Fault diagnosis method and system based on simulation model optimization and multi-domain feature fusion

Abstract

The application belongs to the technical field of fault diagnosis, and discloses a fault diagnosis method and system based on simulation model optimization and multi-domain feature fusion. The method comprises the following steps: constructing an initial tower drum simulation model of a tower drum and a fault diagnosis model of multi-domain feature fusion, optimizing and condensing the initial tower drum simulation model to obtain a condensed tower drum simulation model; using the condensed tower drum simulation model to generate simulation data of the tower drum, and performing multi-domain feature fusion on the simulation data and corresponding measured data to obtain a comprehensive feature vector; and using the fault diagnosis model to perform fault diagnosis on the comprehensive feature vector to obtain a corresponding real-time fault diagnosis result. The application solves the problems of low accuracy, poor effect and poor comprehensiveness in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a fault diagnosis method and system based on simulation model optimization and multi-domain feature fusion. Background Technology

[0002] The wind turbine tower, especially the tower of a large wind turbine, is a crucial structure supporting key components such as the hub, blades, and nacelle. Its operational status directly affects the stability and safety of the entire wind turbine. During long-term operation, the tower may suffer structural damage or performance degradation due to various factors such as material fatigue, corrosion, wind load impact, and foundation settlement, leading to malfunctions. Therefore, real-time and accurate fault diagnosis of the tower has significant engineering and economic value.

[0003] The existing technology has the following drawbacks:

[0004] 1) Low accuracy: In existing technologies, fault diagnosis models based on physical models or data-driven methods will have significantly reduced diagnostic accuracy if the model accuracy is low or the training data is insufficient.

[0005] 2) Poor performance: Existing technologies directly compare simulation model data with measured data and make fault judgments based on the comparison results. However, there are large errors between simulation model data and measured data, which can easily lead to false alarms, resulting in poor performance. Furthermore, it requires a lot of manual intervention, which is inefficient.

[0006] 3) Poor comprehensiveness: Traditional data comparison methods only compare signals in the time domain, which will lose a lot of important information reflecting the operating status of the equipment, resulting in insufficient basis for fault judgment and easy to miss some hidden faults. Summary of the Invention

[0007] To address the problems of low accuracy, poor effectiveness, and lack of comprehensiveness in existing technologies, the present invention aims to provide a fault diagnosis method and system based on simulation model optimization and multi-domain feature fusion.

[0008] The technical solution adopted in this invention is as follows:

[0009] A fault diagnosis method based on simulation model optimization and multi-domain feature fusion includes the following steps:

[0010] An initial tower simulation model and a fault diagnosis model based on multi-domain feature fusion were constructed. The initial tower simulation model was then optimized and condensed to obtain the condensed tower simulation model.

[0011] Using the condensed tower simulation model, simulation data of the tower is generated, and the simulation data and the corresponding measured data are fused together using multi-domain features to obtain a comprehensive feature vector.

[0012] Using a fault diagnosis model, fault diagnosis is performed on the comprehensive feature vector to obtain the corresponding real-time fault diagnosis results.

[0013] Furthermore, an initial tower simulation model and a fault diagnosis model based on multi-domain feature fusion are constructed. The initial tower simulation model is then optimized and condensed to obtain a condensed tower simulation model, including the following steps:

[0014] Based on the basic simulation information and 3D scanning data of the tower, an initial tower simulation model is constructed using 3D simulation technology.

[0015] Using machine learning algorithms, an initial fault diagnosis model is constructed, and a training sample set containing different working conditions is input for optimization training to obtain the final fault diagnosis model.

[0016] The initial tower simulation model was optimized and condensed to obtain the condensed tower simulation model.

[0017] Furthermore, the simulation data includes the simulation displacement, simulation acceleration, and simulation tilt angle data output from the optimized tower simulation model;

[0018] The measured data includes the measured displacement, measured acceleration, and measured tilt angle data of the corresponding positions on the tower collected by the sensors.

[0019] Furthermore, the fault diagnosis model is constructed based on the RF-Attention-SVM algorithm, and the fault diagnosis model includes a key feature screening module constructed based on the RF algorithm, an attention weight module constructed based on the Attention mechanism, and a fault diagnosis module constructed based on the SVM algorithm, which are connected in sequence.

[0020] Furthermore, the initial tower simulation model is optimized and condensed to obtain the condensed tower simulation model, including the following steps:

[0021] Sensitivity calculations were performed on the initial tower simulation model, and model parameters whose sensitivity exceeded the threshold were taken as target model parameters that needed to be optimized.

[0022] A multi-objective optimization algorithm is used to optimize the target model parameters of the initial tower simulation model to obtain the optimized tower simulation model.

[0023] The dynamic condensation method is used to reduce the degrees of freedom of the optimized tower simulation model, resulting in a condensed tower simulation model.

[0024] Furthermore, a multi-objective optimization algorithm is used to optimize the target model parameters of the condensed tower simulation model to obtain the optimized tower simulation model, including the following steps:

[0025] Using the condensed tower simulation model, a simulation sample of the tower is generated, and frequency domain analysis is performed to obtain the simulation natural frequency and simulation mode shape of the simulation sample.

[0026] Using sensors, the actual samples corresponding to the simulated samples are collected, and the EFDD algorithm is used to obtain the actual natural frequencies and actual mode shapes of the actual samples.

[0027] Define the objective function based on the simulated natural frequencies, simulated mode shapes, measured natural frequencies, and measured mode shapes;

[0028] Based on the objective function, the fitness function of the MOSGA algorithm is set, and the target model parameters are iteratively optimized based on the fitness function to obtain the optimal target model parameters;

[0029] Based on the optimal model parameters, the simulation model of the condensed tower is optimized to obtain the optimized tower simulation model.

[0030] Furthermore, based on the objective function, the fitness function of the MOSGA algorithm is defined, and the target model parameters are iteratively optimized based on the fitness function to obtain the optimal target model parameters, including the following steps:

[0031] Based on the objective function, the fitness function of the MOSGA algorithm is set, and the initial model parameters are encoded into individual vectors of the MOSGA algorithm.

[0032] Based on the individual vectors, a number of initial solutions are generated using a chaotic mapping sequence algorithm; each initial solution corresponds to an initial set of model parameters.

[0033] Based on the fitness function, the MOSGA algorithm is used to iteratively optimize several initial solutions and retain the best individual in each iteration;

[0034] If the number of iterations is greater than or equal to the iteration threshold or the fitness value of the best individual is less than the fitness threshold, then the best individual will be output as the optimal solution.

[0035] Decoding the individual vectors of the optimal solution yields the optimal model parameters for the condensed tower simulation model.

[0036] Furthermore, using the condensed tower simulation model, simulation data for the tower is generated, and the simulation data and corresponding measured data are fused using multi-domain features to obtain a comprehensive feature vector, including the following steps:

[0037] Using a condensed tower simulation model, simulation data of the tower is generated, and sensors are used to collect the corresponding measured data.

[0038] The simulation data and the corresponding measured data are synchronized by adjusting the time step and the number of sampling points to obtain synchronized simulation data and synchronized measured data.

[0039] The time-domain and frequency-domain features of the simulated data and measured data after synchronous processing are extracted, and multi-domain features are fused to obtain a comprehensive feature vector.

[0040] Furthermore, using a fault diagnosis model, fault diagnosis is performed on the comprehensive feature vector to obtain the corresponding real-time fault diagnosis results, including the following steps:

[0041] Using the pre-trained RF structure in the key feature filtering module of the fault diagnosis model, several key feature components are extracted from the comprehensive feature vector.

[0042] Based on the dynamic attention weights, the attention weight module of the fault diagnosis model is used to weight and concatenate several key feature components to obtain the concatenated features.

[0043] Based on the splicing characteristics, the fault diagnosis module of the fault diagnosis model is used to perform fault diagnosis prediction and obtain the corresponding real-time fault diagnosis results.

[0044] A fault diagnosis system based on simulation model optimization and multi-domain feature fusion is provided to implement a fault diagnosis method. The system includes a model building and optimization unit, a multi-domain feature fusion unit, and a fault diagnosis unit connected in sequence.

[0045] The beneficial effects of this invention are as follows:

[0046] This invention discloses a fault diagnosis method and system based on simulation model optimization and multi-domain feature fusion. The comprehensive feature vector obtained through multi-domain feature fusion and the fault diagnosis model constructed based on machine learning algorithms achieve high accuracy and can uncover deep relationships between data and faults, improving diagnostic accuracy. The initial simulation model is optimized through dynamic convergence and multi-objective optimization algorithms, enabling the optimized tower simulation model to more accurately reflect the dynamic characteristics of the actual tower, improving the reliability of simulation data, avoiding large errors between simulation model data and measured data, reducing false alarms, and improving effectiveness. Combined with automated fault diagnosis, no manual intervention is required, improving efficiency. In addition to utilizing traditional vibration acceleration data, it also integrates multi-domain information such as displacement and tilt angle, which can more comprehensively reflect the structural state of the tower, improving the comprehensiveness of feature information and the robustness of diagnosis. Furthermore, by comparing data in the time and frequency domains and fusing the time-frequency domain features of multi-source signals, most of the important information reflecting the equipment's operating status is retained, avoiding the omission of some hidden faults due to insufficient fault judgment criteria.

[0047] Other beneficial effects of the present invention will be further explained in the specific embodiments. Attached Figure Description

[0048] Figure 1 This is a flowchart of the fault diagnosis method based on simulation model optimization and multi-domain feature fusion in this invention.

[0049] Figure 2 This is a structural block diagram of the fault diagnosis system based on simulation model optimization and multi-domain feature fusion in this invention. Detailed Implementation

[0050] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1:

[0052] like Figure 1 As shown in the figure, this embodiment provides a fault diagnosis method based on simulation model optimization and multi-domain feature fusion, including the following steps:

[0053] S1: Construct an initial tower simulation model and a fault diagnosis model based on multi-domain feature fusion, and optimize and condense the initial tower simulation model to obtain a condensed tower simulation model, including the following steps:

[0054] S1-1: Based on the basic simulation information and 3D scanning data of the tower, construct the initial tower simulation model using 3D simulation technology, including the following steps:

[0055] S1-1-1: Input the basic simulation information and 3D scanning data of the tower into the Computer-Aided Engineering (CAE) software;

[0056] Basic simulation information includes data such as the tower's type, size, material, shape, and structure;

[0057] S1-1-2: Based on the basic simulation information of the tower, use 3D simulation technology to construct a 3D simulation model of the tower;

[0058] S1-1-3: Based on the three-dimensional scanning data of the tower, the three-dimensional simulation model is corrected, and several mesh nodes are set in the three-dimensional simulation model to obtain the initial tower simulation model.

[0059] The mesh nodes correspond to the actual sensor installation locations on the tower; each mesh node is equivalent to a sensor, and the initial tower simulation model can output the simulated displacement, simulated acceleration, and simulated tilt angle data corresponding to the mesh node location.

[0060] S1-2: Using machine learning algorithms, an initial fault diagnosis model is constructed, and a training sample set containing different working conditions is input for optimization training to obtain the final fault diagnosis model.

[0061] The fault diagnosis model is built based on the Random Forest (RF)-Attention-Support Vector Machine (SVM) algorithm, and the fault diagnosis model includes a key feature screening module built based on the RF algorithm, an attention weight module built based on the Attention mechanism, and a fault diagnosis module built based on the SVM algorithm, which are connected in sequence.

[0062] The comprehensive feature vector typically contains a large number of features, which may include redundant information or noise. The key feature selection module evaluates the importance of each feature by calculating the sum of impurities (such as Gini impurities) reduced across all trees, thus filtering key feature components. This significantly reduces the data dimensionality of subsequent processing, reduces computational burden, and accelerates model training and prediction. The attention weight module assigns different weights to the selected key feature components. The attention mechanism learns the relationships between these features and automatically determines which features are more important at a given moment or for a particular type of fault, based on the needs of the current diagnostic task. Through weighting, the model can more effectively utilize the information in key features, especially those features with higher discriminative power under specific conditions, thereby improving the overall expressive power of the feature vector. By intelligently allocating weights, the model can more accurately capture subtle features of the fault, thereby improving the accuracy and discriminative power of fault diagnosis. The fault diagnosis module performs the final fault category judgment based on the concatenated features output by the attention weight module. It performs excellently in high-dimensional space and can effectively process feature vectors that may be high-dimensional and information-rich after the first two steps of processing.

[0063] S1-3: Optimize and condense the initial tower simulation model to obtain the condensed tower simulation model, including the following steps:

[0064] S1-3-1: Perform sensitivity calculations on the initial tower simulation model and use the model parameters whose sensitivity exceeds the threshold as the target model parameters that need to be optimized.

[0065] The formula for calculating sensitivity is:

[0066]

[0067] In the formula, Sensitivity; This is the natural frequency of the actual wind turbine. The natural frequency of the initial tower simulation model; This represents the adjustment amount for the l-th model parameter; The original value of the l-th model parameter is given. The above formula performs dimensionless sensitivity analysis on the model parameters to eliminate the influence of parameter dimension differences on sensitivity evaluation.

[0068] S1-3-2: Using a multi-objective optimization algorithm, the target model parameters of the initial tower simulation model are optimized to obtain the optimized tower simulation model, including the following steps:

[0069] S1-3-2-1: Using the condensed tower simulation model, generate a simulation sample of the tower and perform frequency domain analysis to obtain the simulation natural frequency and simulation mode shape of the simulation sample;

[0070] S1-3-2-2: Using sensors, acquire the measured samples corresponding to the simulated samples, and use the Enhanced Frequency Domain Decomposition (EFDD) algorithm to obtain the measured natural frequencies and measured mode shapes of the measured samples. This includes the following steps:

[0071] S1-3-2-2-1: Using a sensor, collect the actual sample corresponding to the simulation sample, and perform noise reduction and bandpass filtering preprocessing on the actual acceleration signal of the actual sample to obtain the preprocessed actual acceleration signal, so as to remove noise and irrelevant frequency components;

[0072] S1-3-2-2-2: Perform Fast Fourier Transform (FFT) on the preprocessed measured acceleration signal to obtain the measured acceleration amplitude spectrum and the measured phase spectrum;

[0073] S1-3-2-2-3: Based on the measured acceleration amplitude spectrum and the measured phase spectrum, construct a Hankel matrix to characterize the frequency domain correlation of the signal, with the frequency point as the center. The row dimension of the Hankel matrix is ​​the number of sensors, and the column dimension is the time lag point.

[0074] S1-3-2-2-4: Perform Singular Value Decomposition (SVD) on the matrix to obtain the singular value spectrum and left and right singular vectors, where the peak value of the singular value spectrum corresponds to the measured natural frequency and the left singular vector corresponds to the measured mode shape.

[0075] S1-3-2-2-5: By performing peak detection on the singular value spectrum peaks, the measured natural frequency is determined, and the left singular vector at the corresponding frequency is normalized to obtain the measured mode shape;

[0076] To intuitively reflect the vibration modes of each degree of freedom of the structure, this method efficiently separates noise from the true modes through frequency domain matrix decomposition and statistical characteristic analysis, and is suitable for modal parameter identification under environmental excitation.

[0077] S1-3-2-3: Define the objective function based on the simulated natural frequency, simulated mode shape, measured natural frequency, and measured mode shape;

[0078] The formula is:

[0079]

[0080] In the formula, The objective function, i.e., the error value, is used to balance the importance of natural frequency and mode shape error; The first objective weight and the second objective weight; To simulate the natural frequency and the measured natural frequency; For simulated vibration modes and measured vibration modes; This is a natural frequency indication value; This is a mode shape indicator. This represents the total number of inherent frequencies. This represents the total number of vibration modes;

[0081] S1-3-2-4: Based on the objective function, define the fitness function of the Multi-Objective Snow Geese Algorithm (MOSGA) algorithm, and iteratively optimize the objective model parameters based on the fitness function to obtain the optimal objective model parameters, including the following steps:

[0082] S1-3-2-4-1: Based on the objective function, set the fitness function of the MOSGA algorithm and encode the initial objective model parameters into individual vectors of the MOSGA algorithm;

[0083] The formula is:

[0084]

[0085] In the formula, The fitness function; For MOSGA individuals The corresponding model error value; It is a MOSGA individual; These are the first fitness weight, the second fitness weight, and the third fitness weight; For MOSGA individuals The corresponding model size; For MOSGA individuals The corresponding model training cost;

[0086] S1-3-2-4-2: Based on the individual vectors, use the chaotic mapping sequence algorithm to generate several initial solutions; each initial solution corresponds to an initial model parameter.

[0087] The formula is:

[0088]

[0089] In the formula, The initial MOSGA individuals generated for the Circle chaotic mapping sequence, i.e., the initial solutions; For each MOSGA individual, i is a randomly generated MOSGA individual; For the remainder function;

[0090] S1-3-2-4-3: Based on the fitness function, the MOSGA algorithm is used to iteratively optimize several initial solutions, and the best individual in each iteration is retained. This includes the following steps:

[0091] S1-3-2-4-3-1: Use the fitness function to obtain the initial fitness value of each initial MOSGA individual in the initial MOSGA population, and select the initial MOSGA individual with the lowest fitness value as the leader goose.

[0092] S1-3-2-4-3-2: Entering the exploration phase, a leader goose rotation mechanism, a call guidance mechanism, and a dynamic reverse mechanism are introduced to iteratively update the initial MOSGA population, resulting in an updated MOSGA population, while retaining the best individuals.

[0093] The leader goose rotation mechanism selects a new leader goose in each iteration based on the fitness values ​​of individual MOSGA individuals. This mechanism can prevent the leader goose from getting trapped in local optima too early and enhance the global search capability of the algorithm.

[0094] The formula is:

[0095]

[0096] In the formula, As the leading goose in an update; For the first The initial MOSGA individual with the third-to-last fitness value in the initial MOSGA population after the first iteration; For the first The initial MOSGA individual with the fifth-to-last fitness value in the initial MOSGA population after the first iteration; This represents the current iteration number; The optimal individual; It is the first weighting factor; This is a function for generating random numbers;

[0097] The call guidance mechanism adjusts the individual position update using a sound wave propagation attenuation model based on the distance between the MOSGA individual and the leader goose. MOSGA individuals that are closer to the leader goose have a greater influence on their position update and can quickly move closer to the optimal solution, while MOSGA individuals that are farther away have a smaller influence on their position update and can maintain a certain level of exploration ability. This mechanism can avoid excessive aggregation or dispersion of the group and improve the local search accuracy of the algorithm.

[0098] The formula is:

[0099]

[0100] In the formula, For a newly updated MOSGA individual; For the first The initial MOSGA individuals for the number of iterations; The initial sound intensity received by the MOSGA individual; For sound intensity parameters; The initial sound intensity; The lowest acceptable sound intensity; The convergence factor; The initial MOSGA individual that is furthest away; The parameter is random. Let Brownian motion function be used. These are Brownian motion parameters; This is the XOR operation symbol;

[0101]

[0102] In the formula, is the convergence factor; tanh(.) is the hyperbolic tangent function; This represents the current iteration number; a is the maximum number of iterations. max a min λ represents the maximum and minimum values ​​of the convergence factor, respectively; λ is the deceleration rate parameter. As a decreasing periodic parameter, λ = -2π. =π;

[0103] The dynamic reverse mechanism dynamically reverses the initial MOSGA individuals, increasing the diversity of exploration directions and avoiding getting trapped in local optima;

[0104] The formula is:

[0105]

[0106] In the formula, For a single update of the reverse MOSGA individual; γ is the decreasing inertia coefficient; L max L min These are the maximum and minimum values ​​in the vector space, respectively.

[0107] The leader goose from the first update, several MOSGA individuals from the first update, and several reverse MOSGA individuals from the first update will be integrated to obtain a MOSGA population from the first update, and the MOSGA individual with the lowest fitness value will be retained as the best individual.

[0108] S1-3-2-4-3-3: Entering the development phase, anomaly boundary strategy and Gaussian mutation mechanism are introduced to perform a second update on the MOSGA population updated once, resulting in a second-updated MOSGA population, and the best individual is retained.

[0109] The abnormal boundary strategy calculates the difference between the fitness value of each updated MOSGA individual and the population average fitness value. For MOSGA individuals with fitness values ​​much higher than the population average, their position update method will be adjusted, such as using Gaussian mutation mechanism, larger step size or smaller step size. This mechanism can help individuals avoid getting trapped in local optima and improve the convergence speed and accuracy of the algorithm.

[0110] The formula is:

[0111]

[0112] In the formula, This is a MOSGA individual that has undergone a second update; For a newly updated MOSGA individual; The fitness function; This represents the average fitness value of the population. The individual with the highest fitness value is the MOSGA. These are the second and third weighting factors; These are parameters for the Gaussian mutation mechanism;

[0113] S1-3-2-4-4: If the number of iterations is greater than or equal to the iteration number threshold or the fitness value of the best individual is less than the fitness threshold, then the best individual will be output as the optimal solution.

[0114] S1-3-2-4-5: Decode the individual vectors of the optimal solution to obtain the optimal target model parameters of the condensed tower simulation model;

[0115] S1-3-2-5: Optimize the initial tower simulation model based on the optimal target model parameters to obtain the optimized tower simulation model;

[0116] In addition to the MOSGA algorithm used above, the second-generation non-dominated sorting genetic algorithm (NSGA-II) can also be used for optimization.

[0117] S1-3-3: The dynamic condensation method is used to reduce the degrees of freedom of the optimized tower simulation model, resulting in the condensed tower simulation model.

[0118] The actual structure of the tower has fewer degrees of freedom than the finite element model of the optimized tower simulation model, which causes the modal degrees of freedom of the two to not match. This results in the loss and discontinuity of modes in the dynamic test model. Therefore, it is necessary to condense the degrees of freedom of the finite element model to make the degrees of freedom match.

[0119] S2: Using the condensed tower simulation model, generate tower simulation data, and fuse the simulation data and corresponding measured data through multi-domain feature fusion to obtain a comprehensive feature vector;

[0120] The simulation data includes the simulated displacement, simulated acceleration, and simulated tilt angle data output from the simulation model of the condensed tower.

[0121] The measured data includes the measured displacement, measured acceleration, and measured tilt angle data of the corresponding positions on the tower, collected by the sensors;

[0122] The measured acceleration and tilt data transmitted via optical fiber are filtered to remove noise. The acceleration and tilt data values ​​measured by the acceleration and tilt sensors are transmitted to the host computer via optical fiber. The acceleration signal is subjected to high-pass filtering to remove baseline drift with a cutoff frequency of (0-0.1) Hz, and low-pass filtering to remove high-frequency noise with a cutoff frequency of (50-100) Hz. Wavelet threshold denoising is used on the tilt data to eliminate instantaneous interference caused by gusts.

[0123] Using a condensed tower simulation model, simulation data for the tower is generated. The simulation data and corresponding measured data are then fused using multi-domain feature fusion to obtain a comprehensive feature vector. This process includes the following steps:

[0124] S2-1: Use the condensed tower simulation model to generate tower simulation data, and use sensors to collect the actual measurement data corresponding to the simulation data;

[0125] S2-2: Synchronize the simulation data and the corresponding measured data in terms of time step and number of sampling points to obtain the synchronized simulation data and the synchronized measured data.

[0126] Assume the time step of the simulation data is The number of sampling points is The measured data time step is: The number of sampling points is Interpolation is used to adjust the time steps of both. Number of sampling points This ensures that the data is aligned over time.

[0127] S2-3: Extract the time-domain and frequency-domain features of the simulated data and measured data after synchronous processing, and perform multi-domain feature fusion to obtain a comprehensive feature vector;

[0128] Time-domain characteristics include: peak acceleration variation, mean tilt angle variation, and standard deviation of acceleration or displacement variation, as shown in the formula:

[0129]

[0130] In the formula, This represents the change in peak acceleration. These are measured peak acceleration data. For simulated peak acceleration data; if If the value is greater than 0.3, the bolt loosening time domain is abnormal. Bolt loosening will lead to a decrease in connection stiffness, an increase in vibration and impact response, and a significant difference in peak values. The threshold of 0.3 is selected based on the statistical difference of peak values ​​between the loosening state and the normal state in vibration experiments of similar equipment, and can effectively distinguish between the two states.

[0131]

[0132] In the formula, This represents the mean change in the angle of inclination. These are the average measured tilt angle data. This is the average tilt angle data for simulation; if >0.2, and the energy proportion in the high-frequency region (5-10Hz) increases, indicating excessive tilt; if A value >0.25, and an increase of 40% in the energy proportion of the low-frequency region (below 0.5Hz), is marked as foundation settlement. Excessive tilting changes the structural center of gravity, triggering high-frequency vibrations. Changes in the mean value reflect static offset, while an increase in high-frequency energy reflects abnormal dynamic response. The threshold of 0.2 is set based on experience with allowable tilt deviations and vibration energy distribution changes in engineering. Foundation settlement leads to a decrease in the overall structural stiffness and a significant increase in low-frequency vibration energy. The threshold of 0.25 for the mean change is higher than that for excessive tilting. Because settlement is a gradual and slow change, a larger static offset is needed for identification. The 40% energy increase is based on the abrupt change in the energy proportion of low-frequency modes during settlement.

[0133]

[0134] In the formula, This refers to the change in the standard deviation of acceleration or displacement. For the standard deviation of measured acceleration or displacement, For the standard deviation of acceleration or displacement in the simulation; if If the value is greater than 0.4, the crack time domain is abnormal; the crack causes a sudden change in the local stiffness of the structure, the vibration signal dispersion is enhanced, and the standard deviation is significantly increased; the threshold of 0.4 is determined by simulating the difference in the statistical characteristics of the signal under the crack state, reflecting the significant change in the irregularity of the signal;

[0135]

[0136] In the formula, The mean of the simulated data or the measured data after synchronous processing; This refers to either the simulated data or the measured data after synchronous processing. This represents the total number of sampling points; This is the indicator value for the sampling point;

[0137]

[0138] In the formula, The standard deviation of the simulated data or the measured data after synchronous processing;

[0139]

[0140] In the formula, The peak value of the simulation data or the measured data after synchronous processing;

[0141] Frequency domain characteristics include the energy proportion of different frequency bands:

[0142] The formula is:

[0143]

[0144] In the formula, Frequency point The proportion of energy; For discrete data sequences that have undergone Fast Fourier Transform (FFT) on synchronously processed simulation data or synchronously processed measured data; This represents the total number of sampling points; This is a frequency point indication value;

[0145]

[0146] In the formula, For time series representation of simulated data or measured data after synchronous processing; This represents the frequency domain of the simulated data or measured data after synchronous processing; j is the imaginary unit; e is a natural number.

[0147] Loose bolts: Check for new frequency components at (1-3) Hz;

[0148] Excessive tilt: Check if the energy percentage in the high-frequency region (5-10 Hz) has increased;

[0149] The tilting causes the first-order bending mode frequency of the support component to shift to (5-10) Hz, and the increased energy proportion reflects the activation of the anomalous mode.

[0150] Foundation settlement: Check if the proportion of energy below 0.5Hz in the low-frequency range increases by 40%;

[0151] The overall sway mode during foundation settlement of a building often has a frequency of less than 0.5 Hz; acoustic emission testing should focus on high-frequency signals in the (10-20) kHz band, where the stress waves generated by crack opening and closing are concentrated.

[0152] Cracks: Check if the energy percentage in the high-frequency region (10-20 kHz) has increased significantly;

[0153] Multi-domain feature fusion is performed to obtain a comprehensive feature vector, as shown in the formula:

[0154]

[0155] In the formula, This is a comprehensive feature vector; The characteristic component of bolt loosening is denoted as , where =[ (1-3) Hz new frequency marking, 0,0,0,0]; For the over-sloping characteristic components, where, =[ ,0, ,0,0,0]( (The change in energy percentage is 5-10 Hz). The basic settlement characteristic components, among which, =[ ,0,0, ,0,0]( (This refers to the change in energy percentage below 0.5Hz). Let be the characteristic components of the crack, where =[ ,0,0,0, [0] (x is acceleration or displacement) (This represents the energy percentage change over a range of 10-20 kHz); T is the transpose sign.

[0156] S3: Using a fault diagnosis model, perform fault diagnosis on the comprehensive feature vector to obtain the corresponding real-time fault diagnosis results, including the following steps:

[0157] S3-1: Use the pre-trained RF structure in the key feature filtering module of the fault diagnosis model to extract several key feature components from the comprehensive feature vector;

[0158] S3-2: Based on the dynamic attention weights, the attention weight module of the fault diagnosis model is used to weight and concatenate several key feature components to obtain the concatenated features.

[0159] S3-3: Based on the splicing characteristics, the fault diagnosis module of the fault diagnosis model is used to perform fault diagnosis prediction and obtain the corresponding real-time fault diagnosis results.

[0160] Example 2:

[0161] like Figure 2 As shown, this embodiment provides a fault diagnosis system based on simulation model optimization and multi-domain feature fusion to implement a fault diagnosis method. The system includes a model building and optimization unit, a multi-domain feature fusion unit, and a fault diagnosis unit connected in sequence.

[0162] The model building and optimization unit is used to build the initial tower simulation model and the fault diagnosis model of multi-domain feature fusion, and to optimize and condense the initial tower simulation model to obtain the condensed tower simulation model.

[0163] The multi-domain feature fusion unit is used to generate simulation data of the tower using the condensed tower simulation model, and to fuse the simulation data and the corresponding measured data in multiple domains to obtain a comprehensive feature vector.

[0164] The fault diagnosis unit is used to perform fault diagnosis on the comprehensive feature vector using the fault diagnosis model, and obtain the corresponding real-time fault diagnosis results.

[0165] This invention discloses a fault diagnosis method and system based on simulation model optimization and multi-domain feature fusion. The comprehensive feature vector obtained through multi-domain feature fusion and the fault diagnosis model constructed based on machine learning algorithms achieve high accuracy and can uncover deep relationships between data and faults, improving diagnostic accuracy. The initial simulation model is optimized through dynamic convergence and multi-objective optimization algorithms, enabling the optimized tower simulation model to more accurately reflect the dynamic characteristics of the actual tower, improving the reliability of simulation data, avoiding large errors between simulation model data and measured data, reducing false alarms, and improving effectiveness. Combined with automated fault diagnosis, no manual intervention is required, improving efficiency. In addition to utilizing traditional vibration acceleration data, it also integrates multi-domain information such as displacement and tilt angle, which can more comprehensively reflect the structural state of the tower, improving the comprehensiveness of feature information and the robustness of diagnosis. Furthermore, by comparing data in the time and frequency domains and fusing the time-frequency domain features of multi-source signals, most of the important information reflecting the equipment's operating status is retained, avoiding the omission of some hidden faults due to insufficient fault judgment criteria.

[0166] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault diagnosis method based on simulation model optimization and multi-domain feature fusion, characterized in that: Includes the following steps: An initial tower simulation model and a fault diagnosis model based on multi-domain feature fusion were constructed, and the initial tower simulation model was optimized to obtain the optimized tower simulation model. The fault diagnosis model is constructed based on the RF-Attention-SVM algorithm, and the fault diagnosis model includes a key feature screening module constructed based on the RF algorithm, an attention weight module constructed based on the Attention mechanism, and a fault diagnosis module constructed based on the SVM algorithm, which are connected in sequence. The optimization of the model parameters of the initial tower simulation model to obtain the optimized tower simulation model includes the following steps: The dynamic condensation method is used to reduce the degrees of freedom of the initial tower simulation model, resulting in a condensed tower simulation model. The sensitivity of the condensed tower simulation model is calculated. If the obtained sensitivity is greater than the threshold, the optimized tower simulation model is output; otherwise, proceed to the next step. The model parameters of the condensed tower simulation model are optimized using a multi-objective optimization algorithm to obtain the optimized tower simulation model, including the following steps: Using the condensed tower simulation model, a simulation sample of the tower is generated, and frequency domain analysis is performed to obtain the simulation natural frequency and simulation mode shape of the simulation sample. Using sensors, the actual samples corresponding to the simulated samples are collected, and the EFDD algorithm is used to obtain the actual natural frequencies and actual mode shapes of the actual samples. Define the objective function based on the simulated natural frequencies, simulated mode shapes, measured natural frequencies, and measured mode shapes; Based on the objective function, the fitness function of the MOSGA algorithm is set, and iterative optimization is performed based on the fitness function to obtain the optimal model parameters; Based on the optimal model parameters, the simulation model of the condensed tower is optimized to obtain the optimized tower simulation model. Using the optimized tower simulation model, simulation data of the tower is generated, and the simulation data and the corresponding measured data are fused together using multi-domain features to obtain a comprehensive feature vector. Using a fault diagnosis model, fault diagnosis is performed on the comprehensive feature vector to obtain the corresponding real-time fault diagnosis results.

2. The fault diagnosis method based on simulation model optimization and multi-domain feature fusion according to claim 1, characterized in that: The initial tower simulation model and the multi-domain feature fusion fault diagnosis model are constructed, and the initial tower simulation model is optimized to obtain the optimized tower simulation model. The process includes the following steps: Based on the basic simulation information and 3D scanning data of the tower, an initial tower simulation model is constructed using 3D simulation technology. Using machine learning algorithms, an initial fault diagnosis model is constructed, and a training sample set containing different working conditions is input for optimization training to obtain the final fault diagnosis model. The model parameters of the initial tower simulation model are optimized to obtain the optimized tower simulation model.

3. The fault diagnosis method based on simulation model optimization and multi-domain feature fusion according to claim 2, characterized in that: The simulation data includes the simulation displacement, simulation acceleration, and simulation tilt angle data output by the optimized tower simulation model; The measured data includes the measured displacement, measured acceleration, and measured tilt angle data of the corresponding position of the tower collected by the sensor.

4. The fault diagnosis method based on simulation model optimization and multi-domain feature fusion according to claim 3, characterized in that: Based on the objective function, the fitness function of the MOSGA algorithm is defined, and iterative optimization is performed based on the fitness function to obtain the optimal model parameters, including the following steps: Based on the objective function, the fitness function of the MOSGA algorithm is set, and the initial model parameters are encoded into individual vectors of the MOSGA algorithm. Based on the individual vectors, a number of initial solutions are generated using a chaotic mapping sequence algorithm; each initial solution corresponds to an initial set of model parameters. Based on the fitness function, the MOSGA algorithm is used to iteratively optimize several initial solutions and retain the best individual in each iteration; If the number of iterations is greater than or equal to the iteration threshold or the fitness value of the best individual is less than the fitness threshold, then the best individual will be output as the optimal solution. Decoding the individual vectors of the optimal solution yields the optimal model parameters for the condensed tower simulation model.

5. The fault diagnosis method based on simulation model optimization and multi-domain feature fusion according to claim 4, characterized in that: Using the optimized tower simulation model, simulation data for the tower is generated. The simulation data and corresponding measured data are then fused using multi-domain features to obtain a comprehensive feature vector. This process includes the following steps: The optimized tower simulation model is used to generate tower simulation data, and sensors are used to collect the corresponding measured data. The simulation data and the corresponding measured data are synchronized by adjusting the time step and the number of sampling points to obtain synchronized simulation data and synchronized measured data. The time-domain and frequency-domain features of the simulated data and measured data after synchronous processing are extracted, and multi-domain features are fused to obtain a comprehensive feature vector.

6. The fault diagnosis method based on simulation model optimization and multi-domain feature fusion according to claim 5, characterized in that: Using a fault diagnosis model, fault diagnosis is performed on the comprehensive feature vector to obtain the corresponding real-time fault diagnosis results, including the following steps: Using the pre-trained RF structure in the key feature filtering module of the fault diagnosis model, several key feature components are extracted from the comprehensive feature vector. Based on the dynamic attention weights, the attention weight module of the fault diagnosis model is used to weight and concatenate several key feature components to obtain the concatenated features. Based on the splicing characteristics, the fault diagnosis module of the fault diagnosis model is used to perform fault diagnosis prediction and obtain the corresponding real-time fault diagnosis results.

7. A fault diagnosis system based on simulation model optimization and multi-domain feature fusion, used to implement the fault diagnosis method as described in any one of claims 1-6, characterized in that: The system includes a model building and optimization unit, a multi-domain feature fusion unit, and a fault diagnosis unit connected in sequence.

Images