CN-122020482-A - Intelligent bolt fastening state monitoring method based on machine learning

CN122020482ACN 122020482 ACN122020482 ACN 122020482ACN-122020482-A

Abstract

The invention discloses an intelligent bolt fastening state monitoring method based on machine learning, which comprises the steps of collecting multisource fastening state data to obtain standardized fastening state data, carrying out segmentation and segmentation to form segment-level fastening state feature vectors, forming a fastening state feature vector sequence, constructing a fastening state structure diagram, executing WEISFEILER-Lehman algorithm to form fastening state structure feature vectors, outputting fastening state structure deviation index based on a trained improved soft extreme learning machine model, and generating a bolt fastening state judging result based on a preset grading judging rule. According to the invention, by introducing an improved soft extreme learning machine model and WEISFEILER-Lehman algorithm, low false alarm of the bolt fastening state under the condition of scarce abnormal samples is realized, and the intelligent monitoring and the early identification of loosening risk can be realized by leaning on a line.

Inventors

LI HAIPENG
ZHOU TIANZAI
YANG LI
CHEN YATAO
GU SHAOZU
YIN JIANG
LI ZHE
LIU WEI

Assignees

金翼安达航空科技（北京）有限公司
大唐伊吾清洁能源有限公司

Dates

Publication Date: 20260512
Application Date: 20260206

Claims (9)

1. The intelligent bolt fastening state monitoring method based on machine learning is characterized by comprising the following steps of: the method comprises the steps of collecting multisource fastening state data of a bolt connection structure in the running process, and preprocessing the multisource fastening state data to obtain standardized fastening state data; the standardized fastening state data is segmented to obtain a fastening state segment set, and for each fastening state segment, time domain features, frequency domain features and statistical features are extracted and spliced to form segment-level fastening state feature vectors; Forming a fastening state feature vector sequence based on the segment-level fastening state feature vector, and constructing a fastening state structure diagram; Executing WEISFEILER-Lehman algorithm on the fastening state structure diagram, carrying out numerical interval mapping initialization node labels based on fragment-level fastening state feature vectors on each diagram node, combining the node labels with adjacent node labels, and carrying out hash updating to form fastening state structure feature vectors after the preset iteration times are completed; Constructing a normal sample set based on the historical fastening state structure feature vectors, constructing and training an improved soft-class extreme learning machine model, inputting the real-time fastening state structure feature vectors into the trained improved soft-class extreme learning machine model, and outputting corresponding fastening state structure deviation indexes; and generating a bolt fastening state judgment result based on the fastening state structure deviation index and a preset grading judgment rule, and writing the bolt fastening state judgment result into a state history library.
2. The intelligent monitoring method for bolt tightening state based on machine learning according to claim 1, wherein the multisource tightening state data specifically comprises vibration signal data, acoustic emission signal data, strain related data, structural state auxiliary data and working condition and environment data.
3. The intelligent monitoring method for the bolt tightening state based on machine learning according to claim 1, wherein the preprocessing of the multi-source tightening state data comprises data synchronization and alignment, denoising and filtering processing, trend and drift correction, and amplitude unification and standardization processing.
4. The intelligent monitoring method for bolt tightening state based on machine learning according to claim 1, wherein the forming of the segment-level tightening state feature vector comprises: Arranging the standardized values of various data in the standardized fastening state data in time sequence at each sampling moment to form a continuous data sequence; Carrying out fragmentation and segmentation processing on the continuous data sequence according to the preset fragment length and the preset fragment step length to generate a fastening state fragment set formed by a plurality of mutually independent and time-continuous data subsequences; based on the fastening state segment sets, respectively executing feature extraction processing on continuous data sequences in the time interval corresponding to each fastening state segment to obtain corresponding time domain feature sets, frequency domain feature sets and statistical feature sets; And sequentially combining the time domain feature set, the frequency domain feature set and the statistical feature set corresponding to each fastening state segment to generate a corresponding segment-level fastening state feature vector.
5. The intelligent monitoring method for the fastening state of the bolt based on machine learning according to claim 1, wherein the constructing a fastening state structure diagram comprises: based on the segment-level fastening state feature vectors, arranging the segment-level fastening state feature vectors according to the sampling time sequence corresponding to each segment-level fastening state feature vector to form a fastening state feature vector sequence; Determining a node set of a fastening state structure diagram based on the fastening state feature vector sequence, correspondingly taking each segment-level fastening state feature vector as a node to obtain the node set, and taking the segment-level fastening state feature vector as a node attribute of the corresponding node; Determining an edge set of a fastening state structure diagram based on the node set, establishing a time adjacent connection relationship between adjacent nodes corresponding to the same bolt, establishing a space adjacent connection relationship between nodes corresponding to different bolts when the bolts are adjacent to each other in space position, and establishing a structure connection relationship when the bolts belong to the same connection structure and have a direct structure connection relationship; and integrating the node set, the edge set and the node attribute to generate a fastening state structure diagram.
6. The intelligent monitoring method for the fastening state of the bolt based on machine learning according to claim 1, wherein the forming of the fastening state structural feature vector after the completion of the preset number of iterations includes: Executing WEISFEILER-Lehman algorithm based on the fastening state structure diagram, executing multidimensional interval coding processing on segment-level fastening state feature vectors corresponding to all nodes in the node set, mapping all the dimensional feature values into corresponding interval coding indexes according to preset interval dividing rules, and forming ordered multidimensional coding labels according to dimension sequences to serve as initial structure labels of all the nodes; Under the current structure refinement iteration, acquiring an adjacent node set adjacent to each node based on the edge set, arranging structural labels of each adjacent node in the adjacent node set to form an ordered adjacent label sequence, and cutting off and supplementing the ordered adjacent label sequence to obtain a fixed-length adjacent label sequence table; Under the current structure refinement iteration, combining the structure label of each node with a corresponding fixed-length adjacent label sequence table, and executing hash update on the combined result to update the structure label of each node; Repeatedly executing structure refinement iteration until the preset iteration times are reached, performing cross-layer stacking on the structure labels obtained under each structure refinement iteration round according to the iteration round sequence, constructing a node label iteration stacking matrix, selecting an initial structure label and a structure label after the preset iteration times are completed for each node according to the preset layer-skipping round, generating a layer-skipping label path of each node, and forming a layer-skipping label aggregation structure; based on the node label iteration stacking matrix, the layer jump label aggregation structure and the fastening state structure diagram, a three-dimensional tensor diagram is constructed, multi-dimensional aggregation processing is carried out on the structure labels corresponding to different nodes, adjacent relations and structure refinement iteration rounds based on the three-dimensional tensor diagram, an aggregation result is mapped into a one-dimensional vector representation, and a fastening state structure characteristic vector is generated.
7. The intelligent monitoring method for bolt tightening state based on machine learning according to claim 1, wherein the building and training of the improved soft-class extreme learning machine model comprises: Selecting a structural feature vector of a historical fastening state in a normal fastening state, converging according to time indexes to form a normal sample set, and dividing the normal sample set into a normal sample training subset and a normal sample verification subset; An improved soft-class extreme learning machine model is constructed, and the improved soft-class extreme learning machine model consists of a characteristic bearing module, a mapping generation fusion module, a layered boundary modeling module, a boundary collaborative generation module and a structural deviation evaluation module; In the training stage, performing graph structure sparse processing on a fastening state structure diagram of a normal sample training subset, performing label offset processing on a structure label to generate a disturbance sample set, constructing an out-boundary approximate sample set based on the disturbance sample set, and taking the normal sample training subset, the disturbance sample set and the out-boundary approximate sample set together as a training input set; Respectively training 3 soft-class extreme learning machine sub-models based on different structural feature subspaces, wherein the structural feature subspaces comprise a global structural feature subspace, a local adjacent structural feature subspace and a layer jump label path structural feature subspace, and each soft-class extreme learning machine sub-model completes boundary learning based on a training input set and obtains a corresponding class of discrimination boundary; And determining weight parameters corresponding to the soft-class extreme learning machine sub-models based on the normal sample verification subset, and performing weighted integration on the soft-class extreme learning machine sub-models to obtain the trained improved soft-class extreme learning machine model.
8. The intelligent monitoring method for the fastening state of the bolt based on machine learning according to claim 1, wherein the outputting the corresponding deviation index of the fastening state structure comprises: Inputting the fastening state structure feature vector into a feature bearing module of the trained improved soft-class extreme learning machine model, and executing feature dimension consistency check and structure formatting processing on the fastening state structure feature vector to form an input feature representation; The mapping generation fusion module performs hierarchical channel mapping processing on the input feature representation, divides the input feature representation into 3 structural sub-channels according to structural semantics, performs parallel nonlinear mapping by adopting random mapping cores corresponding to different activation functions in each structural sub-channel, generates hidden layer mapping results corresponding to each structural sub-channel, and performs reconstruction fusion processing to form a unified intermediate representation vector; The hierarchical boundary modeling module respectively constructs soft class discrimination boundaries corresponding to different structure levels based on the intermediate expression vectors, and respectively substitutes the intermediate expression vectors into the soft class discrimination boundaries corresponding to the structure levels to perform boundary discrimination calculation, so as to generate deviation representing the intermediate expression vectors relative to the soft class discrimination boundaries of the structure levels; The boundary collaborative generation module stacks the deviation corresponding to each structure level according to the dimension of the structure level, constructs a three-dimensional boundary stacking representation, and performs cross-layer aggregation processing on the three-dimensional boundary stacking representation to form a unified discrimination boundary representation; The structural deviation evaluation module calculates structural deviation components corresponding to the input characteristic representation under different structural view angles based on the unified type of discrimination boundary representation, generates structural deviation vectors, performs mutual exclusion projection processing on the structural deviation vectors, generates projection residual components among the structural view angles, performs multidimensional fusion operation on the projection residual components, and outputs structural deviation risk indexes corresponding to the fastening state structural characteristic vectors as fastening state structural deviation degree indexes.
9. The intelligent monitoring method for bolt tightening state based on machine learning according to claim 7, wherein the bolt tightening state determination result is generated and written into a state history library; Associating corresponding bolt marks, structure node marks and time indexes for the structure deviation index in the fastening state to form a record to be judged; Constructing a classification decision rule based on a classification threshold set, the classification threshold set comprising a first threshold and a second threshold, the first threshold being less than the second threshold; Comparing the fastening state structure deviation index with a first threshold value and a second threshold value based on a grading judgment rule, and generating a normal fastening state judgment result when the fastening state structure deviation index is smaller than the first threshold value; generating a slight loosening transition state judgment result when the structural deviation index of the fastening state is larger than or equal to a first threshold value and smaller than a second threshold value, and generating an abnormal risk state judgment result when the structural deviation index of the fastening state is larger than or equal to the second threshold value; Writing the record to be judged, the bolt fastening state judging result and the fastening state structure deviation index into a state history library, and storing the history judging record corresponding to the bolt mark in the state history library according to the time index.

Description

Intelligent bolt fastening state monitoring method based on machine learning Technical Field The invention relates to the technical field of machine learning, in particular to an intelligent bolt fastening state monitoring method based on machine learning. Background With the development of large equipment and complex machinery, the method and engineering structure of the invention are developed to high reliability and long-period running direction, the bolt connection structure is used as a key basic connection form, and the stability of the fastening state of the bolt connection structure has an important influence on the safety of the whole structure. In the actual operation process, the bolt connection is subjected to the combined action of vibration load, impact load, temperature change and environmental factors for a long time, and the pre-tightening force attenuation and loosening hidden trouble are easy to generate. The traditional bolt fastening state monitoring technology is mostly based on vibration signals, acoustic emission signals or strain data for analysis, and is combined with an empirical threshold or a simple rule for state judgment. Normally, based on single or small signal characteristics, normal fluctuation caused by working condition change is difficult to effectively respond, the system is sensitive to noise, and false alarm or missing alarm is easy to generate in a complex engineering environment. Part of the prior art starts to introduce a machine learning method, and state identification is realized by extracting characteristics of multi-source monitoring data and training a model. Attempts have also been made to describe the spatial or structural association between bolts using a graph structure to enhance the ability to characterize the overall connection state. However, the existing related method is mostly in a feature level or a simple relation level modeling level, the related evolution of the fastening state under different structure levels is not characterized sufficiently, normal fluctuation and real loosening abnormality are difficult to distinguish reliably under complex working conditions, and the utilization depth and discrimination stability of the model to the structure information are still obviously limited. Therefore, how to provide an intelligent monitoring method for the fastening state of a bolt based on machine learning is a problem to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide an intelligent bolt fastening state monitoring method based on machine learning, which fully utilizes multi-source monitoring data acquisition, structural feature modeling and a machine learning method, and describes a monitoring flow for realizing online intelligent judgment of the bolt fastening state under the condition of scarce abnormal samples in detail. According to the method, the structure diagram representing the evolution relation of the fastening state of the bolt is constructed by uniformly preprocessing and extracting the fragmentation characteristics of the multi-source data, the structural characteristics of the fastening state are subjected to multi-level representation by utilizing WEISFEILER-Lehman algorithm, the soft-class extreme learning machine model is further combined with the improvement, the soft boundary description is formed for the normal fastening state, and the quantitative evaluation of the deviation degree of the structure of the fastening state is realized. The invention realizes the reliable distinction between normal fluctuation and real loosening abnormality in the bolt fastening state, has the advantages of low dependence on abnormal samples, strong structural feature expression capability, high adaptability to complex working conditions, low false alarm rate and the like, and is suitable for online intelligent monitoring of the bolt connection structure in complex engineering environment. According to the embodiment of the invention, the intelligent monitoring method for the bolt fastening state based on machine learning comprises the following steps: the method comprises the steps of collecting multisource fastening state data of a bolt connection structure in the running process, and preprocessing the multisource fastening state data to obtain standardized fastening state data; the standardized fastening state data is segmented to obtain a fastening state segment set, and for each fastening state segment, time domain features, frequency domain features and statistical features are extracted and spliced to form segment-level fastening state feature vectors; Forming a fastening state feature vector sequence based on the segment-level fastening state feature vector, and constructing a fastening state structure diagram; Executing WEISFEILER-Lehman algorithm on the fastening state structure diagram, carrying out numerical interval mapping initialization node labels based on fragment-level fastening s