A deep learning-based multi-bolt loosening fault diagnosis method and device
By optimizing hyperparameters using a dual-branch deep learning network and the Black-winged Kite optimization algorithm, the problems of difficulty in early feature extraction and insufficient cross-scenario generalization ability in multi-bolt loosening fault diagnosis are solved, achieving efficient and accurate multi-bolt loosening fault identification and localization.
Patent Information
- Application Number
- CN202611308239.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to effectively extract early minor fault features in multi-bolt structures, lack cross-scenario generalization capabilities, and are difficult to tune hyperparameters, resulting in insufficient accuracy and stability in diagnosing multi-bolt loosening faults.
A dual-branch deep learning network model is adopted, which combines convolutional neural networks and gated recurrent units to extract spatiotemporal features. The hyperparameters are optimized by adaptive optimization using the Black Kite optimization algorithm to achieve accurate diagnosis of multi-bolt loosening faults.
It improves the sensitivity and accuracy of identifying multi-bolt loosening faults, reduces the false alarm rate and false negative rate, has excellent engineering generalization diagnostic capabilities, and meets the adaptive monitoring needs of industrial equipment.
Smart Images

Figure CN123069115A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mechanical fault diagnosis and structural health monitoring technology, specifically relating to a method and device for diagnosing multi-bolt loosening faults based on deep learning. Background Technology
[0002] High-strength bolted connections, as a crucial fastening and assembly method, are widely used in modern industrial large-scale steel structures such as wind turbine towers, large machinery flanges, aerospace fuselages, and bridge steel structures. Because these large pieces of equipment often operate for extended periods in harsh industrial environments characterized by strong alternating loads, intense mechanical vibrations, multi-axis complex impacts, and multi-physics coupling, bolted connections are highly susceptible to preload decay and early loosening. If localized loosening occurs in a multi-bolted connection structure and is not detected in time, it will lead to asymmetric stiffness degradation and stress redistribution, potentially causing localized fractures or even catastrophic structural collapse. This not only poses a serious threat to public and equipment safety but also results in incalculable economic losses. Therefore, real-time and accurate health monitoring and early fault identification of the fastening status of multi-bolted structures are core requirements for ensuring the safe operation of large industrial equipment and reducing life-cycle maintenance costs.
[0003] Currently, mainstream bolt health monitoring methods primarily rely on vibration signals collected by physical quantity sensors for data analysis. When a multi-bolt structure experiences minor loosening, the contact nonlinearity at the joint surface can lead to extremely weak physical distortions or local impact modulation features mixed into the structural dynamic response signal. To uncover these fault characteristics, existing technologies generally employ traditional signal processing methods (such as Fast Fourier Transform (FFT), wavelet transform, or Empirical Mode Decomposition (EMD)) to extract manually generated statistical indicators in the time-frequency domain, and then combine these with Support Vector Machines (SVM), Random Forests, or traditional shallow machine learning classifiers for fault state identification.
[0004] However, when dealing with actual multi-bolt group connection structures, the above-mentioned existing technologies have the following significant defects and limitations in engineering applications:
[0005] 1. Difficulty in extracting and fully representing early minute fault features: Traditional signal processing methods often rely heavily on expert experience to select specific frequency bands or feature indicators. Due to the large number of bolts and their dense spatial arrangement in multi-bolt structures, the contact nonlinear response caused by the early loosening of a single bolt is extremely weak and easily masked by strong industrial background noise. Traditional methods only perform shallow feature mining from a single time dimension or a single frequency domain dimension, which cannot fully characterize the interwoven spatial topological distortion features and long-period time evolution dependencies in multi-bolt fault signals.
[0006] 2. Insufficient cross-scenario generalization ability and poor model robustness: The operating conditions of actual large-scale mechanical equipment are extremely complex and variable (such as the rotational speed fluctuation of wind turbine towers under different wind speeds). Existing traditional machine learning models are limited by their shallow architecture and have a weak ability to learn high-dimensional nonlinear features. When the operating speed, load conditions change, or when faced with small loose boundaries that have not been trained on, the recognition accuracy and generalization ability of the diagnostic model will drop sharply, making it difficult to meet the requirements of adaptive diagnosis under multiple operating conditions in actual industrial sites.
[0007] 3. Hyperparameters are extremely sensitive, and manual parameter tuning is inefficient: Although some research has attempted to introduce deep neural networks to automatically extract features in recent years, the internal hyperparameters of the network model are extremely complex and highly sensitive when building complex deep networks for multi-bolt spatiotemporal feature extraction. Existing technologies mostly use traditional trial-and-error methods or grid search methods for manual parameter tuning, which not only consumes massive amounts of computational resources and time, but also easily causes the model training to get stuck in local optima. As a result, the adaptive accuracy and training stability of the diagnostic model in actual multi-bolt fault location and severity quantification are difficult to guarantee.
[0008] In summary, existing technologies cannot achieve deep spatiotemporal feature adaptive mining, precise localization, and degree quantification of early minor loosening faults in multi-bolt groups purely from the algorithmic level under conditions of strong noise, variable operating conditions, and no manual parameter tuning intervention. Summary of the Invention
[0009] To address the shortcomings of existing multi-bolt loosening fault diagnosis methods, such as insufficient extraction of early minor fault features, poor generalization ability across operating conditions, and difficulty in manually adjusting hyperparameters, a deep learning-based multi-bolt loosening fault diagnosis method and equipment are proposed.
[0010] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0011] A deep learning-based method for diagnosing multi-bolt loosening faults includes:
[0012] Step 1: Obtain the original vibration response signal of the multi-bolt structure to be tested, and preprocess the original vibration response signal to obtain the characteristic sample to be tested;
[0013] Step 2: Input the feature samples to be tested into the trained dual-branch deep learning network model. The dual-branch deep learning network model includes a convolutional neural network branch and a gated recurrent unit branch set in parallel.
[0014] The spatial topological features of the feature samples to be tested are extracted by branches of a convolutional neural network to obtain the first feature vector.
[0015] The dynamic time-series dependent features of the feature samples to be tested are extracted by the gated recurrent unit branch to obtain the second feature vector;
[0016] The first feature vector and the second feature vector are concatenated and fused to obtain the spatiotemporal fusion feature vector, and the spatiotemporal fusion feature vector is input into the classifier to obtain the predicted probability of each fault category.
[0017] Step 3: Output the fault diagnosis type of the multi-bolt structure under test based on the predicted probability;
[0018] In the offline training phase, the Black-winged Kite Optimization Algorithm is used to globally adaptively optimize the hyperparameters of the dual-branch deep learning network model in order to obtain a well-trained dual-branch deep learning network model.
[0019] To optimize the above technical solution, the specific measures also include:
[0020] In step one above, the original vibration response signal is preprocessed to obtain the feature samples to be tested, including: using a sliding window mechanism to perform overlapping slicing on the original vibration response signal, setting the window length to L and the sliding step size to S, and extracting several one-dimensional feature sample vectors of equal length; wherein, the sliding window mechanism is also used for the sample construction of the training set and validation set in the offline training stage, and labeling multiple categories of fault labels according to the actual bolt loosening position and loosening degree corresponding to each sample.
[0021] The offline training process of the aforementioned dual-branch deep learning network model specifically includes:
[0022] T1. Collect the original time-domain vibration response signals of the multi-bolt structure under different fastening states. The fastening states include the full health state and at least one bolt loosening fault state. The original time-domain vibration response signals are preprocessed and labeled with the corresponding fault state labels to construct training and validation sets.
[0023] T2. Construct the initial two-branch deep learning network model, which includes a convolutional neural network branch and a gated recurrent unit branch.
[0024] T3. Input the feature samples from the training set into the convolutional neural network branch and the gated recurrent unit branch in parallel to extract the spatial topological feature vector and the dynamic time series feature vector, respectively.
[0025] Extracting spatial feature vectors from branches of convolutional neural networks The formula for calculating the forward propagation of its one-dimensional convolution is:
[0026]
[0027] In the formula, It is the hyperbolic tangent activation function; The input feature vector; Indicates the convolution operation; The kernel weight matrix; This is the bias vector; simultaneously, deep temporal feature vectors are extracted from the gated recurrent unit branches, and their hidden state dynamic update formula is:
[0028]
[0029]
[0030]
[0031]
[0032] In the formula, and These are the update gate and reset gate state vectors, respectively; Use the Sigmoid activation function; , , These are the corresponding weight matrices; The hidden state vector from the previous time step; The candidate hidden state vector; This is the hidden state vector output at the current time step;
[0033] T4. The extracted spatial feature vector is concatenated and fused with the temporal feature vector of the last time step, and then mapped and input into the classifier through a fully connected layer. A loss function is constructed based on the deviation between the prediction result and the fault state label.
[0034] T5. The Black-winged Kite Optimization Algorithm is used to optimize the hyperparameters of the dual-branch deep learning network model. With the goal of minimizing the loss function, the optimal hyperparameter combination is output, and the trained dual-branch deep learning network model is obtained based on the optimal hyperparameter combination.
[0035] The above-mentioned extraction of spatial topological features of the test feature sample through a convolutional neural network branch to obtain the first feature vector includes: inputting the test feature sample into a convolutional neural network branch containing multiple one-dimensional convolutional layers and pooling layers; extracting contact nonlinear spatial distortion features caused by bolt loosening in the local receptive field through the sliding operation of the convolutional kernel; performing nonlinear mapping using an activation function; performing dimensionality reduction processing through a pooling layer; and obtaining the first feature vector after processing by a flattening layer.
[0036] The above-mentioned extraction of dynamic time-series dependent features of the test feature sample through the gated loop unit branch to obtain the second feature vector includes: inputting the test feature sample into the gated loop unit branch, updating the degree to which the previous state information is retained in the current state by updating the gate, and resetting the degree of fusion between the historical state information and the current input information by resetting the gate to capture the long-range periodic evolution trend of the vibration signal on the continuous time axis to obtain the second feature vector.
[0037] The above-mentioned optimization of hyperparameters for a two-branch deep learning network model using the Black-winged Kite optimization algorithm includes:
[0038] T51. Set the initial parameters of the black-winged kite optimization algorithm and initialize the black-winged kite population in the preset hyperparameter search space. The position vector of each black-winged kite individual corresponds to a set of candidate hyperparameter solutions of a dual-branch deep learning network model. The hyperparameters include at least one of the following: convolution kernel size, number of network channels, learning rate, and batch size.
[0039] T52. Substitute each candidate hyperparameter solution into the dual-branch deep learning network model, train it using the training set, calculate the fitness value of each Black-winged Kite individual based on the model performance metrics on the validation set, and select the position vector of the individual with the best current fitness value. ;
[0040] T53. Local refined search and update based on the mathematical model of attack behavior of black-winged kites, and global exploration and update based on the mathematical model of migration behavior of black-winged kites combined with the Cauchy mutation mechanism, iteratively updating the position vector of each individual black-winged kite.
[0041] T54. Determine whether the maximum number of optimization iterations T has been reached. If not, return to step T52 to continue iterating. If it has been reached, stop iterating and output the globally optimal individual as the target hyperparameter configuration, thus solidifying the trained dual-branch deep learning network model.
[0042] In step T51, the black-winged kite population is initialized within a preset hyperparameter search space, using the following mathematical model:
[0043]
[0044]
[0045] In the formula, For the black-winged kite population matrix; This represents the dimension of the hyperparameter optimization problem; From 1 to Integers between [a certain number] Indicates the size and number of the black-winged kite population. Represents the first in the optimization search space There are 10 candidate solutions; and They represent the first The black-winged kite in the first Lower and upper bounds of a position in 3D space; It is a random number between [0,1].
[0046] In step T53 above, a local refined search update is performed based on the mathematical model of the black-winged kite's attack behavior. The updated mathematical model is as follows:
[0047] If r < 0.9, then execute the formula:
[0048]
[0049] If r ≥ 0.9, then execute the formula:
[0050]
[0051] In the formula, This represents the current iteration number; and They represent the current generation and the next generation, respectively. Position vectors of individual black-winged kites; For the first The position vector of the individual with the best fitness value in the population; A random number between [0, 1]; The dynamic nonlinear exponential decay parameter for controlling the optimization step size is calculated using the following formula:
[0052]
[0053] In the formula, T is the preset maximum number of optimization iterations;
[0054] A global exploration and update is performed based on a mathematical model of the migration behavior of the black-winged kite combined with the Cauchy mutation mechanism. The updated mathematical model is as follows:
[0055] If F i >F ri Then execute the formula:
[0056]
[0057] If F i ≤ F ri Then execute the formula:
[0058]
[0059] In the formula, F i Let F be the fitness value of the i-th individual black-winged kite. riAnother black-winged kite individual selected randomly. The fitness value; C(0,1) is a random mutation operator conforming to the standard Cauchy distribution, used to give the model the ability to escape local optima by introducing abrupt perturbations; m is the direction and magnitude amplification control factor, defined by the following formula:
[0060] .
[0061] The spatiotemporal fusion feature vector is input into the classifier to obtain the predicted probability of each fault category, including: inputting the concatenated and fused spatiotemporal fusion feature vector into the Softmax classification layer, and calculating the classification probability of each fault category using the following formula:
[0062]
[0063] In the formula, This is the output fault category probability vector, where the values of each dimension in the vector represent the probability that the input sample belongs to the corresponding multi-bolt state. This is the weight matrix of the output layer; This is the bias vector for the output layer; The time feature vector output by the gated loop unit branch at the last time step T;
[0064] Output the fault diagnosis type of the multi-bolt structure under test, including: the category corresponding to the maximum probability in the classification probability vector is used as the final fault diagnosis type output. The fault diagnosis type includes the full health status and the loose location label of the specific bolt number.
[0065] A deep learning-based multi-bolt loosening fault diagnosis device includes: a signal acquisition module for acquiring the original vibration response signal of the multi-bolt structure under test; a data preprocessing module for performing sliding window slicing preprocessing on the original vibration response signal to obtain the feature sample to be tested; a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned deep learning-based multi-bolt loosening fault diagnosis method and outputs the fault diagnosis type of the multi-bolt structure under test.
[0066] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned deep learning-based multi-bolt loosening fault diagnosis method.
[0067] Compared with the prior art, the present invention has the following beneficial effects:
[0068] 1. A dual-branch spatiotemporal feature fusion architecture was constructed to achieve deep representation of minor loosening features: This invention abandons single-dimensional feature extraction, utilizes a convolutional neural network (CNN) branch to capture spatial topological nonlinear distortions, and uses a gated recurrent unit (GRU) branch to mine long-sequence temporal evolution dependencies. Through high-dimensional splicing and fusion, the boundaries of loosening fault features are locked, significantly improving the sensitivity and accuracy of early minor loosening localization.
[0069] 2. Introducing the Black-winged Kite Optimization Algorithm to Achieve Adaptive Optimization of the Model and Avoid Getting Stuck in Local Optimum: This invention achieves multi-dimensional linkage optimization of network hyperparameters by interweaving and iterating the local attack behavior of the Black-winged Kite Optimization Algorithm (BKA) with the global migration behavior based on Cauchy mutation. This completely eliminates the blindness of manual experience-based parameter tuning and ensures the rapid convergence and training stability of the model in different multi-bolt structure application scenarios.
[0070] 3. Possesses excellent engineering generalization diagnostic capabilities, meeting the adaptive monitoring needs of industrial equipment: The end-to-end intelligent diagnostic mechanism enables this method to adaptively process nonlinear vibration response data under complex working conditions. By outputting intuitive fault categories and location labels, it significantly reduces the false alarm rate and false negative rate of multi-bolt groups, providing reliable technical support for the life assessment and maintenance of large high-strength bolt structures. Attached Figure Description
[0071] Figure 1 This is a flowchart illustrating the overall implementation of the deep learning-based multi-bolt loosening fault diagnosis method provided in this embodiment of the invention.
[0072] Figure 2 The original time-domain vibration response signal waveforms of the multi-bolt structure under different health conditions provided in the embodiments of the present invention are shown.
[0073] Figure 3 A complete topology diagram of the dual-branch deep learning network model provided in this embodiment of the invention;
[0074] Figure 4 The iterative control flowchart for multidimensional hyperparameter optimization of the Black-winged Kite optimization algorithm provided in this embodiment of the invention is shown.
[0075] Figure 5 This is an intermediate feature response map of adaptive topology evolution during the hyperparameter optimization process provided in this embodiment of the invention;
[0076] Figure 6 The diagram shows the loss function and accuracy convergence curve of the adaptive deep learning model provided in this embodiment of the invention during the training process.
[0077] Figure 7 This is a schematic diagram of the confusion matrix of the multi-bolt loosening fault diagnosis results provided in an embodiment of the present invention. Detailed Implementation
[0078] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0079] The method provided by the present invention will be described in detail below with reference to specific embodiments:
[0080] This embodiment first uses accelerometers deployed on key measuring points of multi-bolted connection structures (such as the flange pair connecting wind turbine towers) to collect the original time-series vibration response signals of the structure under different bolt loosening states.
[0081] To provide comprehensive physical source data for training deep learning models, this embodiment simulates and collects data on various typical fastening states of multi-bolt structures during the experimental phase. For example... Figure 2 As shown, the original time-domain vibration response signal waveforms of the multi-bolt structure collected in this embodiment are presented in the fully healthy state, the early slight loosening state of a single bolt (such as bolt No. 1), and the combined loosening state of multiple bolts.
[0082] In this embodiment, the standard initial rated preload torque of the bolts is set to the fully healthy state; when the tightening torque of one or more bolts is less than 90% of the standard initial rated preload torque, it is defined as an early minor loosening fault state. The sensor sampling frequency is set to 16kHz to obtain the original one-dimensional time-domain data stream containing nonlinear impact modulation distortion of the contact surface under different operating speed conditions (including 1000rpm and 3000rpm).
[0083] To transform one-dimensional continuous-time signals into training feature samples suitable for batch matrix operations in deep learning networks, this embodiment employs a dynamic sliding window mechanism to perform overlapping slicing on the original time-domain vibration response signal vector. The sampling point length of the sliding window is set to L=1024, and the sliding step size is S. By continuously overlapping the window on the original one-dimensional time series, several one-dimensional feature sample vectors of equal length containing overlapping data segments are extracted.
[0084] Based on the actual physical loosening location of each one-dimensional feature sample vector at the time of acquisition (such as bolt No. 1, bolt No. 2) and the severity of loosening (such as healthy, slightly loose, severely loose), the slice sample vectors are labeled with multi-class fault labels. Finally, a standard fault diagnosis training set and validation set are constructed for subsequent offline training and optimization of the model.
[0085] The dual-branch deep learning network model constructed in this embodiment consists of a spatial feature extraction branch of a convolutional neural network (CNN) and a temporal feature extraction branch of a gated recurrent unit (GRU). The feature sample vectors constructed in the above steps are synchronously and in parallel input into the two feature mining channels of the dual-branch deep learning network model. Its specific hierarchical topology and core parameter configuration are shown in Table 1.
[0086] Table 1
[0087]
[0088] In Table 1, NK represents the number of neurons or convolutional kernels; KS represents the convolutional kernel size and stride; OS represents the output dimension; and AC represents the activation function and configuration.
[0089] Combined with Table 1 Figure 3 The network architecture shown, with its forward propagation logic for parallel extraction and fusion of spatiotemporal features in dual-branch deep learning, unfolds as follows:
[0090] (1) The spatial local feature extraction feature sample vector of the CNN branch enters the L2 layer of the CNN branch. The one-dimensional convolution kernel performs a local sliding weighted summation operation on the time axis, automatically mining the spatial topology and stiffness distortion features caused by contact nonlinearity due to bolt loosening in the vibration signal. The specific forward propagation convolution calculation formula is as follows:
[0091]
[0092] In the formula, It is the hyperbolic tangent activation function; The input feature vector; Indicates the convolution operation; The kernel weight matrix; This is the bias vector. The data is then subjected to dimensionality reduction and downsampling through a max pooling layer. At the L6 layer at the end of the convolutional branch, a global average pooling layer is used for feature flattening, outputting a spatial local feature vector with dimension [Batch_Size, 64]. .
[0093] (2) The deep temporal sequence feature extraction of the GRU branch is performed in parallel with the CNN branch, and the feature sample vector is synchronously input into the multi-layer stacked gated recurrent unit network. Utilizing its unique gated adaptive update mechanism, the deep dynamic temporal evolution and long-short-term time dependence trends of multiple bolt loosening are captured on a continuous time axis. The calculation formulas for the control gating state and hidden state update within each gated recurrent unit cell are as follows:
[0094]
[0095]
[0096]
[0097]
[0098] In the formula, and These are the update gate and reset gate state vectors, respectively; Use the Sigmoid activation function; , , These are the corresponding weight matrices; The hidden state vector from the previous time step; The candidate hidden state vector; This is the hidden state vector output at the current time step.
[0099] After undergoing continuous temporal propagation, the GRU feature extraction branch outputs a sequence of dimensions [Batch_Size, 128] at the last time step T. , This is the time feature vector output by the gated loop unit branch at the last time step T.
[0100] (3) Feature matrix concatenation and fusion and Softmax probability mapping are shown in Table 1 as F1 levels. In order to seamlessly merge the extracted spatial topological features and time series dependencies across dimensions, this embodiment uses the spatial feature vector output by the CNN branch. The time feature vector output by the GRU branch Perform horizontal matrix concatenation (Concat) alignment and directly merge them into a comprehensive spatiotemporal feature vector with dimension [Batch_Size, 192].
[0101] After linear flattening and dimensionality reduction of the feature vector through the F2 fully connected layer, it is fed into the F3 Softmax classification output layer to calculate the classification probability distribution of each preset fault category. The specific matrix mapping formula is as follows:
[0102]
[0103] In the formula, This is the output fault category probability vector, where the values of each dimension in the vector represent the probability that the input sample belongs to the corresponding multi-bolt state. This is the weight matrix of the output layer; This is the bias vector for the output layer.
[0104] To eliminate the randomness and limitations of blindly adjusting hyperparameters of deep learning models based on human experience, this embodiment introduces the Black-winged Kite Optimization Algorithm (BKA) during the offline training phase to perform linked global adaptive optimization of the core multidimensional hyperparameters of the dual-branch deep learning network model. The specific hyperparameter optimization and control process is as follows: Figure 4 As shown, this can be broken down into the following five standardized sub-steps:
[0105] T51. Definition of Hyperparameter Search Space and Population Initialization
[0106] The population size of the Black-winged Kite optimization algorithm is set to pop, and the spatial dimension of the hyperparameters of the deep learning model to be optimized is d. Within the closed intervals of the upper and lower bounds of the preset one-dimensional convolutional kernel size, number of network channels, initial learning rate, and batch size (Batch_Size), the th hyperparameter in the optimization search space is randomly initialized and generated using the following formula. One candidate solution:
[0107]
[0108]
[0109] In the formula, For the black-winged kite population matrix; This represents the dimension of the hyperparameter optimization problem; From 1 to Integers between [a certain number] Indicates the size and number of the black-winged kite population. Represents the first in the optimization search space There are 10 candidate solutions; and They represent the first The black-winged kite in the first Lower and upper bounds of a position in 3D space; It is a random number between [0,1].
[0110] T52, Adaptive Model Training and Current Generation Best Fitness Evaluation
[0111] The candidate hyperparameter configurations for each individual black-winged kite are substituted into a dual-branch deep learning network model. The model is then trained and updated using the constructed training set through forward and backward propagation. The cross-entropy loss function value or classification error rate on the validation set is used as the fitness value F for the i-th individual black-winged kite. i By comparing the diagnostic accuracy of each solution, the individual position vector with the best fitness value in the current generation of the population is selected. .
[0112] T53, Localized Refined Search Iteration to Simulate Attack Behavior
[0113] During the local development phase of the algorithm execution, the aerial attack behavior of a black-winged kite within its foraging area is simulated. The algorithm automatically generates a random number r between 0 and 1. Based on the constraint relationship between the random number r and a preset attack mode switching threshold (set to 0.9 in this embodiment), a segmented local position update is performed. The mathematical model is as follows:
[0114] If r < 0.9, then execute the formula:
[0115]
[0116] If r ≥ 0.9, then execute the formula:
[0117]
[0118] In the formula, This represents the current iteration number; and They represent the current generation and the next generation, respectively. Position vectors of individual black-winged kites; For the first The position vector of the individual with the best fitness value in the population; A random number between [0, 1]; The dynamic nonlinear exponential decay parameter for controlling the optimization step size is calculated using the following formula:
[0119]
[0120] In the formula, T is the preset maximum number of optimization iterations. For example... Figure 5 As shown, the intermediate feature responses of adaptive topological evolution during hyperparameter optimization are presented. The evolutionary step size of individuals increases with the number of iterations. Gradually converge to the optimal point to achieve a refined search within the neighborhood.
[0121] T54, Global Large-Scale Exploration and Iteration of Simulated Migration Behavior
[0122] To prevent the hyperparameter optimization process from getting stuck in a local infinite loop or reaching an extreme point, the algorithm simulates the cross-regional migration behavior of a black-winged kite population during the global exploration phase. The algorithm randomly selects another black-winged kite individual from the current t-th generation population. The fitness values of the current individual and random individuals are compared, and a segmented positional mutation update is performed based on the fitness deviation. The mathematical model is as follows:
[0123] If F i >F ri Then execute the formula:
[0124]
[0125] If F i ≤ F ri Then execute the formula:
[0126]
[0127] In the formula, F i Let F be the fitness value of the i-th individual black-winged kite. ri Another black-winged kite individual selected randomly. The fitness value; C(0,1) is a random mutation operator conforming to the standard Cauchy distribution, used to give the model the ability to escape local optima by introducing abrupt perturbations; m is the direction and magnitude amplification control factor, defined by the following formula:
[0128]
[0129] T55. Iteration Termination Judgment and Target Optimal Model Fixation: Determine if the current optimization iteration count t has reached the preset maximum optimization iteration count T. If not, set t = t + 1 and automatically return to step T52 to continue the iteration loop; if it has reached the maximum, terminate the optimization loop and output the globally optimal individual. The optimal hyperparameter configuration scheme of the dual-branch network model is used as the target, and the network structure under this optimal configuration is solidified as the optimal multi-bolt loosening fault diagnosis model.
[0130] After offline training and hyperparameter optimization are completed, the method of this invention can be directly applied to online real-time health monitoring of the multi-bolted structure under test. The original vibration response signal of the multi-bolted structure under test is acquired online in real time through sensors, and the resulting feature sample vector is obtained through the same sliding window mechanism. This vector is then directly input into the aforementioned pre-defined optimal deep learning model. The Softmax classification layer directly outputs the category index corresponding to the maximum probability, which serves as the final online fault diagnosis type for the multi-bolted structure.
[0131] like Figure 6 As shown, the loss function and accuracy convergence curve of the optimal adaptive deep learning model during formal training are presented. With the steady progression of training epochs, the loss values of both the training and validation sets exhibit a rapid and stable adaptive exponential decrease without any overfitting anomalies. Simultaneously, the diagnostic accuracy curve rapidly rises and converges to an extremely high level, demonstrating that the dual-branch network model after BKA hyperparameter optimization possesses excellent convergence speed and training robustness.
[0132] Finally, the trained model is used for inference and diagnosis on independent test sample sets, and the output localization results are statistically compared with the actual real labels to obtain the following graphs: Figure 7 The confusion matrix shown is from... Figure 7 As clearly demonstrated, the method provided by this invention achieves exceptionally high classification accuracy on the main diagonal of the confusion matrix for all healthy states in multi-bolt group connection structures, early minor loosening locations of individual bolts (such as bolts 1 and 2 occurring alone), and complex loosening faults occurring simultaneously in multiple bolts. No significant false positives, false negatives, or label overlap were observed. This strongly validates the significant industrial practical value and intellectual property protection barrier of the pure deep learning adaptive diagnostic method of this invention in the identification of defects in multi-bolt structures.
[0133] This invention also proposes a deep learning-based multi-bolt loosening fault diagnosis device, comprising: a signal acquisition module for acquiring the original vibration response signal of the multi-bolt structure under test; a data preprocessing module for performing sliding window slicing preprocessing on the original vibration response signal to obtain the feature sample to be tested; a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned deep learning-based multi-bolt loosening fault diagnosis method and outputs the fault diagnosis type of the multi-bolt structure under test.
[0134] The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described deep learning-based multi-bolt loosening fault diagnosis method.
[0135] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.
[0136] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A multi-bolt loosening fault diagnosis method based on deep learning, characterized in that, include: Step 1: Obtain the original vibration response signal of the multi-bolt structure to be tested, and preprocess the original vibration response signal to obtain the feature sample to be tested; Step 2: Input the feature sample to be tested into the trained dual-branch deep learning network model, which includes a convolutional neural network branch and a gated recurrent unit branch set in parallel. The spatial topological features of the feature sample to be tested are extracted through the branches of the convolutional neural network to obtain the first feature vector; The dynamic time-series dependent features of the feature sample to be tested are extracted through the branch of the gated loop unit to obtain the second feature vector; The first feature vector and the second feature vector are concatenated and fused to obtain a spatiotemporal fusion feature vector, and the spatiotemporal fusion feature vector is input into a classifier to obtain the predicted probability of each fault category. Step 3: Output the fault diagnosis type of the multi-bolt structure under test based on the predicted probability; In the offline training phase, the dual-branch deep learning network model employs the Black-winged Kite Optimization Algorithm to globally adaptively optimize the hyperparameters of the dual-branch deep learning network model in order to obtain the trained dual-branch deep learning network model.
2. The method for diagnosing multi-bolt loosening faults based on deep learning according to claim 1, characterized in that, In step one, the preprocessing of the original vibration response signal to obtain the feature samples to be tested includes: using a sliding window mechanism to perform overlapping slicing on the original vibration response signal, setting the window length to L and the sliding step size to S, and extracting several one-dimensional feature sample vectors of equal length; wherein, the sliding window mechanism is also used for the sample construction of the training set and the validation set in the offline training stage, and labeling multiple categories of fault labels according to the actual bolt loosening position and degree of loosening corresponding to each sample.
3. The method for diagnosing multi-bolt loosening faults based on deep learning according to claim 1, characterized in that, The offline training process of the dual-branch deep learning network model specifically includes: T1. Collect the original time-domain vibration response signals of the multi-bolt structure under different fastening states, including the full healthy state and at least one bolt loosening fault state; preprocess the original time-domain vibration response signals and label them with the corresponding fault state labels to construct training and validation sets. T2. Construct an initial dual-branch deep learning network model, wherein the dual-branch deep learning network model includes the convolutional neural network branch and the gated recurrent unit branch; T3. Input the feature samples from the training set into the convolutional neural network branch and the gated recurrent unit branch in parallel to extract the spatial topological feature vector and the dynamic time series feature vector, respectively. Extracting spatial feature vectors from branches of convolutional neural networks The formula for calculating the forward propagation of its one-dimensional convolution is: In the formula, It is the hyperbolic tangent activation function; The input feature vector; Indicates the convolution operation; The kernel weight matrix; This is the bias vector; simultaneously, deep temporal feature vectors are extracted from the gated recurrent unit branches, and their hidden state dynamic update formula is: In the formula, and These are the update gate and reset gate state vectors, respectively; Use the Sigmoid activation function; , , These are the corresponding weight matrices; The hidden state vector from the previous time step; The candidate hidden state vector; This is the hidden state vector output at the current time step; T4. The extracted spatial feature vector is concatenated and fused with the temporal feature vector of the last time step, and then mapped and input into the classifier through a fully connected layer. A loss function is constructed based on the deviation between the prediction result and the fault state label. T5. The Black-winged Kite Optimization Algorithm is used to optimize the hyperparameters of the dual-branch deep learning network model, with the goal of minimizing the loss function, and the optimal hyperparameter combination is output. The trained dual-branch deep learning network model is obtained based on the optimal hyperparameter combination.
4. A method for diagnosing multi-bolt loosening faults based on deep learning according to claim 2 or 3, characterized in that, The step of extracting the spatial topological features of the test feature sample through the convolutional neural network branch to obtain the first feature vector includes: inputting the test feature sample into the convolutional neural network branch containing multiple one-dimensional convolutional layers and pooling layers; extracting the contact nonlinear spatial distortion features caused by bolt loosening in the local receptive field through the sliding operation of the convolutional kernel; performing nonlinear mapping using an activation function; performing dimensionality reduction processing through a pooling layer; and obtaining the first feature vector after processing by a flattening layer. The step of extracting the dynamic time-series dependent features of the test feature sample through the gated loop unit branch to obtain the second feature vector includes: inputting the test feature sample into the gated loop unit branch, updating the degree to which the previous state information is retained in the current state by updating the gate control, and resetting the degree of fusion between the historical state information and the current input information by resetting the gate control, so as to capture the long-range periodic evolution trend of the vibration signal on the continuous time axis and obtain the second feature vector.
5. The method for diagnosing multi-bolt loosening faults based on deep learning according to claim 3, characterized in that, The optimization of hyperparameters of the dual-branch deep learning network model using the Black-winged Kite Optimization Algorithm includes: T51. Set the initial parameters of the black-winged kite optimization algorithm and initialize the black-winged kite population in the preset hyperparameter search space, wherein the position vector of each black-winged kite individual corresponds to a set of candidate hyperparameter solutions of the dual-branch deep learning network model, and the hyperparameters include at least one of the following: convolution kernel size, number of network channels, learning rate and batch size. T52. Substitute each of the candidate hyperparameter solutions into the dual-branch deep learning network model, train it using the training set, calculate the fitness value of each Black-winged Kite individual based on the model performance metrics on the validation set, and select the individual position vector with the best current fitness value. ; T53. Local refined search and update based on the mathematical model of attack behavior of black-winged kites, and global exploration and update based on the mathematical model of migration behavior of black-winged kites combined with the Cauchy mutation mechanism, iteratively updating the position vector of each individual black-winged kite. T54. Determine whether the maximum number of optimization iterations T has been reached. If not, return to step T52 to continue iterating. If it has been reached, stop iterating and output the globally optimal individual as the target hyperparameter configuration, thus solidifying the trained dual-branch deep learning network model.
6. The method for diagnosing multi-bolt loosening faults based on deep learning according to claim 5, characterized in that, In step T51, the initialization of the black-winged kite population within a preset hyperparameter search space is achieved using the following mathematical model: In the formula, For the black-winged kite colony matrix; This represents the dimension of the hyperparameter optimization problem; From 1 to Integers between [a certain number] Indicates the size and number of the black-winged kite population. Represents the first in the optimization search space There are 10 candidate solutions; and They represent the first The black-winged kite in the first Lower and upper bounds of a position in 3D space; It is a random number between [0,1].
7. The method for diagnosing multi-bolt loosening faults based on deep learning according to claim 5, characterized in that, In step T53, the mathematical model of attack behavior based on the black-winged kite is updated through a local fine-grained search. The updated mathematical model is as follows: If r < 0.9, then execute the formula: If r ≥ 0.9, then execute the formula: In the formula, This represents the current iteration number; and They represent the current generation and the next generation, respectively. Position vectors of individual black-winged kites; For the first The position vector of the individual with the best fitness value in the population; A random number between [0, 1]; The dynamic nonlinear exponential decay parameter for controlling the optimization step size is calculated using the following formula: In the formula, T is the preset maximum number of optimization iterations; The mathematical model of migration behavior based on the black-winged kite is combined with the Cauchy mutation mechanism for global exploration and updating. The updated mathematical model is as follows: If F i >F ri Then execute the formula: If F i ≤ F ri Then execute the formula: In the formula, F i Let F be the fitness value of the i-th individual black-winged kite. ri Another black-winged kite individual selected randomly. The fitness value; C(0,1) is a random mutation operator conforming to the standard Cauchy distribution, used to give the model the ability to escape local optima by introducing abrupt perturbations; m is the direction and magnitude amplification control factor, defined by the following formula: 。 8. The method for diagnosing multi-bolt loosening faults based on deep learning according to claim 1, characterized in that, The step of inputting the spatiotemporal fusion feature vector into the classifier to obtain the predicted probability of each fault category includes: inputting the concatenated and fused spatiotemporal fusion feature vector into the Softmax classification layer, and calculating the classification probability of each fault category using the following formula: In the formula, This is the output fault category probability vector, where the values of each dimension in the vector represent the probability that the input sample belongs to the corresponding multi-bolt state. This is the weight matrix of the output layer; This is the bias vector for the output layer; This refers to the time feature vector output by the gated loop unit branch at the last time step T. The output of the fault diagnosis type of the multi-bolt structure under test includes: taking the category corresponding to the maximum probability in the classification probability vector as the final fault diagnosis type output, wherein the fault diagnosis type includes the full health status and the loosening location label of the specific bolt number.
9. A multi-bolt loosening fault diagnosis device based on deep learning, characterized in that, include: The signal acquisition module is used to acquire the original vibration response signal of the multi-bolt structure under test; The data preprocessing module is used to perform sliding window slicing preprocessing on the original vibration response signal to obtain the feature sample to be tested; The system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a deep learning-based multi-bolt loosening fault diagnosis method as described in any one of claims 1 to 8, and outputs the fault diagnosis type of the multi-bolt structure under test.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a deep learning-based multi-bolt loosening fault diagnosis method as described in any one of claims 1 to 8.
