CN-121743782-B - Bridge disease dynamic sensing method and system based on vibration and acoustic time sequence signals

CN121743782BCN 121743782 BCN121743782 BCN 121743782BCN-121743782-B

Abstract

The invention provides a bridge disease dynamic sensing method and system based on vibration and acoustic time sequence signals, which relate to the technical field of bridge disease detection and comprise the following steps of S1, synchronously collecting multi-mode data of a bridge; S2, preprocessing the multi-mode data to obtain a multi-mode feature sequence, S3, performing cross-mode time sequence regulation and feature enhancement on the vibration feature sequence and the acoustic feature sequence by taking a causal direction as constraint to obtain fusion time sequence features with aligned semantics and distinguishable causal effects, S4, inputting the fusion time sequence features into an asymmetric double-process neural network to perform decoupling modeling to generate load state representation and structure response state representation, S5, performing event-level causal discovery and causal effect quantification based on the load state representation and the structure response state representation, and automatically generating a causal diagnosis report with confidence assessment so as to reveal disease causes, quantify load contributions and distinguish active diseases.

Inventors

LI NING
ZOU HAIYUN
BAI HAO
WEN YAN
ZHANG GUOQIANG
HE JIE
MA QIANG
ZHANG JIE
KUANG JING
LI XIN
JIA DONGLIN
XUE QIANLI
Xiang Baoshan
LI TAO
QIU NINGTAO
FAN DONGMING
HE XIAOYAN
SONG XUAN
CHEN CHENG
ZENG YAN
LIAO ZEMING
XU SHICONG
WANG JUN
ZHANG GUO
LI YAN
Xu Fangya
Yuan feiyun
LAN Fuan
SUN CHUN
CHEN XIAOCHONG
CI BIN
FU HONGJIE

Assignees

四川省公路规划勘察设计研究院有限公司
四川乐西高速公路有限责任公司
四川高速公路建设开发集团有限公司
四川川西投资管理有限责任公司

Dates

Publication Date: 20260508
Application Date: 20260226

Claims (9)

1. The bridge disease dynamic sensing method based on the vibration and acoustic time sequence signals is characterized by comprising the following steps of: S1, synchronously acquiring multi-mode data of a bridge, wherein the multi-mode data comprises a vibration original signal and an acoustic original signal; S2, preprocessing the multi-modal data to obtain a multi-modal characteristic sequence, wherein the multi-modal characteristic sequence comprises a vibration characteristic sequence and an acoustic characteristic sequence; S3, taking a causal direction as a constraint, performing cross-modal time sequence regulation and feature enhancement on the vibration feature sequence and the acoustic feature sequence to obtain fusion time sequence features which are aligned in a semantic manner and can be distinguished in causal manner; s3 specifically comprises the following steps: S31, constructing a micro-time sequence normalization function by taking a public hidden variable as a potential causal alignment anchor point so as to establish time alignment between a vibration characteristic sequence and an acoustic characteristic sequence; S32, constructing a static cross-mode hypergraph structure based on a time alignment result, and obtaining deep fusion characteristics through a multi-layer hypergraph convolution network and multi-scale context information; S33, introducing a causal perception feature selection mechanism, carrying out importance assessment on the deep fusion features, and identifying a core feature subset with potential causal driving force on a target disease variable based on an importance assessment result so as to form fusion time sequence features which are aligned in a semantic manner and can be distinguished in a causal manner; S4, inputting the fusion time sequence characteristics into an asymmetric double-process neural network for decoupling modeling to generate load state representation and structure response state representation, wherein the asymmetric double-process neural network comprises a load excitation modeling flow and a structure response modeling flow of structural asymmetry; S5, based on the load state representation and the structural response state representation, event-level causal discovery and causal effect quantification are carried out, and a bridge disease diagnosis report with quantified causal effect is generated.
2. The bridge defect dynamic sensing method based on vibration and acoustic time sequence signals according to claim 1, wherein the deep fusion characteristic obtaining process specifically comprises the following steps: the method comprises the steps of constructing a static cross-modal hypergraph structure based on a time mapping result, namely representing the characteristics of a vibration characteristic sequence and an acoustic characteristic sequence in each time step and the space-time field of the vibration characteristic sequence and the acoustic characteristic sequence as a hyperedge, wherein the hyperedge represents the characteristic vectors of all the characteristics in a sliding window near the time step; The implicit alignment relation of each feature is learned through a multi-layer hypergraph convolution network, the feature of each time step is fused with the context information of the cross-modal time sequence hyperedge to which the feature belongs, so that the similarity between feature vertexes is output, and the deep fusion feature is dynamically determined by searching the time step with the highest similarity.
3. The method of claim 1 wherein the structural asymmetry is characterized in that the load excitation modeling flow employs a self-attention mechanism with causal timing mask constraints and the structural response modeling flow employs a gated cross-attention mechanism to fuse context information from the load excitation modeling flow.
4. The bridge defect dynamic sensing method based on vibration and acoustic time sequence signals according to claim 3, wherein the process of inputting the fusion time sequence characteristics into an asymmetric double-process neural network for decoupling modeling is specifically as follows: Extracting dynamic modes of the fusion timing features on different time scales by a multi-scale timing encoder in the load excitation modeling flow, and generating the load state representation through a self-attention mechanism constrained by a causal timing mask; And extracting a structural response mode of the fused time sequence characteristic through a multi-scale time sequence encoder in the structural response modeling flow, and generating the structural response state representation through a gating cross attention mechanism fused with the load state representation.
5. The method for dynamically sensing bridge diseases based on vibration and acoustic time sequence signals according to claim 1, wherein a characteristic orthogonal regularization term is introduced when training the asymmetric double-process neural network to promote decoupling of load state representation and structural response state representation in a characteristic space.
6. The method for dynamically sensing bridge diseases based on vibration and acoustic time sequence signals according to claim 1, wherein the process of generating the bridge disease diagnosis report with quantitative causal effect is as follows: Carrying out event level detection and characterization construction on the load state representation and the structure response state representation to obtain a load event set and a structure response event set; based on the condition independence test, a causal graph result is learned from the load event set and the structural response event set, and a causal direction is determined in combination with a time sequence priori to obtain a directed acyclic graph, wherein the directed acyclic graph takes the load event and the structural response event as vertexes, and the causal direction is the direction of a directed edge between the vertexes; And carrying out causal effect quantification on each directed edge in the directed acyclic graph, and generating a bridge disease diagnosis report.
7. The bridge disease dynamic sensing method based on the vibration and acoustic time sequence signals is characterized by comprising the steps of preprocessing the multi-modal data to obtain a multi-modal feature sequence, and restraining time alignment between the vibration feature sequence and the acoustic feature sequence by taking the visual feature sequence as a space-time reference when the cross-modal time sequence regulation and feature enhancement are carried out on the vibration feature sequence and the acoustic feature sequence.
8. The bridge disease dynamic sensing method based on vibration and acoustic time sequence signals according to claim 7, wherein the process of preprocessing the multi-modal data to obtain the multi-modal feature sequence is as follows: zero offset correction and noise suppression processing are carried out on the vibration original signal, so that a vibration characteristic sequence is obtained; Performing multistage noise reduction and amplitude normalization processing on the acoustic original signal to obtain an acoustic feature sequence; and performing lens distortion correction processing on the video original data to obtain a visual characteristic sequence.
9. A bridge defect dynamic sensing system based on vibration and acoustic time sequence signals, which is characterized in that the bridge defect dynamic sensing method based on the vibration and acoustic time sequence signals is applied according to any one of claims 1-8, and comprises the following steps: The acquisition module is used for synchronously acquiring multi-mode data of the bridge, wherein the multi-mode data comprises a vibration original signal and an acoustic original signal; The preprocessing module is used for preprocessing the multi-mode data to obtain a multi-mode characteristic sequence, wherein the multi-mode characteristic sequence comprises a vibration characteristic sequence and an acoustic characteristic sequence; The feature engineering module takes the causal direction as constraint, and performs cross-modal time sequence regulation and feature enhancement on the vibration feature sequence and the acoustic feature sequence to obtain fusion time sequence features which are aligned in semantic and can be distinguished in causal; The dual-process cross-modal time sequence attention module inputs the fusion time sequence characteristics into an asymmetric dual-process neural network for decoupling modeling to generate load state representation and structure response state representation, wherein the asymmetric dual-process neural network comprises a load excitation modeling flow and a structure response modeling flow of structural asymmetry; And the causal discovery and quantitative analysis module is used for carrying out event-level causal discovery and causal effect quantification based on the load state representation and the structural response state representation and generating a bridge disease diagnosis report with quantified causal effect.

Description

Bridge disease dynamic sensing method and system based on vibration and acoustic time sequence signals Technical Field The invention relates to the technical field of bridge disease detection, in particular to a bridge disease dynamic sensing method and system based on vibration and acoustic time sequence signals. Background At present, the bridge disease detection technology mainly expands around three directions, namely, the development of an intelligent detection platform of detection equipment such as an unmanned plane, a wall climbing robot and the like, the refinement of a recognition algorithm including a crack segmentation method based on a transducer, a lightweight recognition method of improvement Segformer and the like, and the fusion of multi-mode data such as the fusion of non-destructive evaluation (NDE) data and an image processing framework. However, the directions are limited in that in the aspect of detection equipment, platforms such as an unmanned plane and a wall climbing robot promote automation and safety, but multiple sides focus on data acquisition and lack deep analysis on disease causes, in the aspect of recognition algorithm, the deep learning technology promotes recognition precision and efficiency, but focuses on apparent feature extraction and does not relate to dynamic association of loads and responses, in the aspect of multi-mode fusion, the existing method realizes data coordination and state analysis, but stays in a feature layer fusion or prediction layer mostly, does not realize process decoupling based on physical cause and effect, and lacks explanation capability of 'why diseases occur'. In summary, although the perceptions and recognition capabilities of the prior art are continuously improved, the dynamic association between traffic load and structural response is not generally established, and the detection is still limited to the "what you see is what you get" static mode, so that the cause and development trend of diseases cannot be really diagnosed, which is the core problem to be solved by the present invention. Disclosure of Invention The invention aims to provide a bridge disease dynamic sensing method based on vibration and acoustic time sequence signals, which is characterized in that the vibration and acoustic signals of a bridge are synchronously collected through hardware, and the physical causal direction is taken as a fundamental constraint to drive the subsequent full-link analysis, wherein firstly, a micro time sequence regular and hypergraph convolution network is adopted to realize the semantic alignment and characteristic enhancement of a cross-mode signal; and finally, carrying out event-level causal discovery and statistics quantification on the decoupled state representation, and automatically generating a causal diagnosis report with confidence assessment, thereby revealing disease causes, quantifying load contribution and distinguishing active diseases. In order to solve the technical problems, the invention adopts the following scheme: A bridge disease dynamic sensing method based on vibration and acoustic time sequence signals comprises the following steps: S1, synchronously acquiring multi-mode data of a bridge, wherein the multi-mode data comprises a vibration original signal and an acoustic original signal; S2, preprocessing the multi-modal data to obtain a multi-modal characteristic sequence, wherein the multi-modal characteristic sequence comprises a vibration characteristic sequence and an acoustic characteristic sequence; S3, taking a causal direction as a constraint, performing cross-modal time sequence regulation and feature enhancement on the vibration feature sequence and the acoustic feature sequence to obtain fusion time sequence features which are aligned in a semantic manner and can be distinguished in causal manner; S4, inputting the fusion time sequence characteristics into an asymmetric double-process neural network for decoupling modeling to generate load state representation and structure response state representation, wherein the asymmetric double-process neural network comprises a load excitation modeling flow and a structure response modeling flow of structural asymmetry; S5, based on the load state representation and the structural response state representation, event-level causal discovery and causal effect quantification are carried out, and a bridge disease diagnosis report with quantified causal effect is generated. Further, S3 specifically includes the following steps: S31, constructing a micro-time sequence normalization function by taking a public hidden variable as a potential causal alignment anchor point so as to establish time alignment between a vibration characteristic sequence and an acoustic characteristic sequence; S32, constructing a static cross-mode hypergraph structure based on a time alignment result, and obtaining deep fusion characteristics through a multi-layer hypergraph convolution network a