Bolt looseness LightGBM diagnosis method and system based on ultrasonic signal feature optimization

Abstract

The invention relates to the technical field of data processing, and provides an ultrasonic signal feature optimized bolt looseness LightGBM diagnosis method, which comprises the following steps: S1, carrying out ultrasonic signal feature optimization preprocessing and signal acquisition based on NGO-ICEEMDAN-MPE algorithm combination, applying pulse excitation and receiving echo signals, optimizing ICEEMDAN parameters by using an NGO algorithm, and screening signal components in combination with MPE; s2, feature extraction and LightGBM algorithm state recognition are carried out, time domain and frequency domain multi-dimensional features of echo signals are extracted and fused, data sets are divided after Min-Max normalization processing, and a LightGBM model hyper-parameter and an unknown fault discrimination threshold are optimized through an SOA algorithm; and S3, result visualization and storage are carried out, diagnosis results, feature vectors and model parameters are stored in a structured manner, bolt state imaging, feature waveform display, historical trend analysis and alarm prompt are realized through a Python GUI interface, and diagnosis report export is supported. Rapid detection of the overall structure of the bolt is achieved, the detection efficiency of looseness of the bolt of the power transmission tower is improved, and the detection success rate is increased.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for diagnosing bolt loosening using LightGBM based on optimized ultrasonic signal characteristics. Background Technology

[0002] ICEEMDAN (Improved Adaptive Noise-Complete Ensemble Empirical Mode Decomposition), as an optimized and upgraded version of CEEMDAN (Adaptive Noise-Complete Ensemble Empirical Mode Decomposition), innovatively introduces white noise of specific intensities and an improved iterative calculation method. This fundamentally reduces reconstruction errors after signal decomposition, ensuring the completeness of the decomposition process and signal integrity. Compared to traditional EMD (Empirical Mode Decomposition) and earlier versions of CEEMDAN, its ability to suppress mode aliasing and control endpoint effects when processing non-stationary and nonlinear signals is significantly improved, making it an important technique in complex signal processing. However, ICEEMDAN's performance is highly sensitive to its core parameters (especially the magnitude and number of times white noise is added): if the white noise amplitude is set too small, it is difficult to effectively break the coupling relationship between modal components in the signal, easily leading to insufficient decomposition, with the useful signal and noise still intertwined; if the amplitude is too large or the number of additions is too high, it may introduce additional redundant noise, causing over-decomposition of the signal and destroying the key feature information of the original signal. Unfortunately, the determination of these key parameters in the industry still largely relies on the past experience of engineers and technicians and repeated trial and error to set them manually. There is a lack of an adaptive parameter optimization mechanism that can dynamically adjust according to the characteristics of the signal itself. This greatly limits the stable application and performance of ICEEMDAN in complex real-world scenarios (such as ultrasonic signal processing in strong interference environments).

[0003] Although the technical approach of combining ICEEMDAN with MPE (Multi-Scale Permutation Entropy) has been proven to be an effective signal denoising path—that is, first using ICEEMDAN to decompose the original complex signal into a series of intrinsic mode functions (IMFs) and residual components, and then using MPE to quantify the randomness and complexity of each IMF component, screen out the dominant signal components, and reconstruct the denoised signal—its performance is still deeply limited by the initial decomposition quality of ICEEMDAN. If the parameters of ICEEMDAN are not configured properly, the useful signal and noise in the decomposed IMF components cannot be effectively separated. Even if multi-scale entropy value screening is performed by MPE later, it is difficult to completely remove noise interference and retain key features, ultimately affecting the overall signal processing effect. This core technical pain point provides a clear and urgent technical motivation for introducing professional optimization algorithms to find the best parameter combination for ICEEMDAN. Through the global search capability of the optimization algorithm, ICEEMDAN parameters that are suitable for the current signal characteristics can be accurately matched, maximizing the component screening efficiency of MPE, and ultimately achieving a significant improvement in the signal-to-noise ratio of the reconstructed signal. This forms a closed-loop, adaptive, and high-performance signal optimization solution, fundamentally solving the core problems of traditional methods such as parameter dependence on experience, unstable denoising effect, and weak anti-interference capability.

[0004] Returning to the actual operation and maintenance scenarios of power systems, transmission lines, as the core carriers of power grid energy transmission, generally employ tall tower structures to support conductors. The overall structural safety and operational stability of these towers largely depend on the tightness of the bolt connections. Bolts, as key connection nodes among the various components of the tower, play a crucial role in transferring loads and distributing stress. During long-term service, tower bolts are highly susceptible to loosening, fatigue damage, and corrosion due to a combination of factors, including complex climatic conditions (such as high-temperature exposure, low-temperature freezing, rain erosion, and strong wind disturbance), continuous vibration loads (such as conductor vibration in the wind and equipment vibration), thermal expansion and contraction caused by alternating temperature changes, and natural aging and corrosion of materials. If these hazards are not detected and addressed in a timely manner, they will gradually exacerbate the stress imbalance of the tower, triggering a chain reaction of abnormal changes in conductor sag and accelerated wear at component connections. In severe cases, this can lead to tower structural instability or even complete tower collapse, directly threatening the safe and stable operation of the power grid and causing major power accidents such as large-scale power outages.

[0005] However, current methods for detecting loose bolts on transmission towers still primarily rely on traditional manual inspections, depending mainly on visual observation and mechanical tapping by inspectors. This inspection model has several insurmountable drawbacks: First, transmission lines are often widely distributed in complex geographical environments such as mountains, hills, forests, and remote suburbs. Inspection routes are rugged and transportation is inconvenient. Each inspection requires a significant investment of manpower, resources, and time. Furthermore, the inspection work is highly repetitive and has a long cycle, making it difficult to adapt to the efficient operation and maintenance requirements after the large-scale development of the power grid. Second, the methods of manual tapping or visual inspection... This method is a qualitative inspection and cannot quantitatively assess the bolt tightness. It can only determine whether there is obvious looseness. It lacks the ability to identify early hidden dangers such as slight looseness and insufficient preload, which can easily lead to missed detection of hidden dangers and missing the best maintenance opportunity. Thirdly, the tower height is generally high, and the inspection process often requires operators to climb to high places or use drones for assistance. Operators are exposed to risks such as falling from heights, electromagnetic radiation and severe weather (such as rainstorms, strong winds and high temperatures) for a long time, which poses significant safety hazards and a great threat to the personal safety of operation and maintenance personnel. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a LightGBM diagnostic method and system for bolt loosening based on optimized ultrasonic signal characteristics. This invention's LightGBM diagnostic method for bolt loosening, based on NGO-ICEEMDAN-MPE ultrasonic signal characteristic optimization, designs an ultrasonic testing scheme suitable for transmission tower bolt structures, achieving rapid detection of the entire bolt structure, improving the detection efficiency and success rate of transmission tower bolt loosening, and further enhancing equipment operability.

[0007] The above-mentioned objective of this invention is achieved through the following technical solutions:

[0008] A LightGBM diagnostic method for bolt loosening based on optimized ultrasound signal characteristics includes the following steps:

[0009] S1: Perform ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms. Apply pulse excitation through a piezoelectric transducer and receive echo signals. Optimize ICEEMDAN parameters using the NGO algorithm and filter signal components using MPE to achieve ultrasonic signal denoising, energy quantization and optimization preprocessing.

[0010] S2: Perform feature extraction and LightGBM algorithm state recognition. Extract and fuse multi-dimensional features of the echo signal in the time and frequency domains. After Min-Max normalization, divide the dataset. Optimize the hyperparameters of the LightGBM model and the unknown fault discrimination threshold through the SOA algorithm to achieve bolt fastening state classification and unknown fault identification.

[0011] S3: Visualizes and stores results, provides structured storage of diagnostic results, feature vectors, and model parameters, and enables bolt condition imaging, feature waveform display, historical trend analysis, and alarm prompts through a Python GUI interface. It also supports exporting diagnostic reports.

[0012] Further, in step S1, ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms are performed. Pulse excitation is applied through a piezoelectric transducer and echo signals are received. The NGO algorithm is used to optimize the ICEEMDAN parameters, and the MPE algorithm is used to filter signal components, thereby achieving ultrasonic signal denoising, energy quantization, and optimization preprocessing. Specifically:

[0013] S11: A pulse excitation signal is applied axially to the bolt using a piezoelectric transducer, while echo signals reflected from the thread interface, nut end face, and tower components are received. The ultrasonic wave propagates in a straight line along the bolt. The loss is extremely low when the bolt is fastened. When there are defects or loosening, the waveform characteristics change and the energy is attenuated. The instrument identifies bolt abnormalities by analyzing the received waveform changes. At the same time, the original noisy ultrasonic echo signal is input to initialize the NGO algorithm, including the population size, maximum number of iterations, and search range of optimization parameters.

[0014] S12: When ultrasound propagates in an elastic medium, the kinetic energy generated by the vibration of particles and the deformation potential energy together achieve energy transfer.

[0015] S121: Define the average energy density of ultrasound within a vibration cycle as the average energy density. Perform noise reduction, filtering and normalization on the acquired signal, and use window function and bandpass filtering to suppress environmental noise.

[0016] S122: Define a basic signal x and add zero-mean, unit-variance white noise to it to generate a series of constructed signals x. (i) For each x (i) Calculate the local mean, and take the average of all local means to obtain the first residual component r. i And calculate the first modal component imf1; iteratively calculate the subsequent modal components imf k The iteration continues until the residual components satisfy the termination condition or the modal components are less than the first three local extrema.

[0017] S123: For each individual in the population, its position represents a set of ICEEMDAN parameters. Using the parameters represented by the current eagle's position, ICEEMDAN decomposition is performed on the original signal to obtain a series of IMF components and a residual component. The multi-scale permutation entropy of each IMF component is calculated, and an entropy threshold is set. Typically, IMF components with higher MPE values ​​are considered noise-dominant components. Based on the MPE values, all IMFs are divided into two categories: a set of signal-dominant IMFs and a set of noise-dominant IMFs. The set of signal-dominant IMFs is summed to reconstruct the denoised signal, and the fitness value is calculated to evaluate the quality of the current parameters.

[0018] S13: Based on the NGO's hunting strategy, which includes the first stage being attack and the second stage being pursuit, update the positions of all individual goshawks and retain the currently globally optimal individual, that is, the goshawk with the best parameters; the position of the northern goshawk population is represented by matrix X, and the objective function value is represented by matrix F.

[0019] S14: Repeat steps S12 and S13 until the maximum number of iterations T is reached or other convergence conditions are met. Finally, the NGO algorithm outputs the globally optimal combination of parameters.

[0020] S15: Using the optimal parameters found by NGO, perform a final ICEEMDAN decomposition on the original signal, calculate the multiscale permutation entropy (MPE) of each IMF, distinguish between signal and noise IMFs using the same thresholding strategy as before, and reconstruct the final optimized ultrasound signal using the signal IMFs.

[0021] In step S121, the formula for calculating the average energy density of ultrasound is:

[0022]

[0023] in, ρ is the average energy density of the ultrasonic wave within one cycle; ρ is the density of the material; A is the amplitude of the particle vibration. It is the angular frequency;

[0024] According to the energy density formula, under the premise that the material density and frequency are constant, the ultrasonic energy E is characterized by the sum of the squares of the signal amplitude, and the calculation formula is as follows:

[0025]

[0026] Where E is the energy of the sampled signal; N is the length of the sampled signal; v i Let be the amplitude of the sampled signal at the i-th point;

[0027] The ultrasonic energy decreases as bolt loosening worsens. Based on the calculated energy value E, the energy attenuation rate η is defined to quantitatively characterize the degree of loosening, and its expression is as follows:

[0028]

[0029] Where E0 is the energy of the ultrasonic response signal under the standard state of the bolt; E is the energy of the ultrasonic response signal under the loose state of the bolt. The greater the degree of bolt loosening, the greater the decrease in ultrasonic signal energy. When the ultrasonic signal energy decreases, the ultrasonic signal energy attenuation rate will increase. Therefore, the degree of bolt loosening can be judged by the ultrasonic signal energy attenuation rate.

[0030] In step S122, a basic signal x is set and added to it to generate a series of construct signals x. (i) , represented as:

[0031] x (i) =x+β0E1(w (i) ), (i = 1, 2, ..., N)

[0032] Where, x (i) Let be the i-th constructed signal, N be the total number of constructed signals, β0 be the initial noise standard deviation of the signal, and w (i) It is the added zero-mean unit variance white noise, and E1(·) represents the operation of calculating the first inner membrane state function (IMF) of the signal.

[0033] Calculate the local mean and average it for each x. (i) Calculate the local means, and then take the average of all local means to obtain the first residual component r. i :

[0034]

[0035] Here, M(·) represents the local mean of the signal, and I is the number of constructed signals;

[0036] The first modal component, imf1, is obtained by subtracting the first residual component r1 from the original signal x.

[0037] imf1 = x - r1

[0038] Iterative calculation of subsequent modal components (IMF) k For the k-th modal component, k≥2, use the residual r from the previous iteration. k-1 Subtract the residual r this time k :

[0039] imf k =r k-1 -r k

[0040] Furthermore, r k The calculation method is as follows:

[0041]

[0042] Among them, E k (·) represents the imf operation corresponding to the k-th calculation, β k-1 Let be the noise standard deviation of the (k-1)th iteration;

[0043] The iteration terminates when the residual component meets a certain termination condition or when the modal component is less than the first three local extrema.

[0044] In step S13, the initialization of the position of the Northern Goshawk is represented by matrix X, as shown below:

[0045]

[0046] In the formula, X represents the population of the Northern Goshawk; X i Indicates the current position of the i-th Northern Goshawk; X i,j Let represent the position of the i-th Northern Goshawk in the j-th dimension; N represents the population size; m represents the dimension of the problem. The objective function value for the Northern Goshawk can be calculated by solving the objective function during the problem-solving process, and is represented by matrix F.

[0047]

[0048] Among them, F i Let be the objective function value for i northern eagles.

[0049] Further, in step S2, feature extraction and LightGBM algorithm state recognition are performed. Multi-dimensional features of the echo signal in the time and frequency domains are extracted and fused. After Min-Max normalization, the dataset is divided. The LightGBM model hyperparameters and unknown fault discrimination threshold are optimized using the SOA algorithm to achieve bolt tightening state classification and unknown fault identification. Specifically:

[0050] S21: Extract features of the echo signal, including time delay, peak amplitude, envelope energy, and spectral centroid, and calculate the waveform difference of multiple detections to assess the tightening trend; at the same time, use the ultrasonic signal feature dataset extracted after optimization by NGO-ICEEMDAN-MPE. The dataset contains bolt status labels of known categories, including normal, slightly loose, and severely loose. Divide the dataset into training set, validation set, and test set.

[0051] S22: Extract and fuse time-domain, frequency-domain, and wavelet-domain features from the original vibration signal of the rolling bearing, reducing the dimension of the original vibration signal to 31 dimensions, effectively removing redundant information, reducing training time, and improving model convergence; randomly generate an initial forest, where the position of each tree represents a set of random LightGBM hyperparameters;

[0052] S23: After feature fusion and before dataset partitioning, Min-Max normalization is used to preprocess the feature data; the known fault category data after feature extraction and fusion are divided into training set, validation set and test set. All unknown new faults that have not appeared and have not been modeled are assigned to the test set and do not participate in the model training process, but are only used to verify the model recognition effect; in each seasonal iteration of SOA, the update rate is calculated according to the formula, new trees are randomly generated in the forest to increase population diversity, the trees are sorted according to their strength and core trees are selected, the fitness of neighboring trees is adjusted, strong trees are selected for planting and weak trees are eliminated;

[0053] S24: Train the LightGBM model using a training set with known categories, establish a known fault identification model, construct an unknown class discrimination mechanism, assume any class in the training set as the unknown class, and the remaining classes as the known classes, retrain the classifier with training data, and optimize the unknown class discrimination threshold T using the validation set and SOA; evaluate the quality of the tree in each generation of SOA, track the tree with the highest historical fitness, and output the globally optimal hyperparameter combination after the iteration terminates;

[0054] S25: Input the test sample into the trained LightGBM model to obtain the probability value of each sample. If the highest confidence ρ(i) of the test sample is less than the threshold T, the sample is classified as an unknown new fault. Otherwise, it is classified as the known fault category corresponding to i. Integrate model interpretability technology. Retrain the model using the optimal hyperparameters found by SOA, and deploy it to the actual detection system after evaluation on the test set.

[0055] In step S21, the time delay feature τ is calculated using a cross-correlation function to determine the time lag between the currently detected ultrasonic echo signal y(t) and the reference state signal x(t), which is then used as the time delay feature τ. The calculation formula is as follows:

[0056]

[0057] in, The transit time variation of ultrasonic waves propagating inside the bolt is characterized by its correlation with the internal stress state and sound velocity variation of the bolt. y(t-τ) is the ultrasonic echo signal based on time lag at the current time.

[0058] Peak amplitude A peakThe maximum absolute value in the discrete sampled signal is obtained as the peak amplitude feature. Let the sampled signal sequence be v. i ,but:

[0059]

[0060] Where N is the number of sampling points, this feature reflects the strongest reflection intensity of the echo signal, and the air gap formed by loose bolts will cause this amplitude to decrease significantly;

[0061] The envelope energy E is based on the principle of ultrasonic energy frequency density. Under the premise that the material density and frequency are constant, the ultrasonic energy is characterized by the sum of the squares of the signal amplitude. The calculation formula is as follows:

[0062]

[0063] Where E is the energy of the sampled signal, N is the length of the sampled signal, and v i Let be the amplitude of the sampled signal at the i-th point;

[0064] The frequency domain amplitude spectrum |S(f)| is obtained by performing a Fourier transform on the time-domain signal, and its spectral centroid f is calculated. c To reflect the shift in frequency distribution:

[0065]

[0066] Among them, f k Let S(f) be the k-th frequency point, and K be the total number of frequency points. k | represents the frequency domain amplitude value corresponding to the k-th frequency point;

[0067] Step S23 also includes: for each seasonal iteration, according to formula P r =P m ax-(y / Y) * (P m ax-P m in) Calculate the update rate P r In the formula, y is the current iteration number, Y is the maximum iteration number, and Pmax and Pmin are the maximum and minimum update rates, respectively; P is randomly generated in the forest. r * Add A new trees to the forest to increase population diversity, where A is a random number; sort the forest according to the strength of the trees and select the top N. c A strong tree is selected as the core tree; each core tree influences the growth position of its neighboring trees, simulating a competition process. The fitness of the neighboring trees is adjusted according to the competition intensity Λ. The top A strong trees are selected, and A = ψ(P). s *N) is used for sowing, that is, replicating these excellent solutions in the hope of producing better offspring; based on the resistance rate P -W Remove the weakest P in the forestw * Use N trees to simulate natural selection and eliminate undesirable solutions;

[0068] In step S24, the specific implementation process and formula derivation are as follows:

[0069] Based on the gradient boosting decision tree principle, a strong classifier is formed by stacking multiple weak classifiers. Let the number of iterations be Y, and the current decision tree of the weak classifiers be f. i (x), whose corresponding weight is h i Then the strong classifier F is constructed. Y (X) is represented as:

[0070]

[0071] Where X is the input feature vector, f0(x) is the initial model, and f i (x) represents the current weak classifier. The model segments the features using a histogram optimization algorithm and grows the decision tree using a depth-limited leaf-wise strategy to prevent overfitting.

[0072] When there are d known fault categories, the LightGBM model outputs a confidence vector P for each of the d categories for each test sample:

[0073] P = [ρ1, ρ2, ..., ρ d ]

[0074] Where, ρ d The model represents the confidence level of classifying a sample into class d, and the highest confidence level ρ of the sample to be tested is calculated. j :

[0075] ρ j =max{ρ1,ρ2,…,ρ d}

[0076] The discrimination threshold is set to T, and the discrimination logic is as follows: If ρ j If ρ > T, then the sample is determined to be of a known category and classified into the j-th category; if ρ j If the value is less than or equal to T, then the sample is determined to be an unknown new fault.

[0077] Furthermore, in step S3, the results are visualized and stored, and the diagnostic results, feature vectors, and model parameters are stored in a structured manner, specifically as follows:

[0078] S31: Structured storage of diagnostic results: The bolt status output by the LightGBM model, including the tightness, looseness level, unknown fault markers, identification results, corresponding signal feature vectors, model confidence probabilities, and timestamp information, is automatically stored in a structured database or a data file of a specified format; at the same time, the model version and parameters used for each diagnosis are recorded to ensure data traceability;

[0079] S32: Bolt Status Visualization Imaging: Based on stored diagnostic results, draw a two-dimensional or three-dimensional layout diagram of the bolt group in the PYTHON GUI interface, and intuitively mark the real-time or historical status of each bolt in the layout diagram with different colors and number labels.

[0080] S33: Feature and Confidence Visualization: Provides a detailed analysis view of a single bolt in the interface. Select a specific bolt to plot its original ultrasonic signal, the waveform of the key IMF components optimized by NGO-ICEEMDAN-MPE, and the time-domain and frequency-domain feature maps during the feature extraction process.

[0081] S34: Historical Trend Analysis and Alarm Log: The system automatically generates and displays the historical trend chart of the status of key bolts for predictive maintenance; when the system detects a loose state or an unknown fault, it automatically triggers an audible and visual alarm and records the event time, location, specific description, and suggested measures in the dedicated alarm log list on the interface;

[0082] S35: Report Generation and Data Export: Provides a one-click function to generate diagnostic reports, allowing users to export the current visualization interface, raw data, feature data, and diagnostic reports in common formats to a specified local path for easy archiving or reporting later.

[0083] Furthermore, in step S3, bolt condition imaging, characteristic waveform display, historical trend analysis, and alarm prompts are implemented through a Python GUI interface, supporting the export of diagnostic reports. Specifically:

[0084] This software development platform uses PYTHON GUI, a graphical user interface design tool. This software can display images of bolt loosening and analyze the bolt loosening state using image recognition algorithms.

[0085] An ultrasonic signal feature-optimized LightGBM diagnostic system for bolt loosening, used to perform the ultrasonic signal feature optimization method for bolt loosening as described above, comprises:

[0086] The ultrasonic signal preprocessing module is used to perform ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms. It applies pulse excitation through a piezoelectric transducer and receives echo signals. It optimizes the ICEEMDAN parameters using the NGO algorithm and filters signal components using MPE to achieve ultrasonic signal denoising, energy quantization and optimization preprocessing.

[0087] The feature state recognition module is used for feature extraction and LightGBM algorithm state recognition. It extracts and fuses multi-dimensional features of the echo signal in the time and frequency domains, divides the dataset after Min-Max normalization, and optimizes the hyperparameters of the LightGBM model and the unknown fault discrimination threshold through the SOA algorithm to realize bolt fastening state classification and unknown fault recognition.

[0088] The results visualization and storage module is used for results visualization and storage. It stores diagnostic results, feature vectors and model parameters in a structured manner. It realizes bolt status imaging, feature waveform display, historical trend analysis and alarm prompts through a Python GUI interface, and supports the export of diagnostic reports.

[0089] A computer device, characterized in that it includes a memory and one or more processors, wherein the memory stores computer code, and when the computer code is executed by the one or more processors, causes the one or more processors to perform the method as described above.

[0090] A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer code, which, when executed, is performed as described above.

[0091] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0092] (1) Upgrading and optimizing the detection method: A portable detection system is built by using ultrasonic sensing and intelligent signal analysis technology to replace the traditional manual knocking detection method. It can quickly complete the detection and intelligent identification of the status of tower bolts. At the same time, by analyzing the time domain, frequency domain and envelope characteristics of ultrasonic waves, the bolt tightness can be quantified and slight loosening trends can be identified, realizing the transformation from qualitative judgment of "whether it is loose" to quantitative assessment of "the degree of looseness".

[0093] (2) Supporting the construction of a digital operation and maintenance system: The integration of ultrasonic detection technology, Internet of Things data acquisition technology and artificial intelligence algorithms provides key technical support for the construction of a digital transmission line operation and maintenance system.

[0094] (3) Convenient and efficient testing: This method is simple in principle and easy to operate. It has good directness and repeatability. Testing can be carried out under normal operating conditions of bolts. The testing cycle can be flexibly arranged according to the actual operating conditions of the equipment, which makes it easy to grasp the bolt deformation trend and detect loosening hazards in advance.

[0095] (4) Improve operation and maintenance work and system stability: It effectively improves the efficiency of equipment testing, enhances the portability of testing equipment and reduces the difficulty of operation, which is of great practical significance for ensuring the long-term stable operation of the power system. Attached Figure Description

[0096] Figure 1 This is an overall flowchart of the LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics, as described in this invention.

[0097] Figure 2 This is a schematic diagram of the ultrasonic wave propagation process of the present invention;

[0098] Figure 3 This is a schematic diagram of the main interface of the system of the present invention;

[0099] Figure 4 This is a schematic diagram of the data detection and import interface of the present invention;

[0100] Figure 5 This is a schematic diagram of the data display interface of the present invention;

[0101] Figure 6 This is an overall structural diagram of the LightGBM diagnostic system for bolt loosening, which optimizes the ultrasonic signal characteristics of this invention. Detailed Implementation

[0102] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0103] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0104] This invention presents a Light GBM diagnostic method for bolt loosening based on ultrasonic signal feature optimization using the NGO-ICEEMDAN-MPE algorithm. The method employs the NGO-ICEEMDAN-MPE algorithm to denoise the echo signal. The NGO-ICEEMDAN algorithm is studied to decompose the acquired ultrasonic signal, and key parameters in the ICEEMDAN model are optimized using the NGO algorithm. Methods for extracting multiple feature parameters from the IMF component are investigated and input into the MPE model to achieve ultrasonic signal denoising optimization. A bolt loosening identification algorithm based on SOA-LightGBM is also developed, integrating the optimization algorithm (SOA) and the lightweight gradient boosting framework (LightGBM). Through efficient feature learning and classification, it achieves accurate identification of bolt loosening conditions, improving the maintenance efficiency and safety of transmission towers and ensuring stable system operation.

[0105] The purpose of this invention is:

[0106] (1) The LightGBM diagnostic method for bolt loosening based on ultrasonic signal feature optimization of NGO-ICEEMDAN-MPE in this invention utilizes ultrasonic sensing and intelligent signal analysis technology to construct a portable detection system, replacing the traditional manual tapping method, and realizing rapid detection and intelligent identification of the status of tower bolts. Through ultrasonic time domain, frequency domain and envelope feature analysis, the bolt tightness is quantified, and slight loosening trends are identified, realizing the transformation from "whether it is loose" to "the degree of loosening".

[0107] (2) The new method for diagnosing bolt loosening using LightGBM based on ultrasonic signal feature optimization of NGO-ICEEMDAN-MPE integrates ultrasonic detection technology, Internet of Things data acquisition and artificial intelligence algorithm to provide technical support for building a digital transmission line operation and maintenance system.

[0108] The following is an illustration through specific examples:

[0109] First Embodiment

[0110] like Figure 1 As shown, this embodiment provides a LightGBM diagnostic method for bolt loosening based on optimized ultrasound signal characteristics, including the following steps:

[0111] S1: Perform ultrasound signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms. Apply pulse excitation through a piezoelectric transducer and receive echo signals. Optimize ICEEMDAN parameters using the NGO algorithm and filter signal components using MPE to achieve ultrasound signal denoising, energy quantization and optimization preprocessing.

[0112] Step S1 combines the ICEEMDAN, MPE, and NGO algorithms. ICEEMDAN is an advanced signal decomposition method that adaptively decomposes complex, non-stationary, nonlinear raw ultrasonic echo signals into a series of intrinsic mode functions (IMFs) ranging from high to low frequencies. Compared to traditional EMD, it effectively suppresses mode aliasing and endpoint effects, resulting in more complete decomposition and less residual noise. The MPE algorithm is a metric for measuring the complexity and randomness of time series data. It analyzes the entropy of signals at multiple scales. The NGO algorithm is a novel metaheuristic optimization algorithm that simulates the hunting behavior of a northern eagle (identifying, attacking, and chasing prey). Combining these three algorithms allows for the construction of an adaptive and automated ultrasonic signal optimization process. The NGO algorithm overcomes the empirical dependence of ICEEMDAN parameters, while MPE provides a clear and quantifiable objective function for the NGO optimization process (i.e., maximizing the signal-to-noise ratio of the reconstructed signal or maximizing the sum of MPE values ​​for the noise components). This combination ensures that the signal denoising effect reaches its optimal level under the current conditions.

[0113] In this embodiment, in step S1, ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms are performed. Pulse excitation is applied through a piezoelectric transducer and echo signals are received. The NGO algorithm is used to optimize the ICEEMDAN parameters, and the MPE algorithm is used to filter signal components, thereby achieving ultrasonic signal denoising, energy quantization, and optimization preprocessing. Specifically:

[0114] S11: A piezoelectric transducer applies a pulse excitation signal axially to the bolt, while simultaneously receiving echo signals reflected from the thread interface, nut end face, and tower components. The ultrasonic waves propagate in a straight line along the bolt, exhibiting extremely low loss under tight conditions. When defects or loosening occur, the waveform characteristics change and energy attenuation. The instrument identifies bolt abnormalities by analyzing the received waveform changes. The ultrasonic wave propagation process is as follows: Figure 2 As shown; simultaneously input the original noisy ultrasonic echo signal to initialize the NGO algorithm, including population size, maximum number of iterations, and search range of optimization parameters;

[0115] S12: When ultrasound propagates in an elastic medium, the kinetic energy generated by the vibration of particles and the deformation potential energy together achieve energy transfer.

[0116] S121: Define the average energy density of ultrasound within a vibration cycle as the average energy density. Perform noise reduction, filtering and normalization on the acquired signal, and use window function and bandpass filtering to suppress environmental noise.

[0117] S122: Define a basic signal x and add zero-mean, unit-variance white noise to it to generate a series of constructed signals x. (i) For each x (i)Calculate the local mean, and take the average of all local means to obtain the first residual component r. i And calculate the first modal component imf1; iteratively calculate the subsequent modal components imf k The iteration continues until the residual components satisfy the termination condition or the modal components are less than the first three local extrema.

[0118] In step S122, the ICEEMDAN-MPE algorithm is used to decompose the original signal into IMF components. The randomness of the multi-scale permutation entropy is detected for each IMF component, and the average MPE value of each component signal is calculated. The signal is then optimized using the NGO optimization algorithm.

[0119] S123: For each individual in the population, its position represents a set of ICEEMDAN parameters. Using the parameters represented by the current eagle's position, perform ICEEMDAN decomposition on the original signal to obtain a series of IMF components and a residual component. Calculate the multi-scale permutation entropy of each IMF component and set an entropy threshold. Generally, IMF components with higher MPE values ​​are considered noise-dominant components. Based on the MPE values, all IMFs are divided into two categories: a set of signal-dominant IMFs and a set of noise-dominant IMFs. The set of signal-dominant IMFs is summed to reconstruct the denoised signal, and the fitness value is calculated to evaluate the quality of the current parameters.

[0120] S13: Based on the NGO's hunting strategy, which includes the first stage being attack and the second stage being pursuit, update the positions of all individual goshawks and retain the currently globally optimal individual, that is, the goshawk with the best parameters; the position of the northern goshawk population is represented by matrix X, and the objective function value is represented by matrix F.

[0121] S14: Repeat steps S12 and S13 until the maximum number of iterations T is reached or other convergence conditions are met. Finally, the NGO algorithm outputs the globally optimal combination of parameters.

[0122] S15: Using the optimal parameters found by NGO, perform a final ICEEMDAN decomposition on the original signal, calculate the multiscale permutation entropy (MPE) of each IMF, distinguish between signal and noise IMFs using the same thresholding strategy as before, and reconstruct the final optimized ultrasound signal using the signal IMFs.

[0123] In step S121, the formula for calculating the average energy density of ultrasound is:

[0124]

[0125] in, ρ is the average energy density of the ultrasound wave over one cycle; ρ is the density of the material; A is the amplitude of the particle vibration. It is the angular frequency;

[0126] According to the energy density formula, under the premise that the material density and frequency are constant, the ultrasonic energy E is characterized by the sum of the squares of the signal amplitude, and the calculation formula is as follows:

[0127]

[0128] Where E is the energy of the sampled signal; N is the length of the sampled signal; v i Let be the amplitude of the sampled signal at the i-th point;

[0129] The ultrasonic energy decreases as bolt loosening worsens. Based on the calculated energy value E, the energy attenuation rate η is defined to quantitatively characterize the degree of loosening, and its expression is as follows:

[0130]

[0131] Where E0 is the energy of the ultrasonic response signal under the standard state of the bolt; E is the energy of the ultrasonic response signal under the loose state of the bolt. The greater the degree of bolt loosening, the greater the decrease in ultrasonic signal energy. When the ultrasonic signal energy decreases, the ultrasonic signal energy attenuation rate will increase. Therefore, the degree of bolt loosening can be judged by the ultrasonic signal energy attenuation rate.

[0132] In step S122, a basic signal x is set and added to it to generate a series of construct signals x. (i) , represented as:

[0133] x (i) =x+β0E1(w (i) ), (i = 1, 2, ..., N)

[0134] Where, x (i) Let be the i-th constructed signal, N be the total number of constructed signals, β0 be the initial noise standard deviation of the signal, and w (i) It is the added zero-mean unit variance white noise, and E1(·) represents the operation of calculating the first inner membrane state function (IMF) of the signal.

[0135] Calculate the local mean and average it for each x. (i) Calculate the local means, and then take the average of all local means to obtain the first residual component r. i :

[0136]

[0137] Here, M(·) represents the local mean of the signal, and I is the number of constructed signals;

[0138] The first modal component, imf1, is obtained by subtracting the first residual component r1 from the original signal x.

[0139] imf1 = x - r1

[0140] Iterative calculation of subsequent modal components (IMF) k For the k-th modal component, k≥2, use the residual r from the previous iteration. k-1 Subtract the residual r this time k :

[0141] imf k =r k-1 -r k

[0142] Furthermore, r k The calculation method is as follows:

[0143]

[0144] Among them, E k (·) represents the imf operation corresponding to the k-th calculation, β k-1 Let be the noise standard deviation of the (k-1)th iteration;

[0145] The iteration terminates when the residual component meets a certain termination condition or when the modal component is less than the first three local extrema.

[0146] In step S13, the initialization of the position of the Northern Goshawk is represented by matrix X, as shown below:

[0147]

[0148] In the formula, X represents the population of the Northern Goshawk; X i Indicates the current position of the i-th Northern Goshawk; X i,j Let represent the position of the i-th Northern Goshawk in the j-th dimension; N represents the population size; m represents the dimension of the problem. The objective function value for the Northern Goshawk can be calculated by solving the objective function during the problem-solving process, and is represented by matrix F.

[0149]

[0150] Among them, F i Let be the objective function value for i northern eagles.

[0151] Furthermore, preferably, a neural network-based learning model is used to automatically classify the bolt tightness and quantify the degree of looseness. The training samples are derived from laboratory calibration data and field-collected data, and the SOA-LightGBM optimization algorithm is used for training.

[0152] S2: Perform feature extraction and LightGBM algorithm state recognition. Extract and fuse multi-dimensional features of the echo signal in the time and frequency domains. After Min-Max normalization, divide the dataset. Optimize the hyperparameters of the LightGBM model and the unknown fault discrimination threshold through the SOA algorithm to achieve bolt fastening state classification and unknown fault identification.

[0153] Step S2 combines the SOA and LightGBM algorithms. LightGBM is an efficient and fast gradient boosting framework based on decision trees. It's used to solve classification (e.g., bolt tightness, looseness, severe looseness) and regression problems. It boasts fast training speed, low memory consumption, high accuracy, and supports parallel learning. SOA is a metaheuristic optimization algorithm inspired by the changing seasons. It simulates the growth, competition, reproduction, and death behaviors of trees in a forest during spring, summer, autumn, and winter. It possesses strong global exploration and local exploitation capabilities, effectively escaping local optima to find the global optimum or near-optimal solution. By simulating natural evolution, SOA can automatically and intelligently search a vast parameter space to find a set of hyperparameters that best enables the LightGBM model to perform in bolt loosening diagnosis tasks. This combination not only improves the accuracy and robustness of the diagnostic model but also automates model construction, reducing reliance on subjective human experience.

[0154] In this embodiment, in step S2, feature extraction and LightGBM algorithm state recognition are performed. Multi-dimensional features of the echo signal in the time and frequency domains are extracted and fused. After Min-Max normalization, the dataset is divided. The LightGBM model hyperparameters and unknown fault discrimination threshold are optimized using the SOA algorithm to achieve bolt tightening state classification and unknown fault identification. Specifically:

[0155] S21: Extract features of the echo signal, including time delay, peak amplitude, envelope energy, and spectral centroid, and calculate the waveform difference of multiple detections to assess the tightening trend; at the same time, use the ultrasonic signal feature dataset extracted after optimization by NGO-ICEEMDAN-MPE. The dataset contains bolt status labels of known categories, including normal, slightly loose, and severely loose. Divide the dataset into training set, validation set, and test set.

[0156] S22: Extract and fuse time-domain, frequency-domain, and wavelet-domain features from the original vibration signal of the rolling bearing, reducing the dimension of the original vibration signal to 31 dimensions, effectively removing redundant information, reducing training time, and improving model convergence; randomly generate an initial forest, where the position of each tree represents a set of random LightGBM hyperparameters;

[0157] S23: After feature fusion and before dataset partitioning, Min-Max normalization is used to preprocess the feature data; the known fault category data after feature extraction and fusion are divided into training set, validation set and test set. All unknown new faults that have not appeared and have not been modeled are assigned to the test set and do not participate in the model training process, but are only used to verify the model recognition effect; in each seasonal iteration of SOA, the update rate is calculated according to the formula, new trees are randomly generated in the forest to increase population diversity, the trees are sorted according to their strength and core trees are selected, the fitness of neighboring trees is adjusted, strong trees are selected for planting and weak trees are eliminated;

[0158] S24: Train the LightGBM model using a training set with known categories, establish a known fault identification model, construct an unknown class discrimination mechanism, assume any class in the training set as the unknown class, and the remaining classes as the known classes, retrain the classifier with training data, and optimize the unknown class discrimination threshold T using the validation set and SOA; evaluate the quality of the tree in each generation of SOA, track the tree with the highest historical fitness, and output the globally optimal hyperparameter combination after the iteration terminates;

[0159] S25: Input the test sample into the trained LightGBM model to obtain the probability value of each sample. If the highest confidence ρ(i) of the test sample is less than the threshold T, the sample is classified as an unknown new fault. Otherwise, it is classified as the known fault category corresponding to i. Integrate model interpretability technology. Retrain the model using the optimal hyperparameters found by SOA, and deploy it to the actual detection system after evaluation on the test set.

[0160] In step S21, the time delay feature τ is calculated using a cross-correlation function to determine the time lag between the currently detected ultrasonic echo signal y(t) and the reference state signal x(t), which is then used as the time delay feature τ. The calculation formula is as follows:

[0161]

[0162] in, The transit time variation of ultrasonic waves propagating inside the bolt is characterized by its correlation with the internal stress state and sound velocity variation of the bolt. y(t-τ) is the ultrasonic echo signal based on time lag at the current time.

[0163] Peak amplitude A peak The maximum absolute value in the discrete sampled signal is obtained as the peak amplitude feature. Let the sampled signal sequence be v. i ,but:

[0164]

[0165] Where N is the number of sampling points, this feature reflects the strongest reflection intensity of the echo signal, and the air gap formed by loose bolts will cause this amplitude to decrease significantly;

[0166] The envelope energy E is based on the principle of ultrasonic energy frequency density. Under the premise that the material density and frequency are constant, the ultrasonic energy is characterized by the sum of the squares of the signal amplitude. The calculation formula is as follows:

[0167]

[0168] Where E is the energy of the sampled signal, N is the length of the sampled signal, and v i Let be the amplitude of the sampled signal at the i-th point;

[0169] The frequency domain amplitude spectrum |S(f)| is obtained by performing a Fourier transform on the time-domain signal, and its spectral centroid f is calculated. c To reflect the shift in frequency distribution:

[0170]

[0171] Among them, f k Let S(f) be the k-th frequency point, and K be the total number of frequency points. k | represents the frequency domain amplitude value corresponding to the k-th frequency point;

[0172] Step S23 also includes: for each seasonal iteration, according to formula P r =P m ax-(y / Y) * (P m ax-P m in) Calculate the update rate P r In the formula, y is the current iteration number, Y is the maximum iteration number, and Pmax and Pmin are the maximum and minimum update rates, respectively; P is randomly generated in the forest. r * Add A new trees to the forest to increase population diversity, where A is a random number; sort the forest according to the strength of the trees and select the top N. c A strong tree is selected as the core tree; each core tree influences the growth position of its neighboring trees, simulating a competition process. The fitness of the neighboring trees is adjusted according to the competition intensity Λ. The top A strong trees are selected, and A = ψ(P). s *N) is used for sowing, that is, replicating these excellent solutions in the hope of producing better offspring; based on the resistance rate P -W Remove the weakest P in the forest w * Use N trees to simulate natural selection and eliminate undesirable solutions;

[0173] In step S24, the specific implementation process and formula derivation are as follows:

[0174] Based on the gradient boosting decision tree principle, a strong classifier is formed by stacking multiple weak classifiers. Let the number of iterations be Y, and the current decision tree of the weak classifiers be f. i(x), whose corresponding weight is h i Then the strong classifier F is constructed. Y (X) is represented as:

[0175]

[0176] Where X is the input feature vector, f0(x) is the initial model, and f i (x) represents the current weak classifier. The model segments the features using a histogram optimization algorithm and grows the decision tree using a depth-limited leaf-wise strategy to prevent overfitting.

[0177] When there are d known fault categories, the LightGBM model outputs a confidence vector P for each of the d categories for each test sample:

[0178] P = [ρ1, ρ2, ..., ρ d ]

[0179] Where, ρ d The model represents the confidence level of classifying a sample into class d, and the highest confidence level ρ of the sample to be tested is calculated. j :

[0180] ρ j =max{ρ1,ρ2,…,ρd}

[0181] The discrimination threshold is set to T, and the discrimination logic is as follows: If ρ j If ρ > T, then the sample is determined to be of a known category and classified into the j-th category; if ρ j If the value is less than or equal to T, then the sample is determined to be an unknown new fault.

[0182] Preferably, in step S24, after feature fusion and before dataset partitioning, an explicit data preprocessing step is added, using Min-Max normalization to ensure that all features participate in model training on a fair scale, thereby ensuring the stable performance of the LightGBM algorithm.

[0183] Preferably, in step S25, in order to enhance the reliability and practicality of the diagnostic system in practical applications, techniques that go beyond simple confidence output and integrate model interpretability can be used.

[0184] Preferably, in step S25, since the value of the threshold T directly affects the model's recognition accuracy for known classes and the detection rate for unknown classes, it is possible to seek to maximize the macro F1 score in order to balance the algorithm's precision and recall.

[0185] Furthermore, step S24 also includes: in each generation of SOA, the quality of each tree in the forest needs to be evaluated, and a LightGBM model is trained on the training set using the current hyperparameters θ_i. The trained model is then used to predict on the validation set. The SOA algorithm updates the position of trees in the forest according to the seasonal operations and always tracks and retains the tree with the highest fitness in history. When the maximum number of iterations Y is reached or the fitness no longer significantly improves, the SOA loop terminates, and SOA outputs the globally optimal combination of hyperparameters. Step S25 also includes: using the optimal hyperparameters found by SOA, the final LightGBM bolt loosening diagnostic model is retrained on the complete training and validation sets. The performance of the final model is evaluated unbiased using an independent test set. This optimized model is then deployed to an actual transmission tower bolt loosening detection system.

[0186] S3: Visualizes and stores results, provides structured storage of diagnostic results, feature vectors, and model parameters, and enables bolt condition imaging, feature waveform display, historical trend analysis, and alarm prompts through a Python GUI interface. It also supports exporting diagnostic reports.

[0187] Results visualization and storage are the core of the bolt loosening diagnostic system's human-computer interaction and data management. Through a graphical interface, it transforms abstract diagnostic data into intuitive charts, images, and reports, enabling maintenance personnel to quickly and accurately grasp the health status of bolt clusters and providing data support for decision-making. Structured storage ensures the historical traceability of diagnostic results, accumulating valuable data for predictive maintenance and algorithm optimization.

[0188] In this embodiment, in step S3, the results are visualized and stored, and the diagnostic results, feature vectors, and model parameters are stored in a structured manner, specifically as follows:

[0189] S31: Structured storage of diagnostic results: The bolt status output by the LightGBM model, including the tightness, looseness level, unknown fault markers, identification results, corresponding signal feature vectors, model confidence probabilities, and timestamp information, is automatically stored in a structured database or a data file of a specified format; at the same time, the model version and parameters used for each diagnosis are recorded to ensure data traceability;

[0190] S32: Bolt Status Visualization Imaging: Based on stored diagnostic results, draw a two-dimensional or three-dimensional layout diagram of the bolt group in the PYTHON GUI interface, and intuitively mark the real-time or historical status of each bolt in the layout diagram with different colors and number labels.

[0191] S33: Feature and Confidence Visualization: Provides a detailed analysis view of a single bolt in the interface. Select a specific bolt to plot its original ultrasonic signal, the waveform of the key IMF components optimized by NGO-ICEEMDAN-MPE, and the time-domain and frequency-domain feature maps during the feature extraction process.

[0192] S34: Historical Trend Analysis and Alarm Log: The system automatically generates and displays the historical trend chart of the status of key bolts for predictive maintenance; when the system detects a loose state or an unknown fault, it automatically triggers an audible and visual alarm and records the event time, location, specific description, and suggested measures in the dedicated alarm log list on the interface;

[0193] S35: Report Generation and Data Export: Provides a one-click function to generate diagnostic reports, allowing users to export the current visualization interface, raw data, feature data, and diagnostic reports in common formats to a specified local path for easy archiving or reporting later.

[0194] In step S3, bolt condition imaging, characteristic waveform display, historical trend analysis, and alarm prompts are implemented through a Python GUI interface, supporting the export of diagnostic reports. Specifically:

[0195] This software development platform uses PYTHON GUI, a graphical user interface design tool. This software can display images of bolt loosening and analyze the bolt loosening state using image recognition algorithms. Figure 3 This is the system's main interface. Figure 4 This is the interface for data detection and import. Figure 5 This is the data display interface.

[0196] Second Embodiment

[0197] like Figure 6 As shown, this embodiment provides an ultrasonic signal feature-optimized LightGBM diagnostic system for bolt loosening, used to perform the ultrasonic signal feature optimization method for bolt loosening LightGBM diagnostic as described in the first embodiment, comprising:

[0198] The ultrasonic signal preprocessing module is used to perform ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms. It applies pulse excitation through a piezoelectric transducer and receives echo signals. It optimizes the ICEEMDAN parameters using the NGO algorithm and filters signal components using MPE to achieve ultrasonic signal denoising, energy quantization and optimization preprocessing.

[0199] The feature state recognition module is used for feature extraction and LightGBM algorithm state recognition. It extracts and fuses multi-dimensional features of the echo signal in the time and frequency domains, divides the dataset after Min-Max normalization, and optimizes the hyperparameters of the LightGBM model and the unknown fault discrimination threshold through the SOA algorithm to realize bolt fastening state classification and unknown fault recognition.

[0200] The results visualization and storage module is used for results visualization and storage. It stores diagnostic results, feature vectors and model parameters in a structured manner. It realizes bolt status imaging, feature waveform display, historical trend analysis and alarm prompts through a Python GUI interface, and supports the export of diagnostic reports.

[0201] A computer-readable storage medium stores computer code that, when executed, performs the methods described above. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0202] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0203] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0204] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics, characterized in that, Includes the following steps: S1: Perform ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms. Apply pulse excitation through a piezoelectric transducer and receive echo signals. Optimize ICEEMDAN parameters using the NGO algorithm and filter signal components using MPE to achieve ultrasonic signal denoising, energy quantization and optimization preprocessing. S2: Perform feature extraction and LightGBM algorithm state recognition. Extract and fuse multi-dimensional features of the echo signal in the time and frequency domains. After Min-Max normalization, divide the dataset. Optimize the hyperparameters of the LightGBM model and the unknown fault discrimination threshold through the SOA algorithm to achieve bolt fastening state classification and unknown fault identification. S3: Visualizes and stores results, providing structured storage of diagnostic results, feature vectors, and model parameters. It enables bolt condition imaging, feature waveform display, historical trend analysis, and alarm prompts through a Python GUI interface, and supports exporting diagnostic reports.

2. The LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics according to claim 1, characterized in that, In step S1, ultrasonic signal feature optimization preprocessing and signal acquisition are performed based on a combination of NGO-ICEEMDAN-MPE algorithms. Pulse excitation is applied through a piezoelectric transducer and echo signals are received. The NGO algorithm is used to optimize the ICEEMDAN parameters, and the MPE algorithm is used to filter signal components, achieving ultrasonic signal denoising, energy quantization, and optimized preprocessing. Specifically: S11: A pulse excitation signal is applied axially to the bolt using a piezoelectric transducer, while echo signals reflected from the thread interface, nut end face, and tower components are received. The ultrasonic wave propagates in a straight line along the bolt. The loss is extremely low when the bolt is fastened. When there are defects or loosening, the waveform characteristics change and the energy is attenuated. The instrument identifies bolt abnormalities by analyzing the received waveform changes. At the same time, the original noisy ultrasonic echo signal is input to initialize the NGO algorithm, including the population size, maximum number of iterations, and search range of optimization parameters. S12: When ultrasound propagates in an elastic medium, the kinetic energy generated by the vibration of particles and the deformation potential energy together achieve energy transfer. S121: Define the average energy density of ultrasound within a vibration cycle as the average energy density. Perform noise reduction, filtering and normalization on the acquired signal, and use window function and bandpass filtering to suppress environmental noise. S122: Define a basic signal x and add zero-mean, unit-variance white noise to it to generate a series of constructed signals x. (i) For each x (i) Calculate the local mean, and take the average of all local means to obtain the first residual component r. i And calculate the first modal component imf1; iteratively calculate the subsequent modal components imf k The iteration continues until the residual components satisfy the termination condition or the modal components are less than the first three local extrema. S123: For each individual in the population, its position represents a set of ICEEMDAN parameters. Using the parameters represented by the current eagle's position, ICEEMDAN decomposition is performed on the original signal to obtain a series of IMF components and a residual component. The multi-scale permutation entropy of each IMF component is calculated, and an entropy threshold is set. Typically, IMF components with higher MPE values ​​are considered noise-dominant components. Based on the MPE values, all IMFs are divided into two categories: a set of signal-dominant IMFs and a set of noise-dominant IMFs. The set of signal-dominant IMFs is summed to reconstruct the denoised signal, and the fitness value is calculated to evaluate the quality of the current parameters. S13: Based on the NGO's hunting strategy, which includes the first stage being attack and the second stage being pursuit, update the positions of all individual goshawks and retain the currently globally optimal individual, that is, the goshawk with the best parameters; the position of the northern goshawk population is represented by matrix X, and the objective function value is represented by matrix F. S14: Repeat steps S12 and S13 until the maximum number of iterations T is reached or other convergence conditions are met. Finally, the NGO algorithm outputs the globally optimal combination of parameters. S15: Using the optimal parameters found by NGO, perform a final ICEEMDAN decomposition on the original signal, calculate the multiscale permutation entropy (MPE) of each IMF, distinguish between signal and noise IMFs using the same thresholding strategy as before, and reconstruct the final optimized ultrasound signal using the signal IMFs.

3. The LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics according to claim 1, characterized in that, Also includes: In step S121, the formula for calculating the average energy density of ultrasound is: in, ρ is the average energy density of the ultrasonic wave within one cycle; ρ is the density of the material; A is the amplitude of the particle vibration. It is the angular frequency; According to the energy density formula, under the premise that the material density and frequency are constant, the ultrasonic energy E is characterized by the sum of the squares of the signal amplitude, and the calculation formula is as follows: Where E is the energy of the sampled signal; N is the length of the sampled signal; v i Let be the amplitude of the sampled signal at the i-th point; The ultrasonic energy decreases as bolt loosening worsens. Based on the calculated energy value E, the energy attenuation rate η is defined to quantitatively characterize the degree of loosening, and its expression is as follows: Where E0 is the energy of the ultrasonic response signal under the standard state of the bolt; E is the energy of the ultrasonic response signal under the loose state of the bolt. The greater the degree of bolt loosening, the greater the decrease in ultrasonic signal energy. When the ultrasonic signal energy decreases, the ultrasonic signal energy attenuation rate will increase. Therefore, the degree of bolt loosening can be judged by the ultrasonic signal energy attenuation rate. In step S122, a basic signal x is set and added to it to generate a series of construct signals x. (i) , represented as: x (i) =x+β0E1(w (i) ),(i=1,2,…,N Where, x (i) Let be the i-th constructed signal, N be the total number of constructed signals, β0 be the initial noise standard deviation of the signal, and w (i) It is the added zero-mean unit variance white noise, and E1(·) represents the operation of calculating the first inner membrane state function (IMF) of the signal. Calculate the local mean and average it for each x. (i) Calculate the local means, and then take the average of all local means to obtain the first residual component r. i : Here, M(·) represents the local mean of the signal, and I is the number of constructed signals; The first modal component, imf1, is obtained by subtracting the first residual component r1 from the original signal x. imf1 = x - r1 Iterative calculation of subsequent modal components (IMF) k For the k-th modal component, k≥2, use the residual r from the previous iteration. k-1 Subtract the residual r this time k : imf k =r k-1 -r k Furthermore, r k The calculation method is as follows: Among them, E k (·) represents the imf operation corresponding to the k-th calculation, β k-1 Let be the noise standard deviation of the (k-1)th iteration; The iteration terminates when the residual component meets a certain termination condition or when the modal component is less than the first three local extrema. In step S13, the initialization of the position of the Northern Goshawk is represented by matrix X, as shown below: In the formula, X represents the population of the Northern Goshawk; X i Indicates the current position of the i-th Northern Goshawk; X i,j Let represent the position of the i-th Northern Goshawk in the j-th dimension; N represents the population size; m represents the dimension of the problem. The objective function value for the Northern Goshawk can be calculated by solving the objective function during the problem-solving process, and is represented by matrix F. Among them, F i Let be the objective function value for i northern eagles.

4. The LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics according to claim 1, characterized in that, In step S2, feature extraction and LightGBM algorithm state recognition are performed. Multi-dimensional features of the echo signal in the time and frequency domains are extracted and fused. After Min-Max normalization, the dataset is divided. The SOA algorithm is used to optimize the hyperparameters of the LightGBM model and the unknown fault discrimination threshold, thereby achieving bolt tightening state classification and unknown fault identification. Specifically: S21: Extract features of the echo signal, including time delay, peak amplitude, envelope energy, and spectral centroid, and calculate the waveform difference of multiple detections to assess the tightening trend; at the same time, use the ultrasonic signal feature dataset extracted after optimization by NGO-ICEEMDAN-MPE. The dataset contains bolt status labels of known categories, including normal, slightly loose, and severely loose. Divide the dataset into training set, validation set, and test set. S22: Extract and fuse time-domain, frequency-domain, and wavelet-domain features from the original vibration signal of the rolling bearing, reducing the dimension of the original vibration signal to 31 dimensions, effectively removing redundant information, reducing training time, and improving model convergence; randomly generate an initial forest, where the position of each tree represents a set of random LightGBM hyperparameters; S23: After feature fusion and before dataset partitioning, Min-Max normalization is used to preprocess the feature data; the known fault category data after feature extraction and fusion are divided into training set, validation set and test set. All unknown new faults that have not appeared and have not been modeled are assigned to the test set and do not participate in the model training process, but are only used to verify the model recognition effect. In each seasonal iteration of SOA, the update rate is calculated according to the formula, new trees are randomly generated in the forest to increase population diversity, trees are sorted according to strength and core trees are selected, the fitness of neighboring trees is adjusted, strong trees are selected for planting and weak trees are eliminated. S24: Train the LightGBM model using a training set with known categories, establish a known fault identification model, construct an unknown class discrimination mechanism, assume any class in the training set as the unknown class, and the remaining classes as the known classes, retrain the classifier with training data, and optimize the unknown class discrimination threshold T using the validation set and SOA; evaluate the quality of the tree in each generation of SOA, track the tree with the highest historical fitness, and output the globally optimal hyperparameter combination after the iteration terminates; S25: Input the test sample into the trained LightGBM model to obtain the probability value of each sample. If the highest confidence ρ(i) of the test sample is less than the threshold T, the sample is classified as an unknown new fault. Otherwise, it is classified as the known fault category corresponding to i. Integrate model interpretability technology. Retrain the model using the optimal hyperparameters found by SOA, and deploy it to the actual detection system after evaluation on the test set.

5. The LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics according to claim 1, characterized in that, Also includes: In step S21, the time delay feature τ is calculated using a cross-correlation function to determine the time lag between the currently detected ultrasonic echo signal y(t) and the reference state signal x(t), which is then used as the time delay feature τ. The calculation formula is as follows: in, The transit time variation of ultrasonic waves propagating inside the bolt is characterized by its correlation with the internal stress state and sound velocity variation of the bolt. y(t-τ) is the ultrasonic echo signal based on time lag at the current time. Peak amplitude A peak The maximum absolute value in the discrete sampled signal is obtained as the peak amplitude feature. Let the sampled signal sequence be v. i ,but: Where N is the number of sampling points, this feature reflects the strongest reflection intensity of the echo signal, and the air gap formed by loose bolts will cause this amplitude to decrease significantly; The envelope energy E is based on the principle of ultrasonic energy frequency density. Under the premise that the material density and frequency are constant, the ultrasonic energy is characterized by the sum of the squares of the signal amplitude. The calculation formula is as follows: Where E is the energy of the sampled signal, N is the length of the sampled signal, and v i Let be the amplitude of the sampled signal at the i-th point; The frequency domain amplitude spectrum |S(f)| is obtained by performing a Fourier transform on the time-domain signal, and its spectral centroid f is calculated. c To reflect the shift in frequency distribution: Among them, f k Let S(f) be the k-th frequency point, and K be the total number of frequency points. k | represents the frequency domain amplitude value corresponding to the k-th frequency point; Step S23 also includes: for each seasonal iteration, according to formula P r =P m ax-(y / Y) * (P m ax-P m in) Calculate the update rate P r In the formula, y is the current iteration number, Y is the maximum iteration number, and Pmax and Pmin are the maximum and minimum update rates, respectively; P is randomly generated in the forest. r * Add A new trees to the forest to increase population diversity, where A is a random number; sort the forest according to the strength of the trees and select the top N. c A strong tree is selected as the core tree; each core tree influences the growth position of its neighboring trees, simulating a competition process. The fitness of the neighboring trees is adjusted according to the competition intensity Λ. The top A strong trees are selected, and A = ψ(P). s *N) is used for sowing, that is, replicating these excellent solutions in the hope of producing better offspring; based on the resistance rate P -W Remove the weakest P in the forest w * Use N trees to simulate natural selection and eliminate undesirable solutions; In step S24, the specific implementation process and formula derivation are as follows: Based on the gradient boosting decision tree principle, a strong classifier is formed by stacking multiple weak classifiers. Let the number of iterations be Y, and the current decision tree of the weak classifiers be f. i (x), whose corresponding weight is h i Then the strong classifier F is constructed. Y (X) is represented as: Where X is the input feature vector, f0(x) is the initial model, and f i (x) represents the current weak classifier. The model segments the features using a histogram optimization algorithm and grows the decision tree using a depth-limited leaf-wise strategy to prevent overfitting. When there are d known fault categories, the LightGBM model outputs a confidence vector P for each of the d categories for each test sample: P=[ρ1,ρ2,…,ρ d ] Where, ρ d The model represents the confidence level of classifying a sample into class d, and the highest confidence level ρ of the sample to be tested is calculated. j : r j =max{ρ1,ρ2,…,ρ d } The discrimination threshold is set to T, and the discrimination logic is as follows: If ρ j If ρ > T, then the sample is determined to be of a known category and classified into the j-th category; if ρ j If the value is less than or equal to T, then the sample is determined to be an unknown new fault.

6. The LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics according to claim 1, characterized in that, In step S3, the results are visualized and stored, and the diagnostic results, feature vectors, and model parameters are stored in a structured manner, specifically as follows: S31: Structured storage of diagnostic results: The bolt status output by the LightGBM model, including the tightness, looseness level, unknown fault markers, identification results, corresponding signal feature vectors, model confidence probabilities, and timestamp information, is automatically stored in a structured database or a data file of a specified format; at the same time, the model version and parameters used for each diagnosis are recorded to ensure data traceability; S32: Bolt Status Visualization Imaging: Based on stored diagnostic results, draw a two-dimensional or three-dimensional layout diagram of the bolt group in the PYTHON GUI interface, and intuitively mark the real-time or historical status of each bolt in the layout diagram with different colors and number labels. S33: Feature and Confidence Visualization: Provides a detailed analysis view of a single bolt in the interface. Select a specific bolt to plot its original ultrasonic signal, the waveform of the key IMF components optimized by NGO-ICEEMDAN-MPE, and the time-domain and frequency-domain feature maps during the feature extraction process. S34: Historical Trend Analysis and Alarm Log: The system automatically generates and displays the historical trend chart of the status of key bolts for predictive maintenance; when the system detects a loose state or an unknown fault, it automatically triggers an audible and visual alarm and records the event time, location, specific description, and suggested measures in the dedicated alarm log list on the interface; S35: Report Generation and Data Export: Provides a one-click function to generate diagnostic reports, allowing users to export the current visualization interface, raw data, feature data, and diagnostic reports in common formats to a specified local path for easy archiving or reporting later.

7. The LightGBM diagnostic method for bolt loosening based on optimized ultrasonic signal characteristics according to claim 1, characterized in that, In step S3, bolt condition imaging, characteristic waveform display, historical trend analysis, and alarm prompts are implemented through a Python GUI interface, supporting the export of diagnostic reports. Specifically: This software development platform uses PYTHON GUI, a graphical user interface design tool. This software can display images of bolt loosening and analyze the bolt loosening state using image recognition algorithms.

8. A bolt loosening LightGBM diagnostic system with ultrasonic signal feature optimization for performing the bolt loosening LightGBM diagnostic method with ultrasonic signal feature optimization as described in any one of claims 1-7, characterized in that, include: The ultrasonic signal preprocessing module is used to perform ultrasonic signal feature optimization preprocessing and signal acquisition based on the combination of NGO-ICEEMDAN-MPE algorithms. It applies pulse excitation through a piezoelectric transducer and receives echo signals. It optimizes the ICEEMDAN parameters using the NGO algorithm and filters signal components using MPE to achieve ultrasonic signal denoising, energy quantization and optimization preprocessing. The feature state recognition module is used for feature extraction and LightGBM algorithm state recognition. It extracts and fuses multi-dimensional features of the echo signal in the time and frequency domains, divides the dataset after Min-Max normalization, and optimizes the hyperparameters of the LightGBM model and the unknown fault discrimination threshold through the SOA algorithm to realize bolt fastening state classification and unknown fault recognition. The results visualization and storage module is used for results visualization and storage. It stores diagnostic results, feature vectors and model parameters in a structured manner. It realizes bolt status imaging, feature waveform display, historical trend analysis and alarm prompts through a Python GUI interface, and supports the export of diagnostic reports.

9. A computer device, characterized in that, The device includes a memory and one or more processors, wherein the memory stores computer code that, when executed by the one or more processors, causes the one or more processors to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer code that, when executed, performs the method as described in any one of claims 1 to 7.

Images