frequency domain matrix to obtain the piecewise acoustic elastic basic model, high-order nonlinear coefficient set, and temperature-compensated acoustic elastic coefficient matrix; perform roughness fractal characterization on the bolt connection interface to obtain the contact surface fractal feature description set; analyze the contact area evolution risk under pressure based on the contact surface fractal feature description set to obtain the pressure-contact area response curve set; perform nonlinear mapping of interface equivalent stiffness based on the pressure-contact area response curve set to obtain the interface stiffness nonlinear equation set; perform discrete modeling of spring-mass network based on the interface stiffness nonlinear equation set and the piecewise acoustic elastic basic model to obtain the interface acoustic propagation network model; calculate ultrasonic transmission characteristics based on the interface acoustic propagation network model to obtain the interface transmittance spectrum characteristic set; perform phase jump and time delay effect analysis on the interface transmittance spectrum characteristic set to obtain the interface acoustic phase characteristic mapping table; integrate the interface acoustic phase characteristic mapping table and the interface acoustic propagation network model to obtain the interface contact acoustic characteristic model; perform multidimensional stress state mapping analysis on the interface contact acoustic characteristic model to obtain the multidimensional stress acoustic mapping matrix. Based on the high-order nonlinear coefficient set, temperature-compensated acoustic elastic coefficient matrix, interface contact acoustic characteristic model, and multidimensional stress acoustic mapping matrix, bolt loosening state identification is performed to obtain bolt structure stress distribution map; Step S4: Based on the bolt structure stress distribution map, multi-source heterogeneous data fusion and edge computing are performed to obtain connection reliability index; Step S5: Based on the connection reliability index, intelligent early warning of bolt status is performed on hydropower unit bolts to obtain bolt reliability prediction results.

2. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Collect characteristic parameters of the bolt to obtain a bolt characteristic parameter set; Step S12: Perform resonant frequency prediction analysis based on the bolt characteristic parameter set to obtain a theoretical resonant frequency mapping table; Step S13: Design a high-frequency excitation waveform based on the theoretical resonant frequency mapping table to obtain an initial excitation waveform library; Step S14: Perform an excitation-response test loop on the bolt based on the initial excitation waveform library to obtain a response characteristic dataset; Step S15: Set excitation parameters based on the response characteristic dataset to obtain an excitation parameter configuration; Step S16: Drive a high-frequency high-voltage exciter to perform ultrasonic excitation processing on the bolt based on the excitation parameter configuration to obtain a real-time resonant response data stream; Step S17: Construct a resonant characteristic spectrum from the real-time resonant response data stream to obtain a bolt resonant characteristic spectrum diagram.

3. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Optimize the deployment of the multi-channel receiver array based on the bolt resonance characteristic spectrum diagram to obtain an optimized spatial distribution scheme for the receivers; Step S22: Acquire multi-dimensional acoustic signals from the bolts based on the optimized spatial distribution scheme to obtain a multi-channel original acoustic waveform set; Step S23: Extract and classify the operating condition noise features from the multi-channel original acoustic waveform set to obtain a monitoring environment noise feature map; Step S24: Perform wavelet decomposition and reconstruction of the multi-channel original acoustic waveform set based on the monitoring environment noise feature map to obtain an acoustic feature enhancement waveform group; Step S25: Perform spatiotemporal synchronization integration of the acoustic feature enhancement waveform group to obtain an acoustic propagation synchronization dataset; Step S26: Extract time-domain and frequency-domain dual-dimensional features from the acoustic propagation synchronization dataset to obtain an acoustic propagation time-domain and frequency-domain matrix.

4. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, Step S3, which involves extracting the bolt acoustic elastic parameters, includes: analyzing and identifying the acoustic feature components of the acoustic propagation time-frequency domain matrix to obtain a set of acoustic feature parameters; performing material acoustoelastic analysis based on the acoustic feature parameter set to obtain the acoustoelastic coefficients of the bolt material; modeling the basic acoustoelastic relationship based on the acoustoelastic coefficients of the bolt material to obtain a segmented acoustoelastic basic model; calibrating the higher-order nonlinear coefficients based on the acoustoelastic coefficients of the bolt material to obtain a set of higher-order nonlinear coefficients; and extracting the temperature-acoustic coupling correction factor based on the segmented acoustoelastic basic model and the set of higher-order nonlinear coefficients to obtain a temperature-compensated acoustoelastic coefficient matrix.

5. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, The multidimensional stress state mapping analysis described in step S3 includes: determining the multiaxial stress of the interface contact acoustic characteristic model to obtain the multiaxial stress data of the acoustoelastic model; performing tensor-type acoustoelastic analysis based on the multiaxial stress data of the acoustoelastic model to obtain the tensor-type acoustoelastic fundamental equation set; deriving the sound velocity-stress tensor relationship based on the tensor-type acoustoelastic fundamental equation set to obtain the sound velocity-stress tensor mapping function set; performing principal stress decomposition and extraction based on the sound velocity-stress tensor mapping function set to obtain the principal stress acoustic response characteristic table; performing acoustoelastic matrix transformation processing based on the principal stress acoustic response characteristic table to obtain acoustoelastic matrix transformation data; performing bolt stress transmission path analysis based on the acoustoelastic matrix transformation data to obtain bolt stress distribution path data; and integrating the bolt stress distribution path, acoustoelastic matrix transformation data, and principal stress acoustic response characteristic table into a multidimensional stress acoustic mapping matrix.

6. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, Step S3, the identification of bolt loosening status, includes: constructing a nonlinear acoustoelastic correction model based on bolt geometry and stress characteristics using a high-order nonlinear coefficient set, a temperature-compensated acoustoelastic coefficient matrix, an interface contact acoustic characteristic model, and a multidimensional stress acoustic mapping matrix; performing preload inversion calculation based on the acoustoelastic nonlinear correction model to obtain preload distribution data; extracting time-series fluctuation characteristics from the real-time preload distribution data to obtain a preload time-series feature set; identifying ultrasonic propagation anomaly patterns in the preload time-series feature set to obtain ultrasonic propagation anomaly features; and analyzing the dynamic changes in contact surface acoustic impedance based on the ultrasonic propagation anomaly features to obtain contact surface state characteristics. The system employs a multi-dimensional loosening index system, which is constructed by analyzing the preload time sequence feature set, ultrasonic propagation anomaly features, and contact surface state feature vectors. Based on this system, a loosening degree grading standard is designed to obtain a loosening level assessment system. Bolt loosening state detection is performed using the preload time sequence feature set, ultrasonic propagation anomaly features, and contact surface state feature vectors to obtain bolt loosening state data. Connection state characteristics are judged based on the bolt loosening state data and the loosening level assessment system to obtain a connection state feature spectrum. Finally, mechanical parameters are reconstructed in three dimensions based on the connection state feature spectrum, and a structural stress distribution map is generated to obtain a bolt structural stress distribution map.

7. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Receive the stress distribution map of the bolt structure, and simultaneously acquire the vibration data of the hydropower unit operation and historical monitoring records to obtain a multi-source monitoring data stream; Step S42: Dynamically allocate edge computing resources based on the multi-source monitoring data stream to obtain a monitoring intelligent computing scheme; Step S43: Perform multi-domain data spatiotemporal alignment and standardization on the multi-source monitoring data stream according to the monitoring intelligent computing scheme to obtain a monitoring aligned dataset; Step S44: Perform feature dimensionality reduction and key parameter extraction on the monitoring aligned dataset to obtain a state key feature vector; Step S45: Construct a dynamic Bayesian network fusion model based on the state key feature vector to obtain a state fusion probability map; Step S46: Construct a multi-dimensional health assessment index system based on the state fusion probability map to obtain a health multi-dimensional assessment matrix; Step S47: Perform a comprehensive calculation of the connection reliability index based on the health multi-dimensional assessment matrix to obtain the connection reliability index.

8. The method for dynamic monitoring of the condition of bolts in a hydropower unit according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Obtain historical bolt monitoring data; construct a training set based on the connection reliability index and historical bolt monitoring data to obtain a bolt state evolution training set; Step S52: Design a residual learning network model based on the bolt state evolution training to obtain a bolt micro-change recognition deep model; Step S53: Use the bolt micro-change recognition deep model to identify bolt micro-loosening features to obtain an early abnormal feature map; Step S54: Perform multi-level state classification based on the early abnormal feature map to obtain a fine classification result of the health state; Step S55: Construct a degradation trend time series prediction model based on the fine classification result of the health state to obtain a bolt degradation trend prediction model; Step S56: Perform multi-scenario life prediction simulation based on the bolt degradation trend prediction model to obtain bolt reliability prediction results.

9. A dynamic monitoring system for the condition of bolts in a hydropower unit, characterized in that, To implement the dynamic monitoring method for bolt condition of hydropower units as described in claim 1, the dynamic monitoring system for bolt condition of hydropower units includes: an adaptive resonant excitation module for predicting the resonant frequency of the bolts to obtain an initial excitation waveform library; performing adaptive resonant excitation on the bolts according to the initial excitation waveform library to obtain a bolt resonant characteristic spectrum diagram; an acoustic signal acquisition module for acquiring bolt acoustic signals according to the bolt resonant characteristic spectrum diagram to obtain a multi-channel original acoustic waveform set; performing operating noise preprocessing on the multi-channel original acoustic waveform set to obtain an acoustic propagation time-domain-frequency domain matrix; a nonlinear mechanical characteristic conversion module for extracting bolt acoustic elastic parameters according to the acoustic propagation time-domain-frequency domain matrix to obtain a piecewise acoustic elastic basic model, a high-order nonlinear coefficient set, and a temperature-compensated acoustic elastic coefficient matrix; performing roughness fractal characterization on the bolt connection interface to obtain a contact surface fractal feature description set; performing contact area evolution risk assessment on the contact surface under pressure according to the contact surface fractal feature description set to obtain a pressure-contact area response curve set; and performing nonlinear mapping of interface equivalent stiffness according to the pressure-contact area response curve set to obtain interface stiffness. The system employs a set of nonlinear equations for interface stiffness. Based on the set of nonlinear equations for interface stiffness and the piecewise acoustic elasticity model, a spring-mass network is discretized to obtain an interface acoustic propagation network model. Ultrasonic transmission characteristics are calculated using this model to obtain an interface transmittance spectrum characteristic set. Phase jump and time delay effects are analyzed on the interface transmittance spectrum characteristic set to obtain an interface acoustic phase characteristic mapping table. The interface acoustic phase characteristic mapping table and the interface acoustic propagation network model are integrated to obtain an interface contact acoustic characteristic model. Multidimensional stress state mapping analysis is performed on the interface contact acoustic characteristic model to obtain a multidimensional stress acoustic mapping matrix. Bolt loosening state identification is performed based on the set of high-order nonlinear coefficients, the temperature-compensated acoustic elasticity coefficient matrix, the interface contact acoustic characteristic model, and the multidimensional stress acoustic mapping matrix to obtain a bolt structure stress distribution map. An edge computing module is used to perform multi-source heterogeneous data fusion and edge computing based on the bolt structure stress distribution map to obtain a connection reliability index. An intelligent early warning module is used to provide intelligent early warning of bolt status for hydropower unit bolts based on the connection reliability index, obtaining bolt reliability prediction results.

Images