Bolt fastening state intelligent monitoring method based on machine learning

Abstract

The invention discloses a bolt fastening state intelligent monitoring method based on machine learning, and the method comprises the steps: collecting multi-source fastening state data, and obtaining standardized fastening state data; fragmented segmentation is carried out, and fragment-level fastening state feature vectors are formed; forming a fastening state feature vector sequence, and constructing a fastening state structure chart; a Weisfeier-Lehman algorithm is executed, and a fastening state structure feature vector is formed; outputting a fastening state structure deviation index based on the trained improved soft one-class extreme learning machine model; and based on a preset grading judgment rule, generating a bolt fastening state judgment result. According to the method, the improved soft one-class extreme learning machine model and the Weisfeier-Lehman algorithm are introduced, so that low false alarm, reliable online intelligent monitoring and loosening risk early identification of the bolt fastening state under the condition of abnormal sample scarcity are realized.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based intelligent monitoring method for bolt fastening status. Background Technology

[0002] As large equipment, complex machinery, and engineering structures develop towards higher reliability and longer service life, bolted connections, as a critical foundational connection type, significantly impact overall structural safety due to the stability of their fastening state. In actual operation, bolted connections are subjected to the combined effects of vibration loads, impact loads, temperature changes, and environmental factors, making them prone to preload decay and loosening. Traditional bolt fastening status monitoring technologies often rely on vibration signals, acoustic emission signals, or strain data for analysis, combined with empirical thresholds or simple rules for status judgment. These methods, typically based on single or limited signal characteristics, struggle to effectively address normal fluctuations caused by changes in operating conditions and are sensitive to noise, easily leading to false alarms or missed alarms in complex engineering environments.

[0003] Some existing technologies have begun to incorporate machine learning methods, using feature extraction from multi-source monitoring data to train models for state recognition. Other studies have attempted to utilize graph structures to describe the spatial or structural relationships between bolts, enhancing the representation of the overall connection state. However, existing methods largely remain at the feature-level or simple relation-level modeling, failing to adequately characterize the evolution of fastening states at different structural levels. Furthermore, they struggle to reliably distinguish between normal fluctuations and genuine loosening anomalies under complex conditions, and the depth of structural information utilization and the stability of their discrimination remain significantly limited.

[0004] Therefore, how to provide a machine learning-based intelligent monitoring method for bolt fastening status is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent monitoring method for bolt tightness based on machine learning. This invention fully utilizes multi-source monitoring data acquisition, structured feature modeling, and a class of machine learning methods, detailing a monitoring process for online intelligent determination of bolt tightness under conditions of scarce abnormal samples. The method constructs a structural diagram representing the evolution of bolt tightness by uniformly preprocessing and extracting fragmented features from multi-source data. It then uses the Weisfeiler-Lehman algorithm to perform multi-level representation of the structural features of the tightness state, further combining an improved soft first-class extreme learning machine model to form a soft boundary description for the normal tightness state, achieving a quantitative assessment of the degree of structural deviation from the tightness state. This invention reliably distinguishes between normal fluctuations and true loosening anomalies in bolt tightness, possessing advantages such as low dependence on abnormal samples, strong structural feature expression capability, high adaptability to complex working conditions, and low false alarm rate. It is suitable for online intelligent monitoring of bolted connection structures in complex engineering environments.

[0006] A machine learning-based intelligent monitoring method for bolt fastening status according to an embodiment of the present invention includes:

[0007] Multi-source fastening status data of bolted connection structures during operation are collected, and the multi-source fastening status data is preprocessed to obtain standardized fastening status data.

[0008] The standardized fastening state data is segmented into fragments to obtain a set of fastening state fragments. For each fastening state fragment, time-domain features, frequency-domain features, and statistical features are extracted and concatenated to form a fragment-level fastening state feature vector.

[0009] Based on the segment-level fastening state feature vectors, a fastening state feature vector sequence is formed, and a fastening state structure diagram is constructed.

[0010] The Weisfeiler-Lehman algorithm is executed on the tight state structure graph. The node labels of each graph node are initialized by numerical interval mapping based on the fragment-level tight state feature vector. The node labels are combined with the labels of adjacent nodes and hash updates are performed. After completing a preset number of iterations, the tight state structure feature vector is formed.

[0011] A normal sample set is constructed based on the historical fastening state structure feature vector. An improved soft first-class extreme learning machine model is constructed and trained. The real-time fastening state structure feature vector is input into the trained improved soft first-class extreme learning machine model, and the corresponding fastening state structure deviation index is output.

[0012] Based on the structural deviation index of the fastening state and the preset classification judgment rules, the bolt fastening state judgment result is generated and written into the state history database.

[0013] Optionally, the multi-source fastening status data specifically includes vibration signal data, acoustic emission signal data, strain-related data, structural status auxiliary data, and operating condition and environmental data.

[0014] Optionally, the preprocessing of multi-source fastening state data specifically includes data synchronization and alignment, noise reduction and filtering, trend and drift correction, and amplitude unification and standardization.

[0015] Optionally, forming the fragment-level fastened state feature vector includes:

[0016] The standardized values ​​of various data types in the standardized fastening state data at each sampling time are arranged in chronological order to form a continuous data sequence;

[0017] The continuous data sequence is segmented according to the preset segment length and preset segment step size to generate a set of tight state segments consisting of multiple independent and temporally continuous data subsequences.

[0018] Based on the set of fastened state segments, feature extraction processing is performed on the continuous data sequences within the time interval corresponding to each fastened state segment to obtain the corresponding time-domain feature set, frequency-domain feature set and statistical feature set.

[0019] The time-domain feature set, frequency-domain feature set, and statistical feature set corresponding to each fastened state segment are combined in sequence to generate the corresponding segment-level fastened state feature vector.

[0020] Optionally, the construction of the fastened state structure diagram includes:

[0021] Based on the segment-level fastening state feature vectors, and arranged according to the sampling time order corresponding to each segment-level fastening state feature vector, a fastening state feature vector sequence is formed.

[0022] The node set of the fastening state structure diagram is determined based on the fastening state feature vector sequence. Each segment-level fastening state feature vector is mapped to a node to obtain the node set, and the segment-level fastening state feature vector is used as the node attribute of the corresponding node.

[0023] The edge set of the fastening state structure diagram is determined based on the node set. Temporal adjacency connection is established between adjacent nodes corresponding to the same bolt. Spatial adjacency connection is established between nodes corresponding to different bolts when the bolts are spatially adjacent to each other. Structural connection relationship is established when the bolts belong to the same connection structure and have a direct structural connection relationship.

[0024] By integrating the node set, edge set, and node attributes, a tight state structure diagram is generated.

[0025] Optionally, the step of forming a tight-state structural feature vector after completing a preset number of iterations includes:

[0026] Based on the tight state structure diagram, the Weisfeiler-Lehman algorithm is executed to perform multi-dimensional interval encoding on the fragment-level tight state feature vectors corresponding to each node in the node set. The feature values ​​of each dimension are mapped to the corresponding interval encoding indexes according to the preset interval division rules, and ordered multi-dimensional encoding labels are formed according to the dimensional order as the initial structure labels of each node.

[0027] Under the current structural refinement iteration, the set of adjacent nodes that are adjacent to each node is obtained based on the edge set. The structural labels of each adjacent node in the adjacent node set are arranged to form an ordered adjacent label sequence. The ordered adjacent label sequence is truncated and padded to obtain a fixed-length adjacent label sequence table.

[0028] Under the current structural refinement iteration, the structural label of each node is combined with the corresponding fixed-length adjacency label sequence list, and the combination result is hashed to update the structural label of each node.

[0029] Repeat the structural refinement iteration until the preset number of iterations is reached. Stack the structural labels obtained in each structural refinement iteration in order of iteration to construct a node label iteration stacking matrix. Select the initial structural label and the structural label after completing the preset number of iterations for each node according to the preset skip layer rounds to generate the skip layer label path for each node and form a skip layer label aggregation structure.

[0030] Based on the node label iterative stacking matrix, the jump-level label aggregation structure, and the tight state structure graph, a three-dimensional tensor graph is constructed. Based on the three-dimensional tensor graph, the structural labels corresponding to different nodes, adjacency relationships, and structural refinement iterations are subjected to multi-dimensional aggregation processing. The aggregation result is then mapped to a one-dimensional vector representation to generate a tight state structure feature vector.

[0031] Optionally, the construction and training of the improved soft one-class extreme learning machine model includes:

[0032] Select the structural feature vectors of historical fastening states that are in a confirmed normal fastening state, aggregate them according to time index to form a normal sample set, and divide the normal sample set into a normal sample training subset and a normal sample verification subset;

[0033] An improved soft first-class extreme learning machine model is constructed, which consists of a feature-carrying module, a mapping generation and fusion module, a hierarchical boundary modeling module, a boundary co-generation module, and a structural deviation evaluation module.

[0034] During the training phase, graph structure sparsity processing is performed on the compact state structure graph of the normal sample training subset, and label offset processing is performed on the structure labels to generate a perturbation sample set. Based on the perturbation sample set, an approximate sample set outside the boundary is constructed. The normal sample training subset, the perturbation sample set, and the approximate sample set outside the boundary are used together as the training input set.

[0035] Three soft class-one extreme learning machine sub-models are trained based on different structural feature subspaces. The structural feature subspaces include the global structural feature subspace, the local adjacency structural feature subspace, and the skip label path structural feature subspace. Each soft class-one extreme learning machine model completes boundary learning based on the training input set and obtains the corresponding class-one discrimination boundary.

[0036] The weight parameters corresponding to each soft class I extreme learning machine sub-model are determined based on the normal sample validation subset. Weighted ensemble is then performed on each soft class I extreme learning machine sub-model to obtain the trained improved soft class I extreme learning machine model.

[0037] Optionally, the output corresponding to the structural deviation index of the fastening state includes:

[0038] The feature vector of the tight state structure is input into the feature carrying module of the trained improved soft first-class extreme learning machine model. The feature vector of the tight state structure is subjected to feature dimension consistency verification and structural formatting to form the input feature representation.

[0039] The mapping generation and fusion module performs hierarchical channel mapping processing on the input feature representation, dividing the input feature representation into three structural sub-channels according to structural semantics. In each structural sub-channel, a parallel nonlinear mapping is performed using random mapping kernels corresponding to different activation functions to generate the hidden layer mapping results corresponding to each structural sub-channel and perform reconstruction and fusion processing to form a unified intermediate representation vector.

[0040] The hierarchical boundary modeling module constructs soft first-class discrimination boundaries corresponding to different structural levels based on the intermediate representation vector, and substitutes the intermediate representation vector into the soft first-class discrimination boundary corresponding to each structural level to perform boundary discrimination calculation, generating the deviation of the intermediate representation vector from the soft first-class discrimination boundary of each structural level.

[0041] The boundary collaborative generation module stacks the deviations corresponding to each structural level according to the structural level dimension to construct a three-dimensional boundary stacking representation, and performs cross-layer aggregation processing on the three-dimensional boundary stacking representation to form a unified first-class discriminative boundary representation;

[0042] The structural deviation assessment module is based on a unified single-class discrimination boundary representation. It calculates the structural deviation components corresponding to the input feature representation under different structural perspectives, generates a structural deviation vector, performs mutually exclusive projection processing on the structural deviation vector, generates projection residual components between each structural perspective, performs multi-dimensional fusion operation on the projection residual components, and outputs the structural deviation risk index corresponding to the structural feature vector of the fastened state as the structural deviation degree index of the fastened state.

[0043] Optionally, the step of generating the bolt tightening status determination result and writing it into the status history database is performed.

[0044] To associate the structural deviation index of the fastened state with the corresponding bolt identifier, structural node identifier and time index, a record to be judged is formed.

[0045] A hierarchical judgment rule is constructed based on a hierarchical threshold set, wherein the hierarchical threshold set includes a first threshold and a second threshold, and the first threshold is less than the second threshold;

[0046] Based on the graded judgment rule, the deviation index of the fastened state structure is compared with the first threshold and the second threshold. When the deviation index of the fastened state structure is less than the first threshold, a normal fastening state judgment result is generated.

[0047] When the deviation index of the fastened state structure is greater than or equal to the first threshold and less than the second threshold, a slightly loose transition state judgment result is generated; when the deviation index of the fastened state structure is greater than or equal to the second threshold, an abnormal risk state judgment result is generated.

[0048] Write the record to be judged, the bolt tightness judgment result, and the structural deviation index of the tightness state into the state history database, and store the historical judgment record of the corresponding bolt in the state history database according to the time index.

[0049] The beneficial effects of this invention are:

[0050] This invention proposes an intelligent monitoring method for bolt tightness based on machine learning. It comprehensively utilizes multi-source monitoring data fusion, structured feature modeling, and a class-aspect learning anomaly detection technique to achieve online intelligent determination of bolt tightness. The method collects multi-source data of the bolted connection structure during operation, performs unified preprocessing and fragmented feature extraction to construct a tightness state structure diagram, and employs the Weisfeiler-Lehman algorithm to characterize the multi-level structural features of the diagram, fully exploring the temporal and structural changes in different bolt nodes and their relationships.

[0051] Furthermore, by combining an improved soft-class extreme learning machine model, the soft boundary of the structure in the fastened state is learned based on samples of normal fastening states, enabling quantitative assessment and classification of the degree of deviation of the structure in the fastened state. This invention can effectively distinguish between fluctuations in the normal fastening state and actual loosening anomalies under complex working conditions, even when abnormal samples are scarce or missing. It significantly reduces the risk of false alarms and missed alarms, and possesses beneficial effects such as strong adaptability to changes in working conditions, full utilization of structural information, and stable and reliable monitoring results. It is suitable for long-term online intelligent monitoring of bolted connection structures in complex engineering environments. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a machine learning-based intelligent monitoring method for bolt fastening status proposed in this invention.

[0054] Figure 2 This is a flowchart illustrating the execution of the Weisfeiler-Lehman algorithm in the intelligent monitoring method for bolt fastening status based on machine learning proposed in this invention.

[0055] Figure 3 This is a schematic diagram of the structure of an improved soft first-class extreme learning machine model for a machine learning-based intelligent monitoring method for bolt fastening status proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1 , Figure 2 and Figure 3 A machine learning-based intelligent monitoring method for bolt fastening status includes:

[0058] Multi-source fastening status data of bolted connection structures during operation are collected, and the multi-source fastening status data is preprocessed to obtain standardized fastening status data.

[0059] The standardized fastening state data is segmented into fragments to obtain a set of fastening state fragments. For each fastening state fragment, time-domain features, frequency-domain features, and statistical features are extracted and concatenated to form a fragment-level fastening state feature vector.

[0060] Based on the segment-level fastening state feature vectors, a fastening state feature vector sequence is formed, and a fastening state structure diagram is constructed.

[0061] The Weisfeiler-Lehman algorithm is executed on the tight state structure graph. The node labels of each graph node are initialized by numerical interval mapping based on the fragment-level tight state feature vector. The node labels are combined with the labels of adjacent nodes and hash updates are performed. After completing a preset number of iterations, the tight state structure feature vector is formed.

[0062] A normal sample set is constructed based on the historical fastening state structure feature vector. An improved soft first-class extreme learning machine model is constructed and trained. The real-time fastening state structure feature vector is input into the trained improved soft first-class extreme learning machine model, and the corresponding fastening state structure deviation index is output.

[0063] Based on the structural deviation index of the fastening state and the preset classification judgment rules, the bolt fastening state judgment result is generated and written into the state history database.

[0064] In this embodiment, the multi-source fastening state data specifically includes vibration signal data, acoustic emission signal data, strain-related data, structural state auxiliary data, and working condition and environmental data.

[0065] In this embodiment, the preprocessing of multi-source fastening state data specifically includes data synchronization and alignment, noise reduction and filtering, trend and drift correction, and amplitude unification and standardization.

[0066] In this embodiment, forming a segment-level fastening state feature vector includes:

[0067] The standardized values ​​of various data types in the standardized fastening state data at each sampling time are arranged in chronological order to form a continuous data sequence;

[0068] The continuous data sequence is segmented according to a preset segment length and a preset segment step size to generate a set of tight-state segments consisting of multiple independent and temporally continuous data subsequences, wherein:

[0069] The preset segment length is set to 1024 sampling points. Each tight segment consists of 1024 consecutive sampling points, corresponding to a data analysis window of a fixed time length.

[0070] The preset segment step size is set to 512 sampling points, and the starting sampling positions of two adjacent tight state segments differ by 512 sampling points;

[0071] Based on the set of fastened state segments, feature extraction processing is performed on the continuous data sequences within the corresponding time intervals of each fastened state segment to obtain the corresponding time-domain feature set, frequency-domain feature set, and statistical feature set. The specific execution of the feature extraction processing is as follows:

[0072] For each source signal sequence within the time interval corresponding to each tight state segment, calculate the time domain characteristics, frequency domain characteristics, and statistical characteristics. The time domain characteristics include the mean, root mean square, and peak value. The mean is the arithmetic mean of each sample value within the tight state segment. The root mean square is the square root of the average of the squares of each sample value within the tight state segment. The peak value is the maximum absolute value of the sample value within the tight state segment.

[0073] The frequency domain characteristics include the dominant frequency, the dominant frequency amplitude, and the frequency band energy. The dominant frequency and the dominant frequency amplitude are obtained by performing a fast Fourier transform on the signal of the tight state segment to obtain the amplitude spectrum, and then taking the frequency and amplitude corresponding to the maximum amplitude on the amplitude spectrum. The frequency band energy is obtained by summing the squares of the amplitudes of the amplitude spectrum within the frequency band.

[0074] The statistical characteristics include the maximum value, minimum value, and standard deviation. The maximum value and minimum value are the maximum and minimum values ​​of the sampled values ​​within the fastened state segment, respectively. The standard deviation is a value calculated from the dispersion of the sampled values ​​within the fastened state segment relative to the mean.

[0075] The time-domain feature set, frequency-domain feature set, and statistical feature set corresponding to each fastening state segment are combined sequentially to generate a corresponding segment-level fastening state feature vector. Specifically, generating the corresponding segment-level fastening state feature vector involves:

[0076] For each fastened state segment, the time-domain feature set, frequency-domain feature set, and statistical feature set of the fastened state segment are combined to generate a segment-level fastened state feature vector. Within the same data source, the features are arranged in the order of time-domain features, frequency-domain features, and statistical features. All the arranged features are then concatenated into a one-dimensional vector to obtain the corresponding segment-level fastened state feature vector.

[0077] In this embodiment, constructing the fastened state structure diagram includes:

[0078] Based on the segment-level fastening state feature vectors, and arranged according to the sampling time order corresponding to each segment-level fastening state feature vector, a fastening state feature vector sequence is formed.

[0079] The node set of the fastening state structure diagram is determined based on the fastening state feature vector sequence. Each segment-level fastening state feature vector is mapped to a node to obtain the node set, and the segment-level fastening state feature vector is used as the node attribute of the corresponding node. Specifically, mapping each segment-level fastening state feature vector to a node to obtain the node set involves:

[0080] A unique node identifier is assigned to each segment-level fastening state feature vector in the fastening state feature vector sequence. The node identifier is obtained by combining the bolt identifier, segment time index and segment number. A node set is constructed in such a way that one segment-level fastening state feature vector corresponds to one node. After completing the traversal of the entire fastening state feature vector sequence, the node set is obtained.

[0081] The edge set of the fastening state structure diagram is determined based on the node set. Temporal adjacency connection is established between adjacent nodes corresponding to the same bolt. Spatial adjacency connection is established between nodes corresponding to different bolts when the bolts are spatially adjacent to each other. Structural connection relationship is established when the bolts belong to the same connection structure and have a direct structural connection relationship.

[0082] The node set, edge set, and node attributes are integrated to generate a tight state structure graph. Specifically, generating the tight state structure graph involves:

[0083] Write the node set into the tight state structure graph one by one to complete node registration. Write the fragment-level tight state feature vector corresponding to each node into the node attribute field of the node. Write the edge set into the tight state structure graph one by one. The writing includes specifying the start node identifier and end node identifier of each edge and recording the connection type identifier in the edge attribute for different types of connection relationships. After all nodes and edges are written, the tight state structure graph is obtained.

[0084] In this embodiment, the step of forming a tight-state structural feature vector after completing a preset number of iterations includes:

[0085] Based on the compacted state structure graph, the Weisfeiler-Lehman algorithm is executed to perform multidimensional interval encoding on the fragment-level compacted state feature vectors corresponding to each node in the node set. Each feature value is mapped to its corresponding interval encoding index according to a preset interval partitioning rule, and then arranged into ordered multidimensional encoding labels according to the dimensional order, serving as the initial structure label for each node.

[0086] The preset interval partitioning rule refers to discretizing and mapping each segment-level compact state feature value using a fixed number of intervals. The number of intervals is set to 16. For each feature dimension, the minimum and maximum values ​​on its normal sample training subset are calculated, and the numerical range between the minimum and maximum values ​​is divided into 16 consecutive intervals. Each interval corresponds to a unique interval coding index, forming the interval boundary table and interval coding table of the feature dimension. After obtaining the interval boundary tables of all feature dimensions, the interval boundary tables are used as the preset interval partitioning rule.

[0087] The execution of the multidimensional interval encoding process is specifically as follows:

[0088] For any node in the fastened state structure diagram, read the fragment-level fastened state feature vector of the node, and perform interval encoding dimension by dimension in order of feature dimensions. For each feature dimension, determine the interval number according to the corresponding interval boundary table, and use the interval number as the interval encoding index. When the feature value of the dimension is less than the minimum interval boundary of the dimension, the interval encoding index is set to 0. When the feature value of the dimension is greater than the maximum interval boundary of the dimension, the interval encoding index is set to 15. After completing the interval encoding for all feature dimensions of the node, a hexadecimal integer index sequence arranged in order of dimension is obtained, and the ordered multidimensional encoding label of the node is obtained. This label is written into the node label field as the initial structure label of the node.

[0089] Under the current structural refinement iteration, the set of adjacent nodes connected to each node is obtained based on the edge set. The structural labels of each adjacent node in the adjacent node set are arranged to form an ordered adjacent label sequence. The ordered adjacent label sequence is truncated and padded to obtain a fixed-length adjacent label sequence table. The truncating and padding of the ordered adjacent label sequence specifically involves:

[0090] A target length of 16 is preset for the ordered adjacency label sequence of each node. The structural labels of each neighboring node in the set of neighboring nodes of a node are arranged to form an ordered adjacency label sequence. When the length of the ordered adjacency label sequence is greater than 16, the first 16 adjacency structural labels of the sorted sequence are taken as the truncation result. When the length of the ordered adjacency label sequence is less than 16, padding labels are added to the end of the sequence in order of position until the length reaches 16. The padding labels are set to all-zero multidimensional encoded labels.

[0091] Under the current structural refinement iteration, the structural label of each node is combined with the corresponding fixed-length adjacency label sequence list, and a hash update is performed on the combination result to update the structural label of each node. The execution of the hash update is specifically as follows:

[0092] For any node in the tight state structure diagram, read the node's structure label under the current structure refinement iteration, and read the fixed-length adjacent label sequence corresponding to the node. Connect the structure label and the fixed-length adjacent label sequence sequentially to generate a combined sequence. Convert the combined sequence into a fixed-format encoded string. The fixed format is to connect the dimensional interval encoding indices with separators and insert label separators between adjacent labels. Perform a deterministic hash operation on the encoded string to obtain a hash value. Convert the encoded string into the corresponding integer encoding sequence in character order. Perform a weighted accumulation operation on the integer encoding sequence sequentially, and perform a modulo operation after each accumulation to limit the value range, obtaining a hash value uniquely corresponding to the encoded string. Perform a modulo mapping on the hash value to generate an updated label index. Convert the updated label index into a fixed-length label representation and write it into the node's structure label field as the node's updated structure label in the next structure refinement iteration.

[0093] Repeat the structural refinement iteration until the preset number of iterations is reached. Stack the structural labels obtained in each structural refinement iteration in the order of the iterations, constructing a node label iteration stacking matrix. Select the initial structural label and the structural label after completing the preset number of iterations for each node according to the preset skip-level iterations, generating the skip-level label path for each node, forming a skip-level label aggregation structure, where:

[0094] The preset number of iterations is set to 5. The structural refinement iteration starts from the initial structural label state and continuously executes 5 rounds of structural refinement iteration updates.

[0095] The construction of the node label iterative stacking matrix is ​​specifically as follows:

[0096] For each node in the fastened state structure diagram, the structure labels obtained by the node in the initial state and after the first to fifth rounds of structural refinement iteration are read in sequence. The structure labels are arranged in the order of the iteration rounds to form the node label sequence. The label sequences of all nodes in the fastened state structure diagram are stacked side by side according to the node identifier order to obtain the node label iteration stack matrix. The row index of the node label iteration stack matrix corresponds to the node identifier, the column index corresponds to the iteration round, and the element of the node label iteration stack matrix is ​​the structure label of the corresponding node in the corresponding iteration round.

[0097] The preset skipping rounds are set to the initial round and the 5th round of structural refinement iteration. The two structural labels used for skipping are the initial structural label of the node and the structural label of the node after completing the 5th round of structural refinement iteration.

[0098] The specific steps for generating the jump-level label paths for each node are as follows:

[0099] For any node in the fastened state structure diagram, extract the initial structure label of the node and the structure label after the fifth round of structural refinement iteration from the node label iterative stacking matrix row corresponding to the node, and connect the two in order to form a jump layer splicing label sequence. Write the jump layer splicing label sequence into the jump layer label field of the node to obtain the jump layer label path of the node.

[0100] The formation of the jump-level tag aggregation structure specifically includes:

[0101] The skip label paths of all nodes in the tight state structure diagram are collected in the order of node identification to form a skip label path set. The skip label path set is then associated with the node label iterative stacking matrix and stored together. Each node corresponds to both the cross-layer stacking label sequence and the skip label path. The combined data structure of the node label iterative stacking matrix and the skip label path set is determined as the skip label aggregation structure.

[0102] Based on the node label iterative stacking matrix, the skip label aggregation structure, and the compacted state structure graph, a three-dimensional tensor graph is constructed. Then, based on this three-dimensional tensor graph, multi-dimensional aggregation processing is performed on the structure labels corresponding to different nodes, adjacency relationships, and structural refinement iterations. The aggregation result is mapped to a one-dimensional vector representation to generate a compacted state structure feature vector, where:

[0103] The construction of the three-dimensional tensor graph specifically involves:

[0104] Using the node set of the compact state structure graph as the first-dimensional index, the structure refinement iteration round as the second-dimensional index, and the fixed-length adjacency position index of the node as the third-dimensional index, a three-dimensional tensor data object is created. For any node and any iteration round, the structure label of the node under the iteration round is read, and the fixed-length adjacency label sequence of the node is read. The structure label is written as the center label value of the node under the iteration round and written to the adjacency position index 1 of the three-dimensional tensor. The adjacent structure labels arranged in order in the fixed-length adjacency label sequence are written to the adjacency position indexes 2 to 16 of the three-dimensional tensor. The skip label path of the node is associated with the structure label corresponding to the iteration round and written to the additional field of the three-dimensional tensor graph. After writing all nodes, all iteration rounds, and all adjacency positions, the three-dimensional tensor graph is obtained.

[0105] The generation of the fastened state structural feature vector is specifically as follows:

[0106] Aggregation is performed on the 3D tensor graph along the adjacent position index dimension. To count and summarize the structural label indices along the adjacent position index dimension and form a fixed-length count vector, cross-layer aggregation is performed on the adjacent aggregation representation along the structural refinement iteration round dimension. The fixed-length count vectors under each iteration round are concatenated in the round order to form a cross-layer representation vector. The cross-layer representation vector is then concatenated with the jump layer label path encoding vector corresponding to the node to obtain the node-level structural representation vector. Graph-level aggregation is performed on the node-level structural representation vectors of all nodes in the compact state structure graph along the node dimension. The graph-level average vector is obtained by summing all node-level structural representation vectors dimension by dimension and dividing by the number of nodes, and is used as the compact state structure feature vector.

[0107] In this embodiment, constructing and training the improved soft first-class extreme learning machine model includes:

[0108] Select the structural feature vectors of historical fastening states that are in a confirmed normal fastening state, aggregate them according to time index to form a normal sample set, and divide the normal sample set into a normal sample training subset and a normal sample verification subset;

[0109] An improved soft first-class extreme learning machine (Soft First-Class Extreme Learning Machine) model is constructed. This improved Soft First-Class Extreme Learning Machine model comprises a feature-carrying module, a mapping generation and fusion module, a hierarchical boundary modeling module, a boundary co-generation module, and a structural deviation evaluation module. Specifically, the construction of the improved Soft First-Class Extreme Learning Machine model involves:

[0110] Based on the basic learning chain of an adjacency aggregation representation of input features, random mapping latent, output layer parameter solving, and deviation output, an improved soft-class extreme learning machine model is obtained by adding dimensionality consistency verification and structural formatting processing to the input feature module. Based on the single-channel random mapping latent module, the input feature representation is divided into three structural sub-channels according to structural semantics to obtain the mapping generation fusion module. Based on the single soft-class boundary construction module, a hierarchical boundary modeling module and a boundary co-generation module are formed. Based on the deviation output module, the calculation of structural deviation components under different structural perspectives is introduced to expand and form a structural deviation evaluation module.

[0111] During the training phase, graph structure sparsity processing is performed on the compact state structure graph of the normal sample training subset, and label offset processing is performed on the structure labels to generate a perturbed sample set. Based on the perturbed sample set, an out-of-bounds approximate sample set is constructed. The normal sample training subset, the perturbed sample set, and the out-of-bounds approximate sample set are used together as the training input set, where:

[0112] The generated perturbation sample set is specifically as follows:

[0113] For each compact state structure graph in the normal sample training subset, graph structure sparsity processing is performed. The edge set of the compact state structure graph is processed according to edge type. Temporally adjacent connection edges are kept unchanged. Spatially adjacent connection edges are kept only by the nearest neighbor edge corresponding to each node according to the spatial proximity relationship of the nodes. Structural connection relationship edges are kept only by the direct connection relationship in the connection structure, resulting in a sparse edge set and generating a sparse structure graph.

[0114] Structural label offset processing is performed on the sparse structure graph. The initial structural labels of each node in the sparse structure graph are perturbed dimension by dimension. The interval coding index of each dimension in the multi-dimensional interval coding label is offset by the offset amount and truncated and normalized within the range of interval coding index values ​​to generate the offset initial structural label. The compact state structure graph after the graph structure sparsity processing and structural label offset processing is taken as a perturbation sample. All compact state structure graphs in the normal sample training subset are processed one by one to obtain the perturbation sample set.

[0115] The construction of the approximate sample set outside the boundary is specifically as follows:

[0116] Based on the structure diagram of each perturbation sample in the perturbation sample set, the corresponding perturbation tight state structure feature vector is generated, and the amplitude amplification processing is performed on the perturbation tight state structure feature vector. The feature values ​​of each dimension of the perturbation tight state structure feature vector are multiplied and scaled by the same amplification factor to obtain the feature vector of the approximate sample outside the boundary, which is used as the approximate sample outside the boundary. The same processing is performed on all perturbation samples in the perturbation sample set one by one to obtain the approximate sample set outside the boundary.

[0117] Three soft-class extreme learning machine (Soft-Class IEM) sub-models are trained based on different structural feature subspaces. These structural feature subspaces include a global structural feature subspace, a local adjacency structural feature subspace, and a skip-level label path structural feature subspace. Each Soft-Class IEM model performs boundary learning based on the training input set to obtain a corresponding Class I discrimination boundary. Specifically:

[0118] Each sample in the training input set is mapped to the global structural feature subspace, the local adjacency structural feature subspace, and the skip label path structural feature subspace, resulting in three sets of subspace training data.

[0119] When training a global sub-model for a global structural feature subspace, the global subspace training data is input into the mapping generation and fusion module of the corresponding soft first-class extreme learning machine sub-model to generate an intermediate representation vector of the global sub-model. Based on the intermediate representation vector, a class of soft boundary constraints of the global sub-model is constructed and the output layer parameters are solved to obtain a class of discriminative boundaries of the global sub-model.

[0120] When training the local adjacency sub-model for the local adjacency structure feature subspace, the same training process is used to obtain the first-class discrimination boundary of the local adjacency sub-model. When training the skip-level path sub-model for the skip-level label path structure feature subspace, the same training process is used to obtain the first-class discrimination boundary of the skip-level path sub-model. After training the three sets of subspaces, the first-class discrimination boundaries corresponding to the global sub-model, the local adjacency sub-model and the skip-level path sub-model are obtained respectively.

[0121] The weight parameters for each soft first-class extreme learning machine (Soft First-Class Extreme Learning) sub-model are determined based on a normal sample validation subset. A weighted ensemble is then performed on each Soft First-Class Extreme Learning sub-model to obtain the trained improved Soft First-Class Extreme Learning model, where:

[0122] The determination of weight parameters for each soft-class extreme learning machine sub-model based on a normal sample validation subset is as follows:

[0123] Each sample in the normal sample validation subset is input into the global structural feature sub-model, the local adjacency structural feature sub-model, and the skip label path structural feature sub-model, respectively, to obtain the deviation of the three sub-models from the sample output. The same calculation is repeated for all samples in the normal sample validation subset to obtain the deviation sequence of each sub-model on the validation subset. The mean of the deviation sequence of each sub-model is calculated as the validation deviation mean of the sub-model. The reciprocal of the validation deviation mean of the three sub-models is taken and normalized to obtain the weight parameters corresponding to the three soft class one extreme learning machine sub-models.

[0124] The weighted ensemble of each soft-class extreme learning machine sub-model is specifically performed as follows:

[0125] The global structural feature sub-model, the local adjacency structural feature sub-model, and the skip label path structural feature sub-model are used as three sub-model components of the improved soft first-class extreme learning machine model. The three weight parameters are bound to the corresponding sub-model components. The boundary discrimination outputs of the three sub-model components are connected in parallel to a unified ensemble output. At the ensemble output, the deviations of the outputs of the three sub-models are weighted and summed according to the weight parameters of each sub-model to form the ensemble deviation calculation rule. The combined model containing the three sub-model components, the three weight parameters, and the ensemble deviation calculation rule is determined as the trained improved soft first-class extreme learning machine model.

[0126] In this embodiment, the output corresponding fastening state structural deviation index includes:

[0127] The feature vector of the fastened state structure is input into the feature-carrying module of the trained improved soft first-class extreme learning machine model. Feature dimension consistency verification and structural formatting are performed on the feature vector of the fastened state structure to form an input feature representation. Specifically, forming the input feature representation involves:

[0128] Read the real-time fastened state structure feature vector and obtain the feature dimension number. Compare the feature dimension number with the target dimension number fixed during the training phase. When the feature dimension number is less than the target dimension number, pad the end of the fastened state structure feature vector with padding values ​​in dimension order until the target dimension number is reached. When the feature dimension number is greater than the target dimension number, extract the first target dimension number of feature values ​​in dimension index order as the consistency result. Rearrange the fastened state structure feature vector after consistency processing in dimension order to obtain the input feature representation.

[0129] The mapping generation and fusion module performs hierarchical channel mapping processing on the input feature representation, dividing the input feature representation into three structural sub-channels according to structural semantics. In each structural sub-channel, a parallel nonlinear mapping is performed using random mapping kernels corresponding to different activation functions, generating the hidden layer mapping results for each structural sub-channel. This is then reconstructed and fused to form a unified intermediate representation vector, where:

[0130] The execution of the hierarchical channel mapping process is specifically as follows:

[0131] The input feature representation is divided into three structural sub-channels according to the set of dimension indices. The global structural semantic sub-channel is composed of the feature values ​​corresponding to the set of dimension indices representing the graph-level overall statistics in the input feature representation. The local adjacency structural semantic sub-channel is composed of the feature values ​​corresponding to the set of dimension indices representing the adjacency aggregation statistics in the input feature representation. The jump label path structural semantic sub-channel is composed of the feature values ​​corresponding to the set of dimension indices representing the jump label path aggregation in the input feature representation. The sub-vectors of the three structural sub-channels are written into the input ports of the corresponding channels to obtain the channel input representations of the three structural sub-channels.

[0132] The execution of the reconstruction and fusion process is specifically as follows:

[0133] The channel input representations of the three structural sub-channels are subjected to linear transformation and activation transformation of random mapping kernels respectively to obtain the hidden layer mapping results corresponding to the three structural sub-channels. The results are then concatenated according to the channel order to obtain the concatenated hidden representation. The concatenated hidden representation is then linearly reconstructed according to the output dimension fixed during the training phase. The reconstructed vector is used as a unified intermediate representation vector.

[0134] S523, the hierarchical boundary modeling module constructs soft first-class discrimination boundaries corresponding to different structural levels based on the intermediate representation vector, and substitutes the intermediate representation vector into the soft first-class discrimination boundaries corresponding to each structural level to perform boundary discrimination calculations, generating a value representing the deviation of the intermediate representation vector from the soft first-class discrimination boundaries of each structural level, wherein:

[0135] The construction of soft class discrimination boundaries corresponding to different structural levels is specifically as follows:

[0136] Boundary parameter sets are established for three structural levels: the structural level corresponding to the initial structural label, the structural level after the fifth round of structural refinement iteration, and the structural level corresponding to the jump label path. For any structural level, the intermediate representation vector obtained by the normal sample training subset under the structural level is used as input. The number of hidden nodes is fixed and the hidden layer input weights and biases are randomly generated to obtain the hidden representation results. Based on the hidden representation results, a class of discrimination targets is constructed by constraining the deviation of normal samples from the discrimination boundary. Based on the class of discrimination targets, a class of discrimination parameters for the corresponding structural level are determined to form a class of discrimination boundary for the structural level.

[0137] The boundary discrimination calculation is performed as follows:

[0138] For the intermediate representation vector, the first-class discrimination boundary parameters of the three structural levels are called respectively. The intermediate representation vector is input into the discrimination functions of the initial structural level boundary, the fifth round structural level boundary and the jump path structural level boundary respectively to obtain the discrimination output values ​​corresponding to the three structural levels. The discrimination output value of each structural level is compared with the boundary threshold of the structural level to calculate the deviation of the intermediate representation vector from the soft first-class discrimination boundary of the structural level. The deviation of the initial structural level, the deviation of the fifth round structural level and the deviation of the jump path structural level are output respectively.

[0139] The boundary collaborative generation module stacks the deviations corresponding to each structural level according to the structural level dimension to construct a three-dimensional boundary stacking representation, and performs cross-level aggregation processing on the three-dimensional boundary stacking representation to form a unified discrimination boundary representation, wherein:

[0140] The construction of the three-dimensional boundary stacking representation is specifically as follows:

[0141] A three-dimensional stacked data object is created with the structural hierarchy as the first dimension index, the three types of structural semantic sub-channels as the second dimension index, and the deviation as the third dimension index. The first dimension corresponds to the initial structural hierarchy, the fifth round structural hierarchy, and the jump path structural hierarchy in turn. The second dimension corresponds to the global structural semantic sub-channel, the local adjacent structural semantic sub-channel, and the jump label path structural semantic sub-channel in turn. The third dimension is the position of the corresponding deviation value. For each structural hierarchy, the deviation output by the layer boundary modeling module under the structural hierarchy is written into the corresponding position of the three-dimensional stacked data object to obtain the three-dimensional boundary stacking representation.

[0142] The execution of the cross-layer aggregation process is specifically as follows:

[0143] A weighted summation operation is performed on the three-dimensional boundary stacking representation along the structural hierarchy dimension to obtain the aggregated deviation corresponding to each structural semantic sub-channel. The aggregated deviations corresponding to the three structural semantic sub-channels are then concatenated in sequence to obtain a unified class-aware discrimination boundary representation.

[0144] The structural deviation assessment module, based on a unified single-class discrimination boundary representation, calculates the structural deviation components corresponding to the input feature representation under different structural perspectives, generates a structural deviation vector, performs mutually exclusive projection processing on the structural deviation vector to generate projection residual components between each structural perspective, performs multi-dimensional fusion operations on the projection residual components, and outputs the structural deviation risk index corresponding to the structural feature vector in the fastened state as an indicator of the structural deviation degree in the fastened state, wherein:

[0145] The computational input features represent the structural deviation components corresponding to different structural perspectives, specifically:

[0146] The input feature representation is divided into global structural perspective sub-vectors, local adjacent structural perspective sub-vectors, and skip path structural perspective sub-vectors according to the dimensional index set. The boundary discrimination benchmark values ​​corresponding to the three structural perspectives are read from the unified first-class discrimination boundary representation. For each structural perspective, the corresponding boundary discrimination benchmark value is subtracted from the discrimination output value corresponding to the structural perspective to obtain the structural deviation component. When the structural deviation component is less than 0, the structural deviation component is set to 0, thus obtaining the global structural deviation component, the local adjacent structural deviation component, and the skip path structural deviation component.

[0147] The execution of the mutual exclusion projection process is specifically as follows:

[0148] Based on the global structural deviation component, the local adjacent structural deviation component, and the skip path structural deviation component, one component is taken as the principal component, and the remaining two components are taken as the comparison components to perform projection residual calculation. The projection value of the comparison component in the direction of the principal component is subtracted from the comparison component. Projection residual calculation is performed on the three cases of global structural deviation component as principal component, local adjacent structural deviation component as principal component, and skip path structural deviation component as principal component, respectively, to obtain three sets of projection residual components. The three sets of projection residual components are collected in order to form a projection residual component set.

[0149] The execution of the multidimensional fusion operation is specifically as follows:

[0150] The three structural deviation components of the structural deviation vector and the projection residual components are used as the fusion input. The fusion input is normalized to obtain dimensionless components. The dimensionless structural deviation components are weighted and summed, and the dimensionless projection residual components are weighted and summed to obtain the comprehensive value of structural deviation and the comprehensive value of structural conflict. The comprehensive value of structural deviation and the comprehensive value of structural conflict are then weighted and summed again to obtain the structural deviation risk index, which serves as the indicator of structural deviation degree in the fastened state.

[0151] In this embodiment, the step of generating the bolt tightening status determination result and writing it into the status history database is described.

[0152] To associate the structural deviation index of the fastened state with the corresponding bolt identifier, structural node identifier and time index, a record to be judged is formed.

[0153] A hierarchical judgment rule is constructed based on a hierarchical threshold set, wherein the hierarchical threshold set includes a first threshold and a second threshold, and the first threshold is less than the second threshold. Specifically, the hierarchical judgment rule is constructed based on the hierarchical threshold set as follows:

[0154] Using the deviation index of the fastened state structure as the judgment input, and setting the first threshold and the second threshold as the classification boundary, the value range of the deviation index of the fastened state structure is divided into three non-overlapping judgment ranges according to the first threshold and the second threshold, and the risk status is judged to obtain the classification judgment rule.

[0155] Based on the graded judgment rule, the deviation index of the fastened state structure is compared with the first threshold and the second threshold. When the deviation index of the fastened state structure is less than the first threshold, a normal fastening state judgment result is generated.

[0156] When the deviation index of the fastened state structure is greater than or equal to the first threshold and less than the second threshold, a slightly loose transition state judgment result is generated; when the deviation index of the fastened state structure is greater than or equal to the second threshold, an abnormal risk state judgment result is generated.

[0157] Write the record to be judged, the bolt tightness judgment result, and the structural deviation index of the tightness state into the state history database, and store the historical judgment record of the corresponding bolt in the state history database according to the time index.

[0158] Example 1: To verify the feasibility of this invention in practice, it was applied to the operation monitoring scenario of the bolted connection structure of the wind turbine tower flange in a coastal wind farm. This wind farm is located in a coastal area, where the wind turbines are subjected to complex operating conditions of strong winds, high humidity, and significant diurnal temperature variations. The tower flange has a large number of high-strength bolts. In actual operation, there is a problem of slow decay of preload due to vibration loads and temperature changes. However, very few clear samples of loosening abnormalities can be obtained on-site. Traditional monitoring methods based on fixed thresholds frequently generate false alarms, making it difficult to meet operation and maintenance needs.

[0159] In this scenario, multi-source fastening status data of the bolted connection structure during operation were simultaneously collected and preprocessed in conjunction with operating condition data such as wind speed and ambient temperature. Data was continuously collected for three months at a sampling frequency of 2000 times per second. After fragmentation of the data, fragment-level fastening status features were extracted, and a fastening status structure diagram was constructed. The Weisfeiler-Lehman algorithm was used to generate fastening status structure feature vectors. Without artificially creating loosening samples, only data from the first two months confirmed to be in a normal fastening state were selected as normal samples. An improved soft first-class extreme learning machine model was constructed and trained to establish the soft boundary of the structure in the normal fastening state.

[0160] During the third month of online monitoring, a structural deviation index for the tightening status was output for each time segment. Monitoring results showed that under normal strong wind conditions, the structural deviation index remained stable between 0.18 and 0.32, all considered normal tightening. When one bolt showed a slight decrease in preload after 68 days of continuous operation, the structural deviation index gradually increased to 0.47, and was identified as a slight loosening transition state. Approximately 36 hours before maintenance personnel manually checked the bolt, the structural deviation index further increased to 0.62, triggering an anomaly risk warning. On-site verification showed that the bolt preload had decreased by approximately 18%, verifying the accuracy of the monitoring results. This invention achieves reliable, low-false-report online monitoring of bolt tightening status under conditions of scarce abnormal samples.

[0161] Table 1. Comparison of Intelligent Monitoring Results of Wind Turbine Tower Flange Bolt Tightness Status

[0162] Monitoring time (number of days of operation) Average wind speed (m / s) Structural Deviation Index Judgment Result Does this trigger an alert? Measured preload change rate (%) Traditional threshold method determination results Day 12 8.4 0.21 Normal tightening no -1.2 normal Day 18 9.7 0.27 Normal tightening no -1.8 Minor abnormality Day 25 10.3 0.29 Normal tightening no -2.1 abnormal Day 34 11.6 0.31 Normal tightening no -2.5 abnormal Day 46 9.1 0.33 Normal tightening no -3.0 Minor abnormality Day 58 7.8 0.38 Normal tightening no -5.6 normal Day 62 8.9 0.44 Transition state no -9.3 normal Day 65 10.8 0.47 Transition state no -12.5 normal Day 67 11.2 0.53 Transition state no -15.8 Minor abnormality Day 68 9.5 0.58 Abnormal risks yes -18.2 abnormal Day 69 8.7 0.62 Abnormal risks yes -19.6 abnormal

[0163] As shown in Table 1, during the 69-day continuous monitoring process, the method of this invention consistently outputs the structural deviation index of the tightened state under different wind speeds and operating conditions, and provides stable and consistent state judgment results. From day 12 to day 58, the average wind speed fluctuated between 7.8 m / s and 11.6 m / s, and the structural deviation index remained consistently within the range of 0.21 to 0.38. Although wind speeds were higher and vibrations increased at times, the method of this invention consistently judged the condition as normal and did not trigger any abnormal warnings. In contrast, the traditional threshold method repeatedly gave minor or even abnormal judgments during this period, reflecting that the method of this invention has stronger robustness to normal operating condition fluctuations and a lower tendency for false alarms.

[0164] As the operation progressed, from day 62 to day 67, the structural deviation index gradually increased to the range of 0.44 to 0.53. The method of this invention accurately identified this stage as a slight loosening transition state, reflecting the trend of bolt tightening status evolving from stable to abnormal. In contrast, the traditional threshold method repeatedly judged this stage as normal, failing to reflect potential risks in a timely manner. By day 68 and day 69, the structural deviation index reached 0.58 and 0.62, respectively. The method of this invention promptly triggered an abnormal risk warning. At this time, the measured preload change rate had reached -18.2% and -19.6%, verifying the authenticity of the anomaly judgment results. In summary, the method of this invention can not only identify bolt loosening risks in advance under conditions of scarce abnormal samples, but also did not produce significant false alarms throughout the entire monitoring period, fully demonstrating the practical application value of the method for intelligent monitoring of bolt tightening status under complex working conditions.

[0165] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based intelligent monitoring method for bolt fastening status, characterized in that, include: Multi-source fastening status data of bolted connection structures during operation are collected, and the multi-source fastening status data is preprocessed to obtain standardized fastening status data. The standardized fastening state data is segmented into fragments to obtain a set of fastening state fragments. For each fastening state fragment, time-domain features, frequency-domain features, and statistical features are extracted and concatenated to form a fragment-level fastening state feature vector. Based on the fragment-level fastening state feature vectors, a fastening state feature vector sequence is formed, and a fastening state structure diagram is constructed. The Weisfeiler-Lehman algorithm is executed on the tight state structure graph. The node labels of each graph node are initialized by numerical interval mapping based on the fragment-level tight state feature vector. The node labels are combined with the labels of adjacent nodes and hash updates are performed. After completing a preset number of iterations, the tight state structure feature vector is formed. A normal sample set is constructed based on the historical fastening state structure feature vector. An improved soft first-class extreme learning machine model is constructed and trained. The real-time fastening state structure feature vector is input into the trained improved soft first-class extreme learning machine model, and the corresponding fastening state structure deviation index is output. Based on the structural deviation index of the fastening state and the preset classification judgment rules, the bolt fastening state judgment result is generated and written into the state history database.

2. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The multi-source fastening status data specifically includes vibration signal data, acoustic emission signal data, strain-related data, structural status auxiliary data, and operating condition and environmental data.

3. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The preprocessing of multi-source fastening state data specifically includes data synchronization and alignment, noise reduction and filtering, trend and drift correction, and amplitude unification and standardization.

4. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The formation of the segment-level fastening state feature vector includes: The standardized values ​​of various data types in the standardized fastening state data at each sampling time are arranged in chronological order to form a continuous data sequence; The continuous data sequence is segmented according to the preset segment length and preset segment step size to generate a set of tight state segments consisting of multiple independent and temporally continuous data subsequences. Based on the set of fastened state segments, feature extraction processing is performed on the continuous data sequences within the time interval corresponding to each fastened state segment to obtain the corresponding time-domain feature set, frequency-domain feature set and statistical feature set. The time-domain feature set, frequency-domain feature set, and statistical feature set corresponding to each fastened state segment are combined in sequence to generate the corresponding segment-level fastened state feature vector.

5. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The construction of the fastening state structure diagram includes: Based on the segment-level fastening state feature vectors, and arranged according to the sampling time order corresponding to each segment-level fastening state feature vector, a fastening state feature vector sequence is formed. The node set of the fastening state structure diagram is determined based on the fastening state feature vector sequence. Each segment-level fastening state feature vector is mapped to a node to obtain the node set, and the segment-level fastening state feature vector is used as the node attribute of the corresponding node. The edge set of the fastening state structure diagram is determined based on the node set. Temporal adjacency connection is established between adjacent nodes corresponding to the same bolt. Spatial adjacency connection is established between nodes corresponding to different bolts when the bolts are spatially adjacent to each other. Structural connection relationship is established when the bolts belong to the same connection structure and have a direct structural connection relationship. By integrating the node set, edge set, and node attributes, a tight state structure diagram is generated.

6. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The process of forming a tight-state structural feature vector after completing a preset number of iterations includes: Based on the tight state structure diagram, the Weisfeiler-Lehman algorithm is executed to perform multi-dimensional interval encoding on the fragment-level tight state feature vectors corresponding to each node in the node set. The feature values ​​of each dimension are mapped to the corresponding interval encoding indexes according to the preset interval division rules, and ordered multi-dimensional encoding labels are formed according to the dimensional order as the initial structure labels of each node. Under the current structural refinement iteration, the set of adjacent nodes that are adjacent to each node is obtained based on the edge set. The structural labels of each adjacent node in the adjacent node set are arranged to form an ordered adjacent label sequence. The ordered adjacent label sequence is truncated and padded to obtain a fixed-length adjacent label sequence table. Under the current structural refinement iteration, the structural label of each node is combined with the corresponding fixed-length adjacency label sequence list, and the combination result is hashed to update the structural label of each node; Repeat the structural refinement iteration until the preset number of iterations is reached. Stack the structural labels obtained in each structural refinement iteration in order of iteration to construct a node label iteration stacking matrix. Select the initial structural label and the structural label after completing the preset number of iterations for each node according to the preset skip layer rounds to generate the skip layer label path for each node and form a skip layer label aggregation structure. Based on the node label iterative stacking matrix, the jump-level label aggregation structure, and the tight state structure graph, a three-dimensional tensor graph is constructed. Based on the three-dimensional tensor graph, the structural labels corresponding to different nodes, adjacency relationships, and structural refinement iterations are subjected to multi-dimensional aggregation processing. The aggregation result is then mapped to a one-dimensional vector representation to generate a tight state structure feature vector.

7. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The construction and training of the improved soft first-class extreme learning machine model includes: Select the structural feature vectors of historical fastening states that are in a confirmed normal fastening state, aggregate them according to time index to form a normal sample set, and divide the normal sample set into a normal sample training subset and a normal sample verification subset; An improved soft first-class extreme learning machine model is constructed, which consists of a feature-carrying module, a mapping generation and fusion module, a hierarchical boundary modeling module, a boundary co-generation module, and a structural deviation evaluation module. During the training phase, graph structure sparsity processing is performed on the compact state structure graph of the normal sample training subset, and label offset processing is performed on the structure labels to generate a perturbation sample set. Based on the perturbation sample set, an approximate sample set outside the boundary is constructed. The normal sample training subset, the perturbation sample set, and the approximate sample set outside the boundary are used together as the training input set. Three soft class-one extreme learning machine sub-models are trained based on different structural feature subspaces. The structural feature subspaces include the global structural feature subspace, the local adjacency structural feature subspace, and the skip label path structural feature subspace. Each soft class-one extreme learning machine model completes boundary learning based on the training input set and obtains the corresponding class-one discrimination boundary. The weight parameters corresponding to each soft class I extreme learning machine sub-model are determined based on the normal sample validation subset. Weighted ensemble is then performed on each soft class I extreme learning machine sub-model to obtain the trained improved soft class I extreme learning machine model.

8. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 1, characterized in that, The output corresponding to the fastening state structural deviation index includes: The feature vector of the tight state structure is input into the feature carrying module of the trained improved soft first-class extreme learning machine model. The feature vector of the tight state structure is subjected to feature dimension consistency verification and structural formatting to form the input feature representation. The mapping generation and fusion module performs hierarchical channel mapping processing on the input feature representation, dividing the input feature representation into three structural sub-channels according to structural semantics. In each structural sub-channel, a parallel nonlinear mapping is performed using random mapping kernels corresponding to different activation functions to generate the hidden layer mapping results corresponding to each structural sub-channel and perform reconstruction and fusion processing to form a unified intermediate representation vector. The hierarchical boundary modeling module constructs soft first-class discrimination boundaries corresponding to different structural levels based on the intermediate representation vector, and substitutes the intermediate representation vector into the soft first-class discrimination boundary corresponding to each structural level to perform boundary discrimination calculation, generating the deviation of the intermediate representation vector from the soft first-class discrimination boundary of each structural level. The boundary collaborative generation module stacks the deviations corresponding to each structural level according to the structural level dimension to construct a three-dimensional boundary stacking representation, and performs cross-layer aggregation processing on the three-dimensional boundary stacking representation to form a unified first-class discriminative boundary representation; The structural deviation assessment module is based on a unified single-class discrimination boundary representation. It calculates the structural deviation components corresponding to the input feature representation under different structural perspectives, generates a structural deviation vector, performs mutually exclusive projection processing on the structural deviation vector, generates projection residual components between each structural perspective, performs multi-dimensional fusion operation on the projection residual components, and outputs the structural deviation risk index corresponding to the structural feature vector of the fastened state as the structural deviation degree index of the fastened state.

9. The intelligent monitoring method for bolt fastening status based on machine learning according to claim 7, characterized in that, The bolt fastening status determination result is generated and written into the status history database; To associate the structural deviation index of the fastened state with the corresponding bolt identifier, structural node identifier and time index, a record to be judged is formed. A hierarchical judgment rule is constructed based on a hierarchical threshold set, wherein the hierarchical threshold set includes a first threshold and a second threshold, and the first threshold is less than the second threshold; Based on the graded judgment rule, the deviation index of the fastened state structure is compared with the first threshold and the second threshold. When the deviation index of the fastened state structure is less than the first threshold, a normal fastening state judgment result is generated. When the deviation index of the fastened state structure is greater than or equal to the first threshold and less than the second threshold, a slightly loose transition state judgment result is generated; when the deviation index of the fastened state structure is greater than or equal to the second threshold, an abnormal risk state judgment result is generated. Write the record to be judged, the bolt tightness judgment result, and the structural deviation index of the tightness state into the state history database, and store the historical judgment record of the corresponding bolt in the state history database according to the time index.

Images